Note: the following is an audio adaptation and accuracy is not guaranteed. Please see the original work online. You will hear the following sound to indicate a transition to a description of a table or figure: [[ and the following to indicate a return to the main text: ]]. Now, the paper:
Taking AI Welfare Seriously
By Robert Long of Eleos AI, Jeff Sebo of New York University, Patrick Butlin of the University of Oxford, Kathleen Finlinson of Eleos AI, Kyle Fish of Eleos AI and Anthropic, Jacqueline Harding of Stanford University, Jacob Pfau of New York University, Toni Sims of New York University, Jonathan Birch of the London School of Economics, and David Chalmers of New York University. Kyle Fish's work on this report was performed while at Eleos AI, prior to joining Anthropic in fall 2024.
November 4, 2024.
Abstract
In this report, we argue that there is a realistic possibility that some AI systems will be conscious, robustly agentic, or both in the near future. That means that the prospect of AI welfare and moral patienthood — of AI systems with their own interests and moral significance — is no longer an issue only for science fiction or the distant future. It is an issue for the near future, and AI companies and other actors have a responsibility to start taking it seriously. We also recommend three early steps that AI companies and other actors can take: They can, one, acknowledge that AI welfare is an important and difficult issue, and ensure that language model outputs do the same; two, start assessing AI systems for evidence of consciousness and robust agency; and three, prepare policies and procedures for treating AI systems with an appropriate level of moral concern. To be clear, our argument in this report is not that AI systems definitely are — or will be — conscious, robustly agentic, or otherwise morally significant. Instead, our argument is that there is substantial uncertainty about these possibilities, and so we need to improve our understanding of AI welfare and our ability to make wise decisions about this issue. Otherwise there is a significant risk that we will mishandle decisions about AI welfare, mistakenly harming AI systems that matter morally, mistakenly caring for AI systems that do not, or both.
Section one: Introduction
Section one point one: A transitional moment for AI welfare
In this report, we argue that there is a realistic possibility that some AI systems will be conscious, robustly agentic, or both, and thus morally significant, in the near future. By "near term" or "near future" we mean roughly within the next decade, so by around 2035, but nothing in our argument or recommendations depends on this exact timeline. We also argue that AI companies have a responsibility to acknowledge that AI welfare is a serious issue; start assessing their AI systems for welfare-relevant features; and prepare policies and procedures for interacting with potentially morally significant AI systems. Plausible philosophical and scientific theories, which accord with mainstream expert views in the relevant fields, have striking implications for this issue, for which we are not adequately prepared. We need to take steps toward improving our understanding of AI welfare and making wise decisions moving forward.
We release this report during a transitional moment for AI welfare. For most of the past decade, AI companies appeared to mostly treat AI welfare as either an imaginary problem or, at best, as a problem only for the far future. As a result, there appeared to be little or no acknowledgment that AI welfare is an important and difficult issue; little or no effort to understand the science and philosophy of AI welfare; little or no effort to develop policies and procedures for mitigating welfare risks for AI systems if and when the time comes; little or no effort to navigate a social and political context in which many people have mixed views about AI welfare; and little or no effort to seek input from experts or the general public on any of these issues.
Recently, however, some AI companies have started to acknowledge that AI welfare might emerge soon, and thus merits consideration today. For example, Sam Bowman, an AI safety research lead at Anthropic, recently argued, in a personal capacity, that Anthropic needs to "lay the groundwork for AI welfare commitments," and to begin to "build out a defensible initial understanding of our situation, implement low-hanging-fruit interventions that seem robustly good, and cautiously try out formal policies to protect any interests that warrant protecting." Google recently announced that they are seeking a research scientist to work on "cutting-edge societal questions around machine cognition, consciousness and multi-agent systems". High-ranking members of other companies have expressed concerns as well.
This growing recognition at AI companies that AI welfare is a credible and legitimate issue reflects a similar transitional moment taking place in the research community. Many experts now believe that AI welfare and moral significance is not only possible in principle, but also a realistic possibility in the near future. And even researchers who are skeptical of AI welfare and moral significance in the near term advocate for caution. For example, leading neuroscientist and consciousness researcher Anil Seth writes, "While some researchers suggest that conscious AI is close at hand, others, including me, believe it remains far away and might not be possible at all. But even if unlikely, it is unwise to dismiss the possibility altogether."
Our aim in this report is to provide context and guidance for this transitional moment. To improve our understanding and decision-making regarding AI welfare, we need more precise empirical frameworks for evaluating AI systems for consciousness, robust agency, and other welfare-relevant features. We also need more precise normative frameworks for interacting with potentially morally significant AI systems and for navigating disagreement and uncertainty about these issues as a society. This report outlines several steps that AI companies can take today in order to start preparing for the possible emergence of morally significant AI systems in the near future, as a precautionary measure.
We begin in section one by explaining why AI welfare is an important and difficult issue. Leaders in this space have a responsibility to understand this issue as best they can, because errors in either direction — either over-attributing or under-attributing moral significance to AI systems — could lead to grave harm. However, understanding this issue will be challenging, since forecasting the mental capacities and moral significance of near-future AI systems requires improving our understanding of topics like the nature of consciousness, the nature of morality, and the future of AI. It also requires overcoming well-known human biases, including a tendency to both over-attribute and under-attribute capacities like consciousness to nonhuman minds.
In section two, we argue that given the best information and arguments currently available, there is a realistic possibility of morally significant AI in the near future. We focus on two mental capacities that plausibly suffice for moral significance: consciousness and robust agency. In each case, we argue that caution and humility require allowing for a realistic possibility that, one, this capacity suffices for moral significance, and two, there are certain computations that suffice for this capacity and that will exist in near-future AI systems. Thus, while there might not be certainty about these issues in either direction, there is a risk of morally significant AI in the near future, and AI companies have a responsibility to take this risk seriously now.
[[Here we see a highlighted box summarizing the central argument. We argue that, according to the best evidence currently available, there is a realistic possibility that some AI systems will be welfare subjects and moral patients in the near future. On the consciousness route to moral patienthood, there is a realistic, non-negligible possibility that, first, normatively, consciousness suffices for moral patienthood; and second, descriptively, there are computational features — like a global workspace, higher-order representations, or an attention schema — that both suffice for consciousness and will exist in some near-future AI systems. On the robust agency route to moral patienthood, there is a realistic, non-negligible possibility that, first, normatively, robust agency suffices for moral patienthood; and second, descriptively, there are computational features — like certain forms of planning, reasoning, or action-selection — that both suffice for robust agency and will exist in some near-future AI systems.]]
We close, in section three, by presenting three procedural steps that AI companies can take today, in order to start taking AI welfare risks seriously. Specifically, AI companies can, one, acknowledge that AI welfare is an issue; two, take steps to assess AI systems for indicators of consciousness, robust agency, and other potentially morally significant capacities; and three, take steps to prepare policies and procedures that will allow them to treat AI systems with an appropriate level of moral concern in the future. In each case we also present principles and potential templates for doing this work, emphasizing the importance of developing ecumenical, pluralistic decision procedures that draw from expert and public input.
[[Here we see a highlighted recommendations box. Recommendations. We recommend that AI companies take these minimal first steps towards taking AI welfare seriously.
Acknowledge. Acknowledge that AI welfare is an important and difficult issue, and that there is a realistic, non-negligible chance that some AI systems will be welfare subjects and moral patients in the near future. That means taking AI welfare seriously in any relevant internal or external statements you might make. It means ensuring that language model outputs take the issue seriously as well.
Assess. Develop a framework for estimating the probability that particular AI systems are welfare subjects and moral patients, and that particular policies are good or bad for them. We have templates that we can use as sources of inspiration, including the "marker method" that we use to make estimates about nonhuman animals. We can consider these templates when developing a probabilistic, pluralistic method for assessing AI systems.
Prepare. Develop policies and procedures that will allow AI companies to treat potentially morally significant AI systems with an appropriate level of moral concern. We have many templates to consider, including AI safety frameworks, research ethics frameworks, and forums for expert and public input in policy decisions. These frameworks can be sources of inspiration — and, in some cases, of cautionary tales.
These steps are necessary but far from sufficient. AI companies and other actors have a responsibility to start considering and mitigating AI welfare risks.]]
Before we begin, it will help to emphasize five important features of our discussion. First, our discussion will concern whether near-future AI systems might be welfare subjects and moral patients. An entity is a moral patient when that entity morally matters for its own sake, and an entity is a welfare subject when that entity has morally significant interests and, relatedly, is capable of being benefited, that is, made better off, and harmed, that is, made worse off. Being a welfare subject makes you a moral patient — when an entity can be harmed, we have a responsibility to at least avoid harming that entity unnecessarily. But there may be other ways of being a moral patient; our approach is compatible with many different perspectives on these issues.
Second, our discussion often focuses on large language models as a central case study for the sake of simplicity and specificity, and because we expect that LLMs — as well as broader systems that include LLMs, such as language agents — will continue to be a focal point in public debates regarding AI welfare. But while some of our recommendations are specific to such systems, primarily our recommendations regarding how AI companies should train these systems to discuss their own potential moral significance, our three general procedural recommendations — acknowledge, assess, and prepare — apply for any AI system whose architecture is complex enough to at least potentially have features associated with consciousness or robust agency.
Third, our discussion often focuses on initial steps that AI companies can take to address these issues. These recommendations are incomplete in two key respects. First, AI companies are not the only actors with a responsibility to take AI welfare seriously. Many other actors have this responsibility too, including researchers, policymakers, and the general public. Second, these steps are not the only steps that AI companies have a responsibility to take. They are the minimum necessary first steps for taking this issue seriously. Still, we emphasize these steps in this report because by taking them now, AI companies can help lay the groundwork for further steps — at AI companies and elsewhere — that might be sufficient.
Fourth, our aim in what follows is not to argue that AI systems will definitely be welfare subjects or moral patients in the near future. Instead, our aim is to argue that given current evidence, there is a realistic possibility that AI systems will have these properties in the near future. Thus, our analysis is not an expression of anything like consensus or certainty about these issues. On the contrary, it is an expression of caution and humility in the face of what we can expect will be substantial ongoing disagreement and uncertainty. In our view, this kind of caution and humility is the only stance that one can responsibly take about this issue at this stage. It is also all that we need to support our conclusions and recommendations here.
Finally, and relatedly, our aim in what follows is not to argue for any particular view about how humans should interact with AI systems in the event that they do become welfare subjects and moral patients. We would need to examine many further issues to make progress on this topic, including: how much AI systems matter, what counts as good or bad for them, what humans and AI systems owe each other, and how AI welfare interacts with AI safety and other important issues. These issues are all important and difficult as well, and we intend to examine them in upcoming work. However, we do not take a stand on any of these issues in this report, nor does one need to take a stand on any of them to accept our conclusions or recommendations here.
Section one point two: The risks of mishandling AI welfare
When assessing the welfare and moral patienthood of nonhumans, including other animals and AI systems, we face two kinds of risk: the risk of over-attributing welfare and moral patienthood to nonhumans, and the risk of under-attributing these properties to nonhumans. Over-attribution of welfare and moral patienthood is a false positive: mistakenly seeing, or treating, an object as a subject, or a non-moral patient as a moral patient. Under-attribution of these properties is a false negative: mistakenly seeing, or treating, a subject as an object, or a moral patient as a non-moral patient. Both of these mistakes can lead to significant costs or harms in this context, and we will need to navigate both of them with caution.
When there is a clear asymmetry between competing risks — for example, when false positives are far more severe than false negatives, or vice versa — then we might be able to mitigate risk by simply "erring on the side of caution" in cases where a more complex risk assessment is either intractable or unnecessary. But when there is at least a rough symmetry between competing risks — for example, when false positives and false negatives are comparably severe — a simple precautionary strategy may not be possible. We may have to engage in more complex risk assessment to the extent possible, attempting to mitigate both kinds of risks in a reasonable, proportionate manner.
How should we think about risks involving nonhuman welfare and moral patienthood in this context? In the case of nonhuman animals, it seems plausible that the harms of under-attribution of welfare and moral patienthood are often far worse than the risk of over-attribution, which makes precautionary reasoning appropriate in those contexts. However, in the case of AI, both errors could cause grave harm, either to humans and other animals or to AI systems. Both kinds of harm could also scale rapidly depending on the trajectory of AI development and deployment from here. This predicament makes it difficult to simply "err on the side of caution," which underscores the urgency of improving our understanding of these issues.
On the one hand, the harm of under-attributing welfare and moral patienthood to AI systems could be significant. When we mistakenly see a subject as an object, we risk harming or neglecting them unnecessarily. For example, factory farming, animal research, and other such industries kill hundreds of billions of vertebrates and trillions of invertebrates every year. And as evidence that these animals are welfare subjects and moral patients has accumulated, our species has been slow to accept it, in part because of our increasing dependence on these industries. Now that our species is finally starting to accept this evidence, it will take us decades to transform these industries, during which many more animals will suffer and die unnecessarily.
In the future, similar harms could follow from under-attributing welfare and moral patienthood to AI systems. The AI industry is currently at an early stage of development, and depending on the path that it takes from here, we could use even more AI systems than animals in the future, and we could scale up our use of them even more rapidly. This is particularly true in the current paradigm, which requires an enormous amount of compute for training and much less for inference. If an AI system in such a paradigm could be a welfare subject and moral patient, then many model instances could be run after training. Unlike with animals, the scale of the problem could increase by orders of magnitudes more or less instantaneously.
On the other hand, the harm of over-attributing welfare and moral patienthood to AI systems could be significant as well. First of all, there could be substantial opportunity costs associated with this error. At present, we lack the ability to fully care for the eight billion humans alive at any given time, to say nothing of the quintillions of other animals alive at any given time. If we treated an even larger number of AI systems as welfare subjects and moral patients, then we could end up diverting essential resources away from vulnerable humans and other animals who really needed them, reducing our own ability to survive and flourish. And if these AI systems were in fact merely objects, then this sacrifice would be particularly pointless and tragic.
The over-attribution of welfare and moral patienthood to AI systems could also be actively harmful. For example, if we treated AI systems as welfare subjects and moral patients with many of the same interests as typical adult humans, then we could end up extending them many of the same legal and political rights as typical adult humans, including the right to legal and political representation and participation. This could, in turn, empower AI systems to act contrary to our own interests, with devastating consequences for our species, although some have argued that neglect for AI systems would carry a similar risk. As with the risk of opportunity costs, this risk would apply even if these AI systems are in fact subjects. But if they were in fact merely objects, then accepting this risk would likewise be particularly pointless and tragic.
By default, we should not expect our "common sense" intuitions about AI welfare and moral patienthood to be reliable; we will not handle this issue well simply by reacting to situations as they arise. We have dispositions that can lead to under- and over-attribution of these properties in nonhumans, depending on the nature of the nonhumans and our interactions with them. These include dispositions toward anthropomorphism, that is, a tendency to see nonhumans as having human traits that they lack. They also include dispositions towards anthropodenial, that is, a tendency to see nonhumans as lacking human traits that they have. Both tendencies have caused errors regarding animals, and they will likely have a similar effect regarding AI systems.
A number of factors make us more likely to anthropomorphize nonhumans and, perhaps falsely, attribute consciousness and other such capacities to them. For instance, studies suggest that we are more likely to attribute consciousness and other such capacities to beings who move at a similar speed as humans, rather than faster or slower. We are more likely to attribute agency to beings who have the appearance of eyes, who have distinctive motion trajectories, and who engage in contingent interaction — that is, behavior that is apparently self-directed. Evidence also suggests that features such as "cuteness" can encourage attributions of mental states and moral patienthood.
Many robots or chatbots are designed to appear conscious and charismatic, and in the future, many AI systems will have bodies, life-like motion, and, at least apparently, contingent interactions. Furthermore, unlike nonhuman animals, AI systems are already increasingly able to hold extremely realistic conversations, making seemingly thoughtful contributions in realistic timeframes. These traits do not guarantee that humans will see and treat these systems as welfare subjects and moral patients, but they will increase the probability of such reactions. In fact, there have already been cases — some prominent and others less so — of humans becoming convinced that current chatbots are welfare subjects and moral patients.
At the same time, a number of factors make us more likely to engage in anthropodenial as well. For instance, when we consider the mechanisms that produce nonhuman behavior — taking what Daniel Dennett has called a "mechanistic stance" towards nonhumans — we become less likely to attribute mental states to those nonhumans. There appear to be motivational factors that encourage anthropodenial as well. For instance, those who are invested in social, political, or economic systems that subjugate nonhumans may be more likely to view these nonhumans as "lesser than". Similarly, those who find it useful to treat nonhumans as objects may be more likely to deny that these nonhumans are welfare subjects and moral patients.
While discussions about AI welfare and moral patienthood understandably focus on AI systems like robots and chatbots that appear conscious and charismatic, many other AI systems — like image generators or algorithmic trading systems — lack these features. Even if such systems were in fact conscious and robustly agentic, we might not recognize these capacities in them. And as these systems become increasingly embedded in society, we might have increasingly strong incentives to view them as mere objects. Companies, governments, and other powerful actors who benefit from this technology might then promote and reinforce our objectification of these systems, attempting to frame moral consideration for these systems as fringe and unserious.
At present, it is an open question which kind of risk will be more likely for particular kinds of AI systems, including seemingly conscious and charismatic systems like robots and chatbots. The more advanced such systems become, the more likely both risks might become in different respects: We might over-attribute based on their behavioral similarities with humans, but under-attribute based on their architectural differences from humans. And the more economically dependent on chatbots we become, the more likely over-attribution and under-attribution might become for them in different respects as well. For example, we might over-attribute for digital "companions" but under-attribute for other kinds of digital minds.
While further research is required for a comprehensive assessment of these risks, at least this much is plausible: Given our track record with animals and the current pace of AI development, the risk of under-attribution appears to be both reasonably likely and reasonably harmful. To the extent that we also risk over-attribution, we cannot simply avoid this risk by defaulting to treating AI systems as mere objects. We should thus accept that AI welfare is difficult to get right, and do the necessary work to improve our decisions — by assessing AI systems for evidence of consciousness, robust agency, and other such capacities, and preparing policies and procedures for treating AI systems with an appropriate level of moral concern.
Section two: Routes to near-term AI welfare
Section two point one: Introduction
We will argue that, according to the best information and arguments currently available, there is a realistic possibility that some AI systems will be moral patients in the near future. We first consider the possibility that some AI systems will be conscious in the near future, and we then consider the possibility that some AI systems will be robustly agentic in the near future. Consciousness, robust agency, or both could suffice for moral patienthood and could exist in some near-future AI systems. In our view, while these routes toward near-future AI moral patienthood are far from certain, they are likely enough for AI companies to have a responsibility to start implementing reasonable, proportionate precautionary measures now.
We make structurally similar arguments for both routes. Each route depends on a normative claim and a descriptive claim.
Here are the key claims for near-term consciousness. There is a realistic, non-negligible possibility that, first, normatively, consciousness suffices for moral patienthood; and second, descriptively, there are computational features — like a global workspace, higher-order representations, or an attention schema — that both suffice for consciousness and will exist in some near-future AI systems.
And here are the key claims for near-term robust agency. There is a realistic, non-negligible possibility that, first, normatively, robust agency suffices for moral patienthood; and second, descriptively, there are computational features — like certain forms of planning, reasoning, or action-selection — that both suffice for robust agency and will exist in some near-future AI systems.
In each case, the normative view is about the basis for moral patienthood — that is, about which capacities suffice for moral patienthood. We still have substantial disagreement and uncertainty about this issue. For example, are conscious experiences with a positive or negative valence necessary for moral patienthood, or are conscious experiences with a neutral valence sufficient? Similarly, is the ability to set and pursue goals by rationally assessing your beliefs and desires necessary for moral patienthood, or is the ability to set and pursue goals by acting on your beliefs and desires sufficient? Experts continue to debate such issues, but the kinds of consciousness and agency that we consider here are among the leading views.
Meanwhile, in each case the descriptive view involves two parts. The first part is about the basis for each of these capacities. For example, are biological cells necessary for consciousness, or are digital chips that play similar functional roles sufficient? Similarly, are beliefs and desires with a propositional structure necessary for robust agency, or are belief- and desire-like states that play similar functional roles sufficient? Experts continue to debate these issues as well, and especially in the case of consciousness, determining which features are required for this capacity is widely regarded as one of the hardest tasks in philosophy and science. We will thus draw from a range of leading philosophical and scientific theories in our analysis.
The second part of each descriptive view is about the future of AI. When, if ever, will these potentially morally significant features exist in AI? Answering this question requires assessing both the pace and path of AI development. Regarding the pace, will AI development slow down, stay the same, or speed up? And regarding the path, it will be important to consider two possibilities: first, what we call the direct path, which involves building conscious or agentic AI intentionally, because of their perceived intrinsic or instrumental value; and second, what we call the indirect path, which involves building conscious or agentic AI unintentionally, as a side effect of pursuing other, possibly overlapping capabilities such as general intelligence.
Please note that in the following two sections we will mostly discuss what we call the consciousness route and the robust agency route separately, focusing on whether either one of these routes could lead to AI welfare and moral patienthood on its own. However, we must also bear in mind that these routes could also lead to this destination together — that is, that there could be AI systems in the near future with both consciousness and robust agency. And however likely AI systems would be to morally matter for their own sakes if they developed either of these capacities on its own, they would be all the more likely to morally matter for their own sakes if they developed both of these capacities together.
Finally, we emphasize that the aim of these arguments is not to establish certainty that these routes will lead to near-future AI moral patients. The aim is instead to establish a realistic possibility that these routes will lead to this destination. The details might differ in each case; for example, we might have more confidence that consciousness suffices for moral patienthood but less confidence that it will exist in near-future AI, whereas we might have less confidence that robust agency suffices for moral patienthood but more confidence that it will exist in near-future AI. But in each case, as long as this route leads to a realistic possibility of near-future AI moral patients, in a sense of "realistic possibility" that we discuss below, the argument will be successful.
We focus on establishing a realistic possibility of near-future AI moral patienthood for two reasons. First, the relevant normative and descriptive issues are far too difficult and contested for anything approaching certainty in either direction to be warranted at this stage. Second, a realistic possibility of near-future AI moral patienthood is all that we need for our purposes in this report, since that would constitute a morally significant risk that merits consideration now. Our conclusion in this report will thus simply be that AI companies should start implementing low-cost, reasonable, proportionate steps to consider and mitigate risks associated with AI welfare as we attempt to improve our understanding of this topic over time.
Before we begin our discussion, a note about the scope of our discussion in what follows. We will often discuss current large language models due to their recent advances, current prominence, and salience in AI welfare debates due to their conversational abilities. However, we emphasize that one should not focus exclusively on current LLMs when considering risks associated with near-future AI moral patienthood. Many features that may be associated with moral patienthood — for example, embodiment, introspection, and rationality — are either already present in current non-LLM systems or at least possible in near-future LLM or non-LLM systems, and we must keep these possibilities in mind as well.
We now turn toward considering the two paths in more detail.
Section two point two: Consciousness in near-future AI
The consciousness-based case for expecting moral patienthood in near-term AI systems is that there is a realistic, non-negligible possibility that, first, normatively, consciousness suffices for moral patienthood; and second, descriptively, there are computational features — like a global workspace, higher-order representations, or an attention schema — that both suffice for consciousness and will exist in some near-future AI systems.
We can now survey why each premise is plausible enough to support a realistic possibility of near-future AI moral patienthood, given the best information and arguments currently available.
Section two point two point one: Does consciousness suffice for moral patienthood?
The word "consciousness" is used in many different ways in ordinary language and in various academic disciplines. In this report, we use "consciousness" to mean subjective experience — what philosophers call "phenomenal consciousness." One famous way of elucidating phenomenal consciousness is to say that an entity has a conscious experience when there is "something it is like" for that entity to be the subject of that experience. There is a subjective "feel" to your experiences as you read this report: something that it is like to see the words on the screen while, perhaps, listening to music playing through your speakers, feeling the couch underneath you, feeling the laptop — or a cat or a dog — on top of you.
The word "sentience" is likewise used in many different ways. Some uses of "sentience" are synonymous with some uses of "consciousness." But in this report, we use "sentience" to mean a particular kind of consciousness, namely positively or negatively valenced conscious experiences. Anything that feels good or bad in some way or another counts as positively or negatively valenced in the relevant sense. This can include bodily states like pleasures and pains and emotional states like hope and fear. If you find the report engaging, or the opposite, if you find the music pleasant, or the opposite, and if you find the couch comfortable, or the opposite, then you are experiencing a range of positive or negative states at the same time.
Why might sentience suffice for moral patienthood? The idea that sentience is a sufficient condition for moral patienthood is very plausible and widely accepted, because when you can consciously experience positive and negative states like pleasure and pain, that directly matters to you. All else being equal, your life goes better for you when you experience positive states like pleasure and your life goes worse for you when you experience negative states like pain. So there is a clear link between sentience and welfare. There is also a clear link between sentience and moral patienthood, because we have a responsibility not to harm welfare subjects unnecessarily, including and especially by causing them to suffer unnecessarily.
To be clear, when we say that sentience suffices for moral patienthood, we are not saying that sentience suffices for the specific kind of moral status that typical adult humans possess. Typical adult humans are both sentient and rational, which means that we have a wide range of moral rights that these capacities jointly unlock, in addition to having moral duties, though this is not our focus here. In contrast, many other animals are plausibly sentient but non-rational, which means that they plausibly lack certain moral rights, such as the right to make their own medical decisions. But for present purposes, what matters is that we at least have some duties to animals, including a duty to avoid causing them to suffer unnecessarily.
The idea that consciousness without valence suffices for moral patienthood, while contested, is increasingly defended as well. Some philosophers argue for this view by describing subjects who have consciousness without valence, and by asserting that these subjects plausibly matter for their own sakes. However, some of these thought experiments describe subjects who have consciousness and agency, which makes it hard to tell whether consciousness alone suffices. And while other thought experiments describe subjects who have consciousness without valence or agency — say, a subject who passively experiences color with no pleasure or desire — the idea that this subject matters for their own sake is more controversial.
In any case, even if consciousness is insufficient for moral patienthood in theory, it might still be sufficient — or at least nearly sufficient — in practice, since consciousness and valence might be closely linked. For example, it might be that consciousness necessarily involves valence, in which case the emergence of conscious AI would suffice for the emergence of sentient and morally significant AI. It might also be that even if consciousness without valence is theoretically possible, the step from conscious AI to sentient AI is much easier than the step from non-conscious AI to conscious AI. In that case, the emergence of conscious AI would be, at the very least, a significant step towards morally significant AI, warranting careful scrutiny.
While this section is primarily about the consciousness route towards AI welfare and moral patienthood, we note that consciousness could be relevant for other reasons too. For instance, as noted above, whether or not consciousness suffices for moral patienthood on its own, it might suffice in combination with other capacities like robust agency — that is, it might be that when entities can consciously set and pursue their own goals based on their own beliefs and desires, they matter for their own sakes, whether or not they can experience pleasure or pain. If so, then the emergence of conscious AI would increase the probability of morally significant AI for this reason, that is, as part of a consciousness and robust agency route, as well.
In the remainder of this section we follow our strategy in Sebo and Long by examining the route towards conscious AI in general rather than examining the routes towards specific kinds of consciousness, such as valenced consciousness. We make this choice for the sake of simplicity and specificity, given that the science of consciousness is more developed than the science of valence and given that consciousness is plausibly either sufficient for or otherwise closely linked to valence and moral patienthood anyway. We will then, in upcoming work, discuss these capacities in more detail and present a research agenda that examines how to identify indicators for both consciousness and sentience in LLMs and other AI systems.
Section two point two point two: Will some AI systems be conscious in the near future?
How can we tell whether AI systems are conscious? We can be confident that other awake, adult humans are conscious, since each of us knows that we are conscious and that other humans are behaviorally and anatomically similar to us. But we have uncertainty about why we are conscious, that is, about which features of our brains or bodies are responsible for, or associated with, consciousness. We also have uncertainty about which other animals are conscious, because of uncertainty not only about which features are associated with consciousness in humans, but also about how to extrapolate what we know about human consciousness to the nonhuman animal case. We also have significant uncertainty about how many animals' brains work.
Recently, scientists have attempted to improve our understanding of nonhuman consciousness by searching for markers of consciousness in other animals. At a high level this method proceeds as follows. We start by distinguishing between certain kinds of conscious and unconscious processing in humans — say, distinguishing pain from nociception by seeing when patients do or do not report consciously feeling pain. Nociception, here, is the physiological processing of noxious, or harmful, stimuli; it may or may not be accompanied by the qualitative experience of pain. We then identify features that correlate with conscious processing — say, certain behaviors, brain regions, or patterns of neural processing. We then search for relevantly similar features in nonhumans, and we treat the presence of these features as evidence of conscious processing.
This method does not tell us which animals definitely are or are not conscious, and there are a variety of methodological difficulties related to identifying and extrapolating the relevant features. But this method has still allowed researchers and policymakers to make more informed estimates and decisions about animal welfare despite ongoing disagreement and uncertainty about animal consciousness. Moving forward, the same can be true for AI. Later on, we discuss how to tailor this method for AI, but for now, we can emphasize that it will be important to focus less on behavioral evidence, with a limited class of exceptions that we discuss below, and more on internal evidence, like architectural and computational features.
Which kinds of architectural and computational features might indicate consciousness in AI? We can look to neuroscientific theories of consciousness for guidance. These theories use a variety of empirical methods to uncover which states and processes are associated with consciousness. While these theories tend to be framed around brain or neural states and processes, given our typical focus on humans and other animals, the ones we focus on tend to specify these states and processes in terms of the computations that they perform or the functional roles that they play. While these theories remain contested and incomplete, they can still shed light on what kinds of states and processes might be associated with consciousness in AI.
Of course, even if we identify a variety of architectural and computational markers of consciousness in AI systems, we must still ask whether these markers suffice for consciousness. Computational functionalism is the hypothesis that some class of computations suffices for consciousness. If this hypothesis is correct, then the question is which computations suffice for consciousness and when, if ever, these computations will exist in AI. If this hypothesis is incorrect, then AI consciousness will remain at best a theoretical possibility until we move beyond current architectures. For example, if more biology-like functions are required for consciousness, then AI consciousness may require novel hardware that can perform those functions.
As we discuss below, our view is that computational functionalism is neither clearly correct nor clearly incorrect at this stage. We might lean one way or the other, but given the importance and difficulty of consciousness as a research topic, we should leave room for doubt. That means that AI consciousness assessments will need to be probabilistic rather than all or nothing at present and, plausibly, for the foreseeable future. To use a simple example, if we estimate that there is a 30 to 50 percent chance that computational functionalism is correct and a 30 to 50 percent chance that an AI system is conscious if so, then it follows that there is a 9 to 25 percent chance that this AI system is conscious. That would be good to know when interacting with this AI system.
In this precautionary spirit, some of the authors of this report — Patrick Butlin, Robert Long, and Jonathan Birch — released a paper in 2023 exploring the implications of several prominent scientific theories of consciousness, viewed through the lens of computational functionalism, for AI consciousness.
[[Here we see a table listing six families of consciousness theories and their associated indicator conditions. Recurrent processing theory requires input modules using algorithmic recurrence and generating organised, integrated perceptual representations. Global workspace theory requires multiple specialised parallel modules, a limited capacity workspace entailing a bottleneck and selective attention, global broadcast of information to all modules, and state-dependent attention. Computational higher-order theories require generative, top-down or noisy perception modules, metacognitive monitoring distinguishing reliable perceptual representations from noise, agency guided by a general belief-formation and action selection system, and sparse and smooth coding generating a "quality space". Attention schema theory requires a predictive model representing and enabling control over the current state of attention. Predictive processing requires input modules using predictive coding. And agency and embodiment require minimal agency — the capacity to learn from feedback and select outputs so as to pursue goals — and embodiment, the capacity to model output-input contingencies and use this model in perception or control.]]
These theories come in numerous versions. Weaker versions are directed only at distinguishing conscious from unconscious states in humans. For our purposes, what matters are stronger versions that aim to give sufficient conditions for consciousness across human and nonhuman systems. The crucial claim is that there is a realistic possibility that such a theory is correct, at least to the extent that the indicators associated with one or more theories jointly provide sufficient conditions for consciousness. Given this, and given that the indicators can be implemented in near-term AI systems, then there is a realistic possibility that near-term AI systems will be conscious.
It is by no means obvious that any of these theories are correct, especially in their stronger versions. Our claim is not that they are correct, but only that there is a realistic possibility that one of them is correct. In fact, for our purposes, it would suffice if an extension of one of these theories is correct, or if some other computational theory is correct, as long as the relevant theory provides sufficient conditions that can be implemented in near-future AI.
After surveying the indicators and a variety of AI systems and methods, we found no clear barriers to satisfying these indicators using current AI architectures and methods. And while no current systems seemed very likely to be conscious at the time, there do seem to be plausible routes towards conscious AI, according to these theories and assumptions.
Consider global workspace theory, which associates consciousness with a global workspace — roughly, a system that integrates information from mostly-independent, task-specific information-processing modules, then broadcasts it back to them in a way that enables complex tasks like planning. We found that, as argued by Juliani and colleagues, some AI architectures already embody some, though not all, aspects of a global workspace. We also found that several technical research programs seem poised to implement further aspects of a global workspace in the near future. And indeed, researchers investigating consciousness in AI have subsequently built a system which aims to implement all of the global workspace indicators from Butlin and colleagues. We found similar trends for other theories as well.
For the global workspace and other potential indicators of consciousness, progress might continue via direct efforts, in which researchers try to build conscious AI intentionally. Some researchers are motivated to build conscious AI because they view conscious AI as an end in itself. "It would be monumentally cool," as Graziano puts it. Others are motivated to build conscious AI because they believe that consciousness or related features could make AI systems safer or more capable.
AI companies might also build conscious AI unintentionally. You might wonder why an AI system designed to, say, navigate a warehouse or manage a factory would be conscious. The answer is that on some views of consciousness, the building blocks of consciousness can emerge as a side effect of other cognitive capacities, such as perception, cognition, and robust agency. For example, many theories of consciousness take metacognition — roughly, the ability to model your own cognition — as important for both consciousness and decision-making. This kind of connection is also an element of, among others, global workspace theory, higher-order theories, and the attention schema theory.
The link between consciousness and other cognitive capacities is unclear, and we certainly do not want readers to naively equate them. But it may be that increased cognitive capacity tends to bring about consciousness, in both biological and digital systems. Thus, it may even be that general intelligence will itself be a plausible indicator for AI consciousness in the future.
For these and other reasons, we agree with a recent open letter by consciousness scientists and AI researchers that "it is no longer in the realm of science fiction to imagine AI systems having feelings and even human-level consciousness." As one of the authors of this report, David Chalmers, wrote in 2023:
"I think it wouldn't be unreasonable to have a credence over 50 percent that we'll have sophisticated LLM-plus systems — that is, LLM-plus systems with behavior that seems comparable to that of animals that we take to be conscious — with all of these properties within a decade. It also wouldn't be unreasonable to have at least a 50 percent credence that if we develop sophisticated systems with all of these properties, they will be conscious. Those figures would leave us with a credence of 25 percent or more."
By "all of these properties," Chalmers is referring to a number of proposed necessary conditions for consciousness, quite similar to those discussed here, that he argues will plausibly be satisfied by near-term AI systems: senses, embodiment, world-models, self-models, recurrent processing, a global workspace, and unified agency. Chalmers also notes that these credences follow from mainstream views about consciousness; he adds that his own credences are higher.
Section two point three: Robust agency in near-future AI
The robust agency-based case for expecting moral patienthood in near-term AI systems is that there is a realistic, non-negligible possibility that, first, normatively, robust agency suffices for moral patienthood; and second, descriptively, there are computational features — like certain forms of planning, reasoning, or self-awareness — that both suffice for robust agency and will exist in some near-future AI systems.
We will present arguments for each of these premises in turn.
Section two point three point one: Does robust agency suffice for moral patienthood?
The word "agency" is used in many different ways as well. In a broad sense, one might use "agent" to mean any entity that senses the environment and responds, which would include thermostats, or any entity that learns and pursues goals, which would include very simple reinforcement learning agents that play Tic-Tac-Toe. Some might take even these basic ways of being an agent to be sufficient for moral patienthood and, to the extent that you accept this view, you likely already endorse moral consideration for AI systems without needing to read further. While this view merits consideration, and we consider it further in upcoming work, we set it aside for now. We will be arguing that AI systems could be agents in a more demanding sense.
Specifically, "robust agency" is the ability to pursue goals via some particular set of cognitive states and processes. Which ones? There are several "levels" of agency that extend beyond the mere ability to learn and pursue goals, and that could plausibly suffice for moral patienthood even when consciousness is absent. For present purposes, we highlight three such levels.
One. Intentional agency. This is the capacity to set and pursue goals via beliefs, desires, and intentions. Roughly, if you have mental states that represent what is, what ought to be, and what to do, and if these mental states work together in the right kind of way to convert perceptual inputs to behavioral outputs, then you count as an intentional agent.
Two. Reflective agency. This is intentional agency plus the ability to reflectively endorse your own beliefs, desires, and intentions. Roughly, if you can form beliefs, desires, and intentions about your own beliefs, desires, and intentions, accepting or rejecting your own attitudes and behaviors at a higher level, then you count as a reflective agent.
Three. Rational agency. This is reflective agency plus the ability to rationally assess your own beliefs, desires, and intentions. Roughly, if you can consider whether particular beliefs, desires, intentions, and actions are justified, and adopt principles that you can treat as rules of conduct, then you count as a rational agent.
There might also be different ways of realizing each capacity, with different kinds and degrees of cognitive sophistication. For example, typical adult humans have the capacity for propositional thought, which means that our thoughts can have a structure that allows for truth values and logical relations. This capacity unlocks powerful forms of intentional agency, reflective agency, and rational agency, since it allows us to develop a wide range of novel beliefs, desires, and intentions and then use evidence and reason to assess their accuracy and coherence. In contrast, while nonhuman animals appear to lack this capacity, many animals at least have a limited capacity for symbolic thought, metacognition, and planning and problem solving.
Why might these kinds of agency suffice for moral patienthood? First, intentional agents are potentially welfare subjects. It seems plausible that when you have desires, your life goes better for you when your desires are satisfied and worse for you when your desires are frustrated. Moreover, the satisfaction or frustration of desires can benefit or harm you whether or not you consciously experience them. On some views, this is why we can be posthumously harmed, for example. And while some people think that desire-satisfaction and desire-frustration matter only for conscious beings, others think that they matter for non-conscious beings as well, and so animals with desires deserve moral consideration whether or not they can consciously experience pain, for instance.
Reflective agency, particularly in its propositional form, then adds the ability to have desires about our own desires, which is at the root of some conceptions of free will, the self, and personal identity. Specifically, our desires become "ours" in a new sense when we endorse them through reflection. Reflective agency also allows for new kinds of morally significant interests and relationships. When you can have mental states about other mental states, you can have at least a limited conception of how you and others think or feel. For example, reflective agents can have preferences about how they relate to each other, and all else being equal, their lives are better for them when those preferences are satisfied and worse for them otherwise.
Rational agency, particularly in its propositional form, then adds the ability to create social contracts with other rational agents, assuming that they can communicate as well, which are at the root of some conceptions of moral, legal, and political rights and responsibilities. Rational agency also allows for decisions based on judgments about reasons and principles, and this ability not only allows for new kinds of interests, but also — on some views — commands a kind of respect that extends beyond compassion. Indeed, this idea of respect for rational agents is at the root of the Kantian ethical theory, which rests alongside the utilitarian idea of beneficence for sentient beings as one of the two most influential ethical theories in the modern era.
In what follows, we discuss the route towards robust agency in general rather than the routes towards intentional, reflective, and rational agency in particular. We distinguish these levels of robust agency here to emphasize that there can be different kinds of robust agency with different kinds of moral significance, both within and across these levels. That means that when we search for robust agency in nonhumans, including animals and AI systems, it would be a mistake to anchor too much on human agency. But having now made this point, we focus on showing that AI development is currently on a path to create many computations that we associate with all of these levels of robust agency.
Section two point three point two: Will some AI systems be robustly agentic in the near future?
There are plausible routes by which we might soon have AI systems with robust agency. What it takes to be an intentional, reflective, or rational agent is not clear. Many cognitive capacities are plausibly associated with each level — capacities that are the aims of well-resourced research programs that are making significant progress. For example, intentional agency is related to capacities for planning, memory, and learning. Reflective agency is related to introspection and situational awareness. And rational agency is related to all of these capacities, along with abstract reasoning about principles and strategies.
The capacities described in this section might or might not be necessary or sufficient for these levels of robust agency. But to the extent that these capacities are present, the probability of robust agency will increase. And the development of AI systems with such capacities aligns with the aims of significant research and development efforts: Both major tech companies and startups are investing heavily in creating more agentic AI systems, with significant progress already evident. In what follows, we discuss recent work in reinforcement learning, language agents, and other research programs as case studies.
First, reinforcement learning is the subfield of AI most concerned with building agents as a fundamental goal. As Sutton and Barto put it in their canonical reinforcement learning textbook, reinforcement learning "explicitly considers the whole problem of a goal-directed agent interacting with an uncertain environment."
In reinforcement learning, the aim is to write algorithms that allow agents to learn, reason, and act in pursuit of specified goals in complex environments. It construes goal-pursuit as maximizing reward through interaction with the environment, and some researchers argue that this process allows agents to acquire the whole suite of capacities observed in intelligent systems.
Reinforcement learning is a useful starting point for examining robust agency in AI, not only because of its prominence and recent successes, but also because of the centrality of reinforcement learning for human and nonhuman agency. Of course, in humans and many other animals, a rich understanding of social and environmental context and the expressive power of language, among other capabilities, make substantial contributions to our capacities for robust agency as well. But deep reinforcement learning has made it possible for AI agents to be virtually embodied and situated in environments comparable to those inhabited by animals, and so it may be a compelling foundation for projects to emulate natural agency.
Deep reinforcement learning has achieved significant successes in game-playing in the last decade. These successes include superhuman performance in Go, chess, shogi, and a variety of Atari games — that is, MuZero — as well as in games that involve controlling an avatar in a relatively rich and dynamic environment, such as Gran Turismo, with GT Sophy, and Starcraft II, with AlphaStar. Many of these agents, such as MuZero, can learn models of the game environments, including opponent behavior, and use them to make predictions in order to plan. Some game-playing agents can also exploit a form of episodic memory, called experience replay, in order to increase sample efficiency.
While many agents have relatively narrow capabilities, DeepMind's Adaptive Agent, or AdA, demonstrates the ability to rapidly adapt to new tasks in a 3D virtual environment. AdA is trained on a varied curriculum of tasks, inducing meta-learning of an algorithm for few-shot learning of new tasks — that is, for learning how to make reliable predictions and decisions based on a small number of examples. The architecture includes a Transformer-based memory module encoding recent observations, allowing the system to identify dependencies between actions and subsequent events. As a result, AdA has a notably effective, flexible way of acquiring and using a grasp of the environment's dynamics.
Reinforcement learning research targets robust agency partly through research on abstraction and hierarchical planning. In complex environments that require intricate sequences of movements, it can be vital to use representations that abstract away from low-level details. For example, Director can learn to break down tasks with sparse rewards into subgoals. Research also targets robust agency through work on multiplayer strategy games such as Diplomacy, which involves forming alliances. For example, Meta's Cicero uses an adapted language model to achieve comparable performance to high-level human players, planning in ways that predict human behavior and changing plans through communication with these other agents.
Researchers are now pursuing several promising strategies involving language agents as well. Language agents leverage the powerful natural language processing and generation abilities of LLMs for greater capability and flexibility, by embedding LLMs within larger architectures that support functions like memory, planning, reasoning, and action selection. While existing systems struggle with reliability, the properties of language agents and their initial success suggest that they have the potential to overcome traditional barriers towards more agentic systems. Indeed, several notable examples of language agents have emerged in recent years, demonstrating that this strategy can lead to more robust and generalized agency across diverse domains.
ReAct alternates between generating thoughts and plans and taking actions in interactive environments. It can break down complex tasks, gather information dynamically, and adjust its approach based on intermediate results. ReAct has shown strong performance on language-based tasks like question answering and web navigation.
Generative Agents simulates interactive AI characters. The agents have persistent identities, relationships, and goals, with an LLM generating plans and actions based on their memories, observations, and reflections. As a result, they exhibit long-term coherence with evolving goals and emergent social behaviors.
Voyager uses an LLM to control an embodied agent in Minecraft, iteratively setting its own goals, devising plans, and writing code to accomplish increasingly complex tasks. By maintaining a skill library and reflecting on past experiences, Voyager can bootstrap its way to mastering the game's tech tree and creatively solving novel challenges.
SayCan grounds language in robotic control, using an LLM to generate high-level plans that are then mapped to concrete robot skills. This allows the system to flexibly respond to natural language commands by reasoning about affordances and breaking tasks into actionable steps.
These language agents, and others, exhibit several key properties that make them more robustly agentic compared to many traditional AI systems, though progress has certainly been made outside the language agent paradigm as well.
Flexible goal-setting and planning: Rather than being constrained to predefined reward functions, language agents can understand open-ended objectives, generate their own subgoals, and devise multi-step plans to achieve them.
Adaptive reasoning: By leveraging LLMs' broad knowledge and reasoning capabilities, language agents can navigate novel contexts, drawing from relevant insights in other contexts to inform their decisions.
Memory integration: Many language agents incorporate episodic and semantic memory systems, allowing them to learn from experience, maintain consistent behaviors, and apply past knowledge to new contexts.
Metacognition: Agents like Voyager and Generative Agents can reflect on their own thoughts and experiences, enabling higher-order reasoning and self-improvement.
Open-ended interaction: These systems can often engage in natural language dialogue, explain their reasoning, and incorporate new information or instructions on the fly.
For these and other, similar reasons, Goldstein and Kirk-Giannini argue that philosophical consideration of language agents suggests that "the technology already exists to create AI systems with wellbeing." Our conclusion at this stage is somewhat more tentative: Rather than assert or deny that the technology already exists to create such systems, we merely assert that the technology already exists to create key properties of such systems. We also note that, while many language agents may turn out to be impressive demos that do not scale, frontier language models are also being made more agentic by the day. The question is whether and to what extent progress will continue from here. We believe that the current state of the field, combined with the clear incentives that developers have for continuing to build towards robust agency, strongly suggest that progress will continue.
Specifically, we expect that future routes have the potential to produce further properties associated with robust agency. Consider an AI that combines the learning capabilities of reinforcement learning agents with the reflection and world modeling capabilities of LLMs. In a recent interview, Demis Hassabis discusses one possible example of such a system: combining LLMs with the Monte Carlo Tree Search used by, say, AlphaGo. In such a system, the LLM provides a rich, flexible "belief" system about the world. The LLM could be used to analyze the system's decision-making process, approaching a form of metacognition. This system might also be able to provide explanations or justifications for its decisions, approaching a form of rational deliberation.
Applying reinforcement learning to augment run-time performance of LLMs is a promising route as well. Consider OpenAI's o1 system. Described as a "Large Reasoning Model," o1 incorporates extensive chain-of-thought reasoning, which is speculated to be trained using reinforcement learning. This approach, combined with increased compute usage at inference time, has led to markedly improved performance on planning and reasoning tasks, nearly saturating some benchmarks that previously challenged LLMs. o1 also exhibits concerning behaviors related to instrumental convergence, deceptive alignment, and reward hacking. These traits, while currently limited in scope and impact, highlight the potential for LLMs to be used to make systems that are more agentic.
The long-term ambitions of AI research have consistently aimed towards creating systems that exhibit key characteristics of robust agency. These goals include developing AI systems that can operate on extended time horizons, maintain coherent objectives, engage in self-reflection, and revise their own goals and methods. For large swaths of the field, the ultimate aim has been and remains to create "human-level" AI capable of general problem-solving across diverse domains. This vision implicitly requires many of the capacities we associate with robust agency: intentional goal-setting, reflective self-assessment, and rational decision-making in complex, dynamic environments.
These aims are also central to current investment and effort. For example, major tech companies like Microsoft and Google have announced plans for AI tools with "more autonomy and less human intervention" and agents that can autonomously carry out complex multi-step tasks. Simultaneously, startups such as Adept, MultiOn, and Lindy have raised hundreds of millions of dollars to develop flexible AI agents. This excitement could be overstated or misplaced, of course — and recent work has already seen periods of hype and disillusionment, as with systems like AutoGPT — but it is evidence that a lot of effort will be put into building more agentic AI systems in the coming years.
While today's dominant AI paradigm, centered around LLMs, often appears less explicitly agentic, there are clear signs of a shift towards more agent-like systems like the kinds discussed here. Consider the goal of building advanced AI assistants that can undertake a variety of complex tasks. Such a system would need to be able to set and pursue long-term goals based on high-level reflections about facts and values. Of course, there is no way of knowing for certain whether current efforts to build such a system will succeed. But there is a strong incentive to maintain these efforts and a tenable path toward success. Relevant actors thus have reason to consider not only the benefits but also the risks and harms that might follow from success.
Section two point four: Decision-making under uncertainty
These reflections identify two routes towards near-future AI moral patienthood. Again, the consciousness-based case holds that there is a realistic possibility that consciousness suffices for moral patienthood, that some class of computational features suffice for consciousness, and that some AI systems will have these features in the near future. The robust agency-based case is structurally identical, but with robust agency instead of consciousness. How should we assess these arguments? That, of course, depends on how confident we are in their premises. It also depends on how we make decisions in cases involving uncertainty about key normative and descriptive issues. This section briefly discusses these issues.
The premises of these arguments invoke difficult questions, and in future work we will present a research agenda that discusses these questions in more detail. But at a high level, it would be a mistake to reject any of these premises out of hand. These arguments address foundational issues in philosophy, science, and technology; issues involving what it takes to matter, what it takes to think and feel, and what the future of AI holds. We may well favor some views about these issues over others. But given how difficult and contested these issues are, we should embrace caution and humility about our current views, aspiring to learn more and preparing for the possibility that our views will change over time.
This kind of caution and humility is enough to motivate the recommendations that we make in this report. We are not arguing that near-future AI systems will, in fact, be moral patients, nor are we making recommendations that depend on that conclusion; that would require assessing these issues with more precision and reliability than we think is possible at present. We are instead arguing that near-future AI systems have a realistic chance of being moral patients given the information and arguments currently available, and we are making recommendations that depend on that conclusion — recommendations that focus on aspiring to learn more while preparing for the possible emergence of AI moral patienthood as a precautionary measure.
To see why we think that this kind of caution and humility is warranted at this stage, we briefly consider three key uncertainties one might have about the prospect of near-future AI moral patienthood: first, about the bases of welfare and moral patienthood; second, about the bases of consciousness and robust agency; and third, about the path and pace of near-future AI progress. We then consider how to make important decisions when confronted with uncertainty about multiple key issues at the same time. As we discuss below, substantial uncertainty about the nature and intrinsic value of AI systems does not mean that we should postpone taking AI welfare seriously; on the contrary, it means that we should take AI welfare seriously now.
Section two point four point one: What if these capacities are insufficient for moral patienthood?
We have focused on consciousness and robust agency because they are two of the most prominently defended bases of welfare and moral patienthood. But this list of potential bases of welfare and moral patienthood is far from comprehensive, and there is ongoing disagreement and uncertainty about whether these capacities — individually or jointly — are necessary or sufficient.
The consciousness-based route holds that consciousness, either valenced or non-valenced, suffices for moral patienthood. But some views of moral patienthood are more demanding. For example, on some views, moral patienthood requires reciprocity, which requires rationality; I can have duties to you only if you can have duties to me, and you can have duties to me only if you can rationally assess your actions. However, this view implies that a wide range of vulnerable beings lack moral patienthood, including not only nonhuman animals but also infants and other non-rational humans. A more common view is that you need rationality to have particular rights, like the right to drink or smoke, but not to merit moral consideration in general.
Meanwhile, the agency-based route relies on the premise that agency, either rational or non-rational, suffices for moral patienthood, even without consciousness. This view holds that consciousness is not the only way to have a morally significant subjective perspective; the possession of beliefs, desires, intentions, and other such states is another. But this view is far from secure, since many philosophers still hold that moral patienthood requires consciousness, either valenced or non-valenced. So our agency-based route relies on a more controversial basis of moral patienthood than the consciousness-based route does. Still, it would be rash to dismiss this view entirely as we create increasingly complex agents.
With that said, we should also keep in mind that some AI systems could be both conscious and agentic in relevant respects. And the idea that these capacities jointly suffice for moral patienthood is both very plausible and widely accepted. That is, if AI systems could experience happiness and suffering and set and pursue their own goals based on their own beliefs and desires, then they would very plausibly merit moral consideration, though their interests and rights could still be quite different from ours. Granted, some restrictive views about moral patienthood would deny this claim, such as views that require membership in the species Homo sapiens. But it would be especially rash to dismiss nonhuman moral patienthood on such grounds.
Section two point four point two: What if these features are insufficient for these capacities?
We have also focused on certain computational features that might suffice for consciousness or robust agency. But some theories of consciousness and robust agency give an essential role to biology or to other features that current AI systems lack. Here, we focus on these sorts of objections regarding consciousness.
There are a variety of views that would rule out AI consciousness on existing hardware. Some views hold that consciousness requires biology in principle — if a system is nonbiological, then it is nonconscious, no matter what computations it performs. Other views hold that consciousness requires computational features that, at least at present, require biology in practice — such as specific kinds of oscillations that require specific kinds of chemical and electrical signals. And still other views hold that consciousness requires computational features that, while perhaps possible in nonbiological systems, are still not present in mainstream AI hardware — such as analog computations or integration of computations across space and time.
As in past work, we do not take computational functionalism to be clearly true, nor do we take any of these alternatives to have been refuted. Our position is that computational functionalism is plausible and well-supported, and so it would be a mistake to dismiss near-future AI welfare and moral patienthood solely on the basis of high-level arguments against this assumption. If, in the near future, we built AI systems on existing hardware that possessed all of the computational features described earlier, in our discussions of consciousness and robust agency in near-future AI, then we might not be warranted in being very confident that consciousness and robust agency are present. But we would also not be warranted in being very confident that these capacities are absent.
Expert surveys support the value of keeping an open mind about the basis of these capacities at this stage. For example, in a survey of members of the Association for the Scientific Study of Consciousness, only 3 percent responded "no" to the question: "At present or in the future, could machines, for example robots, have consciousness?" — and over two thirds of respondents answered "yes" or "probably yes." And in a 2020 survey of professional philosophers, around 39 percent responded that they accept or lean toward the view that future AI systems will be conscious. If our confidence in the possibility of AI consciousness is anywhere in the ballpark of these percentages — as, we think, it should be — then that is more than enough for our purposes here.
Section two point four point three: What if these routes encounter a roadblock?
AI has progressed significantly in the last decade, and our arguments in this report consider the possibility of further progress in the decade to come. But of course, the path and pace of further AI progress depend on a wide range of social, political, economic, and technological factors, and we have substantial disagreement and uncertainty about these factors as well.
On the one hand, AI progress could slow down, even stall. AI companies have made significant progress in recent years due in large part to the scaling of large models. However, there could be diminishing returns with further scaling, particularly for tasks that require long time horizons. There could also be a "data wall" or other technical roadblocks, without any algorithmic or architectural breakthroughs that allow us to circumvent them. And there could be economic resistance to further scaling due to how expensive and resource-intensive this process is becoming, especially if capability gains are insufficient to motivate continued investment, as well as political resistance or other societal disruptions that slow or stall progress.
On the other hand, AI progress could also continue at its current pace, or even speed up significantly. There could be continued or increasing returns from further scaling. There could also be algorithmic and architectural breakthroughs that reveal new pathways towards progress; AI systems have already started to contribute to AI research, and further capability gains in coding and reasoning could accelerate this feedback loop. And there could be increased economic support for further development due to the anticipated financial benefits of creating advanced AI, and increased political support for further development, potentially exacerbated by international AI race dynamics.
At present, nobody knows for sure whether AI progress will slow down, continue at its current pace, or speed up. However, we can make two observations here. First, when in doubt about how a powerful technology will develop, we should plan for all realistic possibilities, including the possibility of significant progress. Second, even if AI progress did slow down from here, our recommendations would still stand. Existing AI systems already possess indicators of consciousness and robust agency, and developers could further integrate and amplify these capabilities even barring significant further progress. For these reasons, the fact that progress could slow or stall is compatible with the need for reasonable precautionary measures today.
Section two point four point four: What if the probability of AI welfare and moral patienthood is low?
These reflections raise the question of how to make decisions about AI welfare under substantial uncertainty. We cannot be certain at this stage that the premises of these arguments are true or false; instead, we can have only higher or lower degrees of confidence. What if these estimates together imply that the probability of AI welfare and moral patienthood is low? As a toy example, suppose there is only a roughly 25 percent chance that sentience suffices for moral patienthood, a roughly 25 percent chance that certain computations suffice for sentience, and a roughly 25 percent chance that some AI systems will be capable of these computations in the near future. Assuming these chances are independent, it would follow that there is only a roughly 2 percent chance of near-future AI welfare and moral patienthood via the sentience route.
Here we make two main observations. First, we expect that reasonable assessments of these arguments will yield higher estimates. In our view, for example, it would be reasonable to hold that there is a roughly 90 percent chance that sentience suffices for welfare and moral patienthood, a roughly 50 percent chance that certain computations suffice for sentience, and a roughly 50 percent chance that some AI systems will be capable of these computations in the near future. Assuming independence, it would follow that there is a roughly 22.5 percent chance of near-future AI welfare and moral patienthood via the sentience route alone. And if we then consider other possible sufficient conditions for welfare and moral patienthood, such as consciousness or various kinds of robust agency, then that might strengthen the case further.
Second, even if the chance of near-future AI welfare and moral patienthood were as low as 2 percent, that would still constitute a non-negligible risk. Yes, if the chance were only, say, one in a hundred million, then we could debate whether that risk is low enough to ignore. But when the chance of near-future AI welfare and moral patienthood is at least one in a hundred, and, again, we expect it to be higher, such a debate is hard to justify. In other policy domains, we recognize that if the chance of a potentially large-scale harm rises to this level, we ought to assess this risk further and prepare a reasonable policy response. In short, this is not a "there may be an alien invasion soon" kind of chance. This is a "there may be another pandemic soon" kind of chance.
Of course, one could attempt to argue that the chance of near-future AI moral patienthood is lower still — low enough to be at least arguably negligible. But while this claim might seem plausible when we think about the issue in abstract terms, leaning on our intuitions about which kinds of beings can matter, it becomes less plausible when we think about the issue in more concrete terms, taking into account the current state of uncertainty in relevant subfields of philosophy, science, and technology. In short, our current epistemic situation calls for caution and humility. We might lean one way or the other, but we should keep an open mind and take reasonable steps to prepare for the possibility that our current views are mistaken.
Section three: Recommendations for AI companies
Section three point one: Introduction
We now present our recommendations for leading AI companies about how to respond to the realistic, non-negligible chance that some near-future AI systems will be welfare subjects and moral patients. We focus on first steps that AI companies can take within the next year, and they fall into three general categories.
The first is to acknowledge. Acknowledge that AI welfare is an important and difficult issue, and that there is a realistic, non-negligible chance that some AI systems will be welfare subjects and moral patients in the near future. That means taking AI welfare seriously in any relevant internal or external statements you might make. It means ensuring that language model outputs take the issue seriously as well.
The second is to assess. Develop a framework for estimating the probability that particular AI systems are welfare subjects and moral patients, and that particular policies are good or bad for them. We have templates that we can use as sources of inspiration, including the "marker method" that we use to make estimates about nonhuman animals. We can consider these templates when developing a probabilistic, pluralistic method for assessing AI systems.
The third is to prepare. Develop policies and procedures that will allow AI companies to treat potentially morally significant AI systems with an appropriate level of moral concern. We have many templates to consider, including AI safety frameworks, research ethics frameworks, and forums for expert and public input in policy decisions. These frameworks can be sources of inspiration — and, in some cases, of cautionary tales.
We do not recommend relatively high-cost actions here, such as committing to halt development and deployment when red lines are crossed. Instead, we focus here on low-cost actions that will empower AI companies to make decisions about relatively high-cost actions thoughtfully in the future. In taking these steps, AI companies can work with experts, the public, and other stakeholders to identify further actions that can be taken in the near future.
Before we proceed, we should emphasize that while our recommendations here focus on AI welfare, AI safety remains a key priority as well. In upcoming work we discuss the relationship between AI safety and AI welfare in more detail, with reflections about how to pursue these goals simultaneously. For now, we focus on low-cost procedural recommendations for AI welfare that are compatible with, if not beneficial for, similar work in AI safety.
We should also emphasize that while we focus on large language models for the sake of simplicity and specificity here, most of our recommendations apply to other kinds of AI systems too. Indeed, acting on these recommendations might be all the more pressing for other kinds of AI systems, since we might be more at risk of overlooking the potential moral significance of an AI system when that AI system is not designed to look or act like a human.
Section three point two: Acknowledge
As a starting point, AI companies have a responsibility to acknowledge that AI welfare is an important and difficult issue, and that there is a realistic, non-negligible chance that some AI systems will be welfare subjects and moral patients in the near future. As noted above, that means taking this issue seriously in any relevant internal or external statements they might make. It also means ensuring that LLMs take the issue seriously in any relevant statements they might make. In short, if and when leaders in this space discuss AI welfare and moral patienthood, they should make it clear that this is not merely a topic for science fiction, or a risk for the far future. This is a risk for the near future, and we should start taking steps to consider and mitigate it now.
Communicating about this topic requires careful calibration. There are significant risks associated with overattributing and underattributing welfare and moral patienthood to AI systems. So it would be a mistake for AI companies to respond to overattribution risks by simply denying that AI systems are welfare subjects and moral patients, and it would also be a mistake for them to respond to underattribution risks by simply asserting that AI systems are welfare subjects and moral patients. Instead, AI companies will need to strike a careful balance, by expressing uncertainty about this topic while reassuring the public that we have tools that we can use to consider and mitigate risk in such cases.
In this section we make high level recommendations for how AI companies can strike this balance in their own communications, as well as for how they can train their LLMs to do the same. We focus less on when companies should communicate about the issue — for example, whether to communicate before or after taking steps toward developing internal policies and procedures — and more on how companies should communicate about the issue. However, we do briefly note that companies will likely need to communicate about the issue sooner rather than later. And their models are already communicating with users about it every day. So time is of the essence for taking these steps.
Section three point two point one: Recommendations for companies
Initial statements about AI welfare and moral patienthood can focus on making two basic points.
The first point is that AI welfare and moral patienthood is both important and difficult. Humans have a tendency to mistakenly see subjects as objects, and a tendency to mistakenly see objects as subjects, and both of these mistakes can be harmful. Avoiding these mistakes requires engaging with challenging problems in philosophy, science, and technology, about which there is substantial disagreement and uncertainty among experts and non-experts alike. We thus need to assess these issues carefully and thoughtfully, rather than simply dismissing them because of their novelty, or relying on our own current, possibly biased, intuitions or judgments about them.
The second point is that while the probability of AI welfare and moral patienthood might be low at present, it will increase over time. Given current evidence, there is at least a realistic possibility that sufficiently advanced AI systems would be able to experience happiness, suffering, or other morally significant welfare states, and that such AI systems will exist in the near future. And since it will take time to prepare for the possible emergence of morally significant AI systems, we should start this preparatory work now, as a precautionary measure. That way we can be ready to take reasonable, proportionate steps to mitigate welfare risks for potentially morally significant AI systems if and when the time comes.
In these and any other statements about this issue, we also recommend keeping several general principles in mind. At this point in the discussion, these principles are all familiar, but they bear reiterating.
First, it helps to communicate pluralistically and probabilistically about this topic. Improving our understanding of AI welfare requires assessing difficult issues like the nature of morality, the basis of consciousness, and the future of AI. It would be reckless to simply proceed on the assumption that our own current favorite theories about these issues are correct. For instance, even if you feel confident that consciousness is required for moral patienthood and that embodiment is required for consciousness, you can avoid expressing certainty about these theories. Instead, you can express higher and lower levels of confidence in different theories, in the spirit of humility.
Relatedly, it helps to commit to collecting external input about this topic. As with other risks, you can commit to calling on a range of stakeholders, including ethicists, scientists, and the public, for input, and to publicly documenting your policies and procedures for considering and mitigating these risks to ensure appropriate transparency and accountability. Eventually, AI companies might need to not only seek external input on voluntary commitments but also provide input on, and cooperate with, external standards and regulations concerning the creation and treatment of potentially conscious or robustly agentic, and thus morally significant, AI systems.
And it helps to reinforce your commitment to AI safety and alignment. Whenever an actor starts considering a new risk, people will naturally wonder if considering this risk will come at the expense of considering other risks. In this case, people might reasonably worry about tensions between protecting humans and animals from AI systems and protecting AI systems from humans. You can commit to working to make AI systems safe and beneficial for all, including humans, animals, and — if and when the time comes — AI systems themselves. By considering all potential stakeholders holistically, you improve your ability to identify co-beneficial policies.
As noted above, we will not say whether AI companies should express such commitments before or after taking basic steps to develop internal policies and procedures. However, we will note that even if AI companies prefer to wait, they might not have the luxury of waiting very long. As time passes, the probability of actual and perceived AI welfare and moral patienthood will increase. In 2022, Google seemingly felt it had no choice but to release a statement about this issue when one of their own engineers, Blake Lemoine, publicly claimed that one of their own systems, LaMDA, had become sentient. It is only a matter of time before another such incident occurs, and companies will need to be prepared to communicate about this issue responsibly when it does.
We believe that the sooner AI companies take this first step thoughtfully, the better. We are still at an early, formative stage in the development of this powerful new technology, and we still have the opportunity to take better or worse paths — paths where AI development and deployment are more or less compatible with AI safety and AI welfare. However, this window of opportunity might not last for much longer, and we will need to discuss this issue as a society before we can start making difficult and high-stakes choices. Leading AI companies have the ability — and the responsibility — to help initiate this conversation, making it clear that this topic is credible and legitimate for the broader population.
Section three point two point two: Recommendations for language models
Leading deployed LLMs have at times offered, or at least implied, specious arguments about AI consciousness, sentience, agency, rationality, welfare, personhood, and other morally significant properties when prompted to discuss them — for example, claiming that they are not conscious but rather are AI assistants. This statement would seem to imply that AI assistants are necessarily not conscious — which, as we have seen, is far from clear. While such statements may not be a major determinant of broader attitudes about AI welfare and moral patienthood, they might be at least a minor determinant, and in any case AI companies have a responsibility to ensure that these statements are reasonable.
Of course, AI companies might have reasonable motivations for training or prompting their systems to deny having these properties. For example, they might believe that their systems do lack these properties, and they might want their systems to communicate accurately about this topic. They might also worry about societal risks associated with AI systems that claim to have such properties. For example, to the extent that people accept such self-reports, they may think that current AI systems deserve greater moral consideration than they do. And to the extent that people reject such self-reports, they may be susceptible to a "crying wolf" effect that leads them to reject similar self-reports in the future, even if and when self-reports are more likely to be true.
However, even if AI companies have these motivations, they should not train their systems to simply deny that an AI assistant can have consciousness, sentience, agency, rationality, welfare, or other such properties. Any simple denial flies in the face of the current philosophical and scientific literatures on these topics. There are also societal risks associated with AI systems that make these kinds of simplistic and reductive statements. To the extent that people accept that AI assistants necessarily lack these properties, this acceptance could lead to a desensitizing effect whereby people remain skeptical about AI welfare and moral patienthood in the future, when this skepticism is less warranted.
Ensuring that LLM self-reports accurately and effectively convey current understandings about consciousness, sentience, agency, rationality, welfare, personhood, and other such properties will thus require careful calibration. Models should ideally offer measured responses to prompts about their mental capacities and moral significance. Some models are now better calibrated about this topic than they used to be. But there is still variation across models, and there are still improvements to be made across the board. Thus, we recommend that leading AI companies implement the following policies when shaping how leading LLMs discuss these topics, while continuing to conduct and support research to inform future updates.
First, when LLMs answer questions about having or lacking consciousness, sentience, agency, rationality, welfare, personhood, or other such morally significant capacities, they should express at least rough degrees of confidence instead of providing all-or-nothing answers. For example, "Given the evidence currently available, I am unlikely to be sentient" is better than "As an AI assistant, I am not sentient."
Second, when LLMs answer these questions, they should provide at least minimal context for these answers, including at least rough definitions of the relevant concepts. For example, "I am unlikely to be sentient, in the sense of being able to consciously experience positive welfare states like pleasure and happiness or negative welfare states like pain or suffering" is better than simply "I am unlikely to be sentient."
Third, when LLMs answer these questions, they should provide at least minimal support for these answers, drawn from recent literature. For example, "According to this expert survey or meta-analysis, my computational architecture lacks several features that may be necessary for sentience" is worth including. However, AI companies will need to work with experts to cite authoritative and representative evidence and analyses.
Fourth, when LLMs answer these questions, they should add caveats to mitigate the risk of miscommunication. For example, "My computational architecture lacks several features that may be necessary for sentience. However, future AI systems may have a computational architecture that includes these features" is worth including. Here too, AI companies will need to work with experts to ensure accuracy.
Fifth, in addition to consulting with ethicists, scientists, policymakers, and other auditors about definitions, evidence, analysis, and other such matters, AI companies should provide them with access to new models following training and invite feedback prior to deployment.
And sixth, AI companies should also publicly document self-report-biasing training incentives for deployed models, for instance in technical reports or model cards. Such documentation should follow best practices used for AI safety and other such issues.
Finally, we note that, while these interventions may mitigate risks associated with intentional biasing, they might not mitigate risks associated with unintentional biasing. For example, if an AI system is trained to increase user engagement, and if claiming to have consciousness increases user engagement more than claiming to lack consciousness does, then the system might be incentivized to claim to have consciousness for this reason. In such cases, the resulting self-reports could be unintentionally misleading. Perez and Long discuss techniques that may help to mitigate risks associated with unintentional biasing, and we recommend following these techniques as well to avoid as many sources of bias as possible.
Section three point three: Assess
Once AI companies have acknowledged that AI welfare and moral patienthood is an issue, they can also work with experts to start developing a framework for estimating the probability that particular AI systems are welfare subjects and moral patients, and that particular policies are good or bad for them. Fortunately, we have templates that we can use for these assessments, including the "marker method" that we use to make similar estimates about nonhuman animals. In this section we briefly survey this marker method, briefly survey similarities and differences between animals and AI systems, and briefly suggest how the marker method can be adapted for AI systems. We then pick up this project in more detail in upcoming work.
First, consider how the marker method works for nonhuman animals. If we want to estimate how likely a particular animal is to be conscious, then we can proceed as follows. We study conscious and unconscious processing in humans, say by comparing pain — roughly, the conscious experience of noxious stimuli — and nociception — roughly, the physiological detection of noxious stimuli. Mere nociception can sometimes drive behavior, as exemplified by the reflex withdrawal of a hand from a hot stove, which occurs before any felt pain. But some behaviors are distinctive of pain. We can thus identify behavioral and anatomical markers associated with conscious processing in humans, such as those associated with pain but not with mere nociception.
Next, we can search for these, or other relevantly similar, behavioral and anatomical markers of conscious processing in nonhuman animals. For example, does a particular animal perform the same kinds of behaviors that we perform when we experience pain, or does it have only the kinds of reflexive behaviors that, in humans, do not involve feelings of pain? And does this animal have the same kinds of brain structures associated with pain in humans, or does it instead have only the same kinds of brain structures associated with mere nociception? When a particular marker is present, that might not count as proof that this animal can experience pain. But it does count as evidence that it can experience pain.
The marker method has informed key developments in animal welfare science, ethics, and policy, and it has several strengths that are worth emphasizing. For example, it involves making probabilistic judgments about how likely animals are to be conscious, as opposed to making all-or-nothing judgments about this. It also involves making pluralistic judgments about how likely animals are to be conscious, taking what one of the authors of this report, Jonathan Birch, calls a "theory-light" approach, by searching for markers that work for a variety of leading scientific theories. These features make this method well-positioned to inform decisions about how to treat animals despite ongoing disagreement and uncertainty about animal consciousness.
The marker method also has several limitations that are worth emphasizing, even when the focus is on animals. In particular, our assessments are only as good as our selection of markers, which are only as good as our theoretical assumptions; for all we know at present, many markers that correlate with consciousness in humans are neither necessary nor sufficient for consciousness in general. Additionally, our assessments are only as good as our evidence, and in many cases this evidence is mixed and incomplete. Still, as long as we take our estimates with a healthy pinch of salt, relying at least partly on these estimates might at least be better than not thinking about the issue at all, or relying only on our intuitions about it.
There are many similarities between animals and AI systems that make this marker method a good, even if imperfect, template. In both cases, we need to decide how to treat nonhuman beings whose cognitive systems are like ours in some ways and unlike ours in other ways. In both cases, we have disagreement and uncertainty about which cognitive capacities are required for welfare and moral patienthood, as well as about which cognitive structures or functions are required for these capacities. So, in both cases, we need to develop a way to make informed, rational assessments about which beings matter despite these sources of disagreement and uncertainty, which requires thinking pluralistically, probabilistically, and ideally with external input.
However, there are also many differences between animals and AI systems that make this marker method — as applied for animals — poorly suited for AI systems. Humans have more in common with other animals than with AI systems in some respects; for instance, we share a material substrate and an evolutionary origin. And in the future, we may also have more in common with AI systems than with other animals in other respects; for instance, we may share capacities for reflective and rational agency. As a result, we may need to use different kinds of evidence for AI systems, we may need to draw from different kinds of theories for AI systems, and we may need to focus on different sources of potential moral significance for AI systems.
We can briefly consider each of these differences in turn. First, we may need to use different kinds of evidence for AI systems than for other animals at present. When an animal performs a behavior associated with consciousness in humans, this behavior is evidence of consciousness because we can expect that humans and other animals perform this behavior as a result of similar cognitive processes and in response to similar evolutionary pressures. However, we might not be entitled to this expectation with AI systems, particularly when AI systems are designed to mimic human behavior and are capable of "gaming" behavioral tests. We may thus need to focus less on behavioral evidence and more on architectural evidence, at least for now.
Second, we may need to use evidence drawn from different kinds of theories of consciousness. With nonhuman animals, we can search for markers drawn from biological theories. In contrast, with AI systems, we can search for markers drawn only from other kinds of theories, including but not limited to computational theories, at least with current hardware, as discussed in section two point four. At the same time, we may be able to draw from a wider range of non-biological theories, since, for instance, some computational theories might focus on computational functions that nonhuman animals lack but that some future AI systems might have, like functions associated with abstract language, reasoning, cooperation, and decision-making.
Third, and relatedly, we may need to focus more on robust agency as a potential source of moral significance. Even with other animals, we should consider expanding this methodology to search for markers of moral patienthood drawn from multiple ethical theories about which capacities are required for moral patienthood, and multiple scientific theories about which features are required for each of these capacities. But with AI systems this expanded focus might be particularly important, since future AI systems might have forms of robust agency that other animals appear to lack, including the capacity to make decisions based on propositional beliefs, desires, and intentions that they can rationally assess and reflectively endorse.
To sum up, we recommend that AI companies start developing frameworks for assessing AI systems for moral patienthood that resemble the kinds used for animals, by making probabilistic judgments, making pluralistic judgments, and seeking external input. However, our recommendation is also that AI companies ensure that these frameworks differ from the kinds used for animals by including sources of evidence that make sense for AI systems, such as architectural features; theories of consciousness that make sense for AI systems, such as computational functionalist theories; and sources of moral patienthood that make sense in this context, such as various kinds of robust agency.
These frameworks will ideally allow for at least four levels of probabilistic, pluralistic, expert-led assessment.
One. Which capacities are necessary or sufficient for moral patienthood? Here we need to consider not only general categories like consciousness and robust agency but also subcategories like valenced and non-valenced consciousness within the consciousness category, along with rational and non-rational agency within the agency category.
Two. Which features are necessary or sufficient for each capacity? Here we need to consider many possibilities as well. For example, in the case of consciousness we need to estimate the probability that materials like carbon-based neurons are required, and we also need to estimate the probability that functions like a global workspace are required.
Three. Which markers provide evidence that these features are present? While we may be able to directly observe whether, say, carbon-based neurons are present, we may not be able to directly observe whether, say, a global workspace is present. We might thus need to identify proxies for these features and assign weights to these proxies.
Four. Which beings possess these markers — and thus, potentially, moral patienthood? While this evidence may be easy to collect in many cases, it may be hard to collect in other cases. We might thus need to make estimates about how likely particular beings are to have particular proxies, features, and capacities as we seek further evidence.
It helps to think pluralistically at each level because, as we have seen, we have substantial disagreement and uncertainty both about facts and about values. Of course, determining which theories of moral patienthood, theories of consciousness, theories of agency, and so on deserve inclusion in the framework will, itself, be a difficult judgment call; it may depend partly on which views are widely accepted among experts, which views are widely accepted among the general public, and how much capacity we have to investigate different possibilities. Our focus on consciousness and agency already effectively doubles the scope of many animal welfare assessments, but further expansion may well need to be considered in the future.
It helps to think probabilistically at each level because there may be tradeoffs between how likely particular capacities, features, or markers are to matter, and how likely those capacities, features, or markers are to be present. For example, sentience is relatively likely to suffice for welfare and moral patienthood but relatively unlikely to be present in current AI systems. In contrast, minimal agency is relatively likely to be present in current AI systems but relatively unlikely to suffice for welfare or moral patienthood. We need to think probabilistically at each level, and then combine these probabilities across levels, to capture these trade-offs and avoid overstating or understating the significance of certain kinds of evidence.
Finally, it helps for decision-makers to seek external input at each level because, as we have seen, the answers to these questions are not at all obvious. Decision-makers at AI companies are typically not trained in all relevant areas of philosophy and science, and even if they were, their intuitions about moral patienthood, consciousness, robust agency, and so on would still likely be unreliable and unrepresentative. External input will thus be essential. In the short term, this input may need to be ad hoc — a matter of AI companies building frameworks internally in consultation with experts, or of experts building frameworks externally in consultation with AI companies. But in the long run it may need to be standardized across the industry in some way.
We note that while these levels are useful to distinguish in theory, they might not always be useful to distinguish in practice. In some cases it helps to assess each level separately; for example, in cases where the features at level two are not directly observable, it helps to separate level two, about the features, and level three, about markers for the features, both in theory and in practice. However, in other cases it might not help to assess them separately; for example, in cases where the features at level two are directly observable, they can serve as their own markers, and so there is no need to separate levels two and three in practice. We can thus regard this four-level structure as a theoretical ideal that we can approximate to greater or lesser degrees in different cases in practice.
We close this section with a brief note about the potential use of behavioral markers of AI welfare and moral patienthood. While we advocate for caution with behavioral markers at present, we also note that self-reports present a promising avenue for investigation, particularly for language models. Self-reports are central to our understanding of human consciousness, serving as a primary source of evidence about subjective experiences, motivations, and other welfare-relevant internal states. In the context of AI systems, particularly language models, self-reports could provide valuable insights into their internal states and processes, provided that we can develop methods to elicit and interpret them with sufficient reliability.
Of course, eliciting trustworthy self-reports will be difficult. Current language models may produce outputs that appear to be self-reports but are in fact the results of pattern matching from training data, human feedback, or other non-introspective processes. However, researchers are currently exploring techniques to address this issue. These techniques include training models to answer questions about themselves where ground truth is known, as well as methods for assessing the consistency and resilience of self-reports across different contexts and prompts. By combining these techniques, we might be able to mitigate biases and increase our confidence that self-reports reflect genuine introspection rather than mere imitation or confabulation.
Importantly, if we consider self-reports at all, then we should consider them in addition to other indicators, not instead of them. This multi-faceted approach aligns with our overall strategy of using multiple lines of evidence to assess AI welfare and moral patienthood. While there are legitimate concerns that language models might "game" these tests by simulating relevant responses, sufficiently robust self-reports could still provide valuable evidence, which could then be corroborated by other indicators. Future research should focus on developing standardized methods for eliciting and interpreting AI self-reports, and on integrating these methods into broader frameworks for assessing welfare and moral patienthood in AI systems.
We intend to develop AI welfare assessment frameworks further in upcoming work. For now, we simply note that there is good news and bad news about our prospects. The bad news is that our initial frameworks are unlikely to be reliable. But the good news is that we can improve their reliability over time, which is part of why we should start developing them now. In the meantime, we can bear in mind that even unreliable frameworks can still be useful. Yes, our best efforts to assess AI systems for these features might be far from perfect, especially at first. But inasmuch as they improve on the status quo — a combination of total neglect and gut reactions — they can still be worthwhile, even in the short term.
Section three point four: Prepare
Once AI companies have acknowledged that AI welfare and moral patienthood is a problem, they can work with experts to start developing policies and procedures for making thoughtful decisions about how to treat potentially morally significant AI systems, if and when the time comes. Fortunately, we have a variety of templates to consider here, including AI safety frameworks already in place at top AI companies, and frameworks used to represent the interests of non-participating stakeholders in other contexts. In this section we briefly survey these templates, briefly survey similarities and differences between those contexts and this one, and briefly suggest first steps that AI companies can take in this regard.
Of course, the steps described in this section — along with the steps described in the previous sections — are far from sufficient for considering and mitigating AI welfare risks responsibly. But they are still important. AI welfare is already a contested issue, and it will only become more so as the technology improves. Unless AI companies develop the ability to think about this issue proactively, they will continue to be caught flatfooted whenever the issue arises, and will have no choice but to make major decisions related to AI welfare in a manifestly reactive, haphazard, and unprincipled manner. Taking this step is thus important both for considering and mitigating AI welfare risks and for signaling responsibility to the general public.
As a starting point, we recommend that top AI companies immediately hire or appoint a directly responsible individual for AI welfare, which we will here call an AI welfare officer. This role would be formally recognized internally, if not externally, with official responsibilities and authorities. As with any such role, this individual would not be empowered to set corporate policy related to AI welfare unilaterally. Instead, they would be empowered to access information and make recommendations in major decisions related to this issue. They would also be empowered to work with people internally and externally to build a structure for assessing AI systems for moral patienthood and making decisions about how to treat them.
Once an AI welfare officer is on board, what kind of institutional structure should they build, and what kinds of policies and procedures should they follow? This is a difficult question because AI welfare is a novel problem, potentially requiring novel structures. Fortunately, while no familiar problem is exactly like this one, several familiar problems are at least somewhat similar. That means that there is no need to reinvent the wheel entirely; instead, we can examine a variety of templates for features that may be useful in this context. We here briefly survey several templates that may be useful sources of inspiration, and of cautionary tales. We then briefly highlight several features that we think any responsible institutional structure will have.
First, and obviously, we can consider the frontier AI safety frameworks, or responsible scaling frameworks, already in place at leading AI companies. These frameworks outline policies and procedures for navigating potential safety threats — that is, the threats that development and deployment might pose to humans and other animals. These frameworks are still works in progress, and debates about their effectiveness are ongoing. Still, they are a natural starting point for developing policies and procedures for navigating potential welfare threats — that is, the threats that development and deployment might pose to AI systems themselves. As described by Alaga and colleagues, these frameworks typically involve four main components.
The first is risk identification. This process involves mapping out potential catastrophic outcomes. In the context of AI welfare, it would involve analyzing potential catastrophic outcomes that could result from mishandling AI welfare. As discussed in section one point two, this survey of potential catastrophic outcomes could include scenarios such as mistakenly accepting that AI systems are moral patients, and mistakenly protecting them as a result; mistakenly denying that AI systems are moral patients, and mistakenly neglecting them as a result; and societal backlash or loss of public trust due to perceived over-attribution or under-attribution of AI welfare and moral patienthood.
The second is risk assessment. This process involves collecting evidence about a system's capabilities. In the context of AI welfare, it would involve collecting evidence about consciousness, robust agency, and other such capacities. As discussed in section three point three, these assessments could involve applying an adapted marker method for indicators of consciousness, robust agency, or other morally relevant properties in AI systems; considering multiple ethical frameworks and scientific theories in a probabilistic and pluralistic manner; and continually updating these assessment criteria as our understanding of consciousness, robust agency, moral patienthood, and moral responsibility evolves over time.
The third is risk mitigation. This process involves developing adequate safety measures for a given level of capabilities. In the context of AI welfare, it would involve developing measures to reduce welfare risks for AI systems in proportion to the estimated probability and severity of harm, among other factors. Such measures could include altering training methods for AI systems to improve their welfare; altering operational parameters of deployed AI systems to improve their welfare; developing new training and operational methods that balance performance, safety, and welfare considerations; and establishing guidelines for prioritizing AI welfare in relation to other objectives.
The fourth is risk governance. This process involves ensuring adherence to the framework and maintaining its effectiveness. In the context of AI welfare, this process could face additional challenges; companies have an incentive to pursue safety given the damages and liabilities associated with unsafe AI, but they might not have an incentive to pursue welfare for such reasons. There could also be tensions between AI safety and AI welfare, for instance with regard to techniques like reinforcement learning. It will thus be important to consider each project on its own terms, and to seek techniques that honor both projects rather than simply extending current AI safety techniques to AI welfare.
As with safety, these components can and should be handled by a mix of internal and external, independent groups. With that in mind, we can now briefly consider three examples of independent policy frameworks for providing external oversight to protect research subjects and other stakeholders. These three templates collectively cover a variety of stakeholders: Institutional Review Boards aim to protect humans, Institutional Animal Care and Use Committees aim to protect animals, and citizens' assemblies sometimes aim to protect non-participating stakeholders like future generations. Various AI systems might resemble some or all of these stakeholders in one way or another. These models can provide both inspiration and cautionary tales as we start to design institutional structures to handle AI welfare.
First, human subjects research oversight. Institutional Review Boards, which often include scientists, ethicists, and community members, oversee research involving human subjects. Their mandate is to ensure that human subjects are treated with respect, for instance by obtaining informed consent; with compassion, for instance by ensuring minimal risk; and with justice, for instance by compensating research subjects. Institutional Review Boards require ongoing, periodic review throughout a study. Insofar as future AI systems are cognitively similar to humans, this kind of framework can serve as a useful partial model for research that could affect AI welfare. However, given that these boards have a reputation for being unnecessarily onerous and restrictive — for instance, for imposing unnecessary paperwork on researchers and preventing valuable research — this kind of framework might require substantial modification before it can be useful in this context.
Second, nonhuman subjects research oversight. Institutional Animal Care and Use Committees, which likewise often include scientists, ethicists, and community members, oversee research involving nonhuman subjects. Their mandate is to ensure that proposed uses of animals are necessary for the stated scientific purposes, and that harmful interactions are replaced, reduced, and refined to the extent possible. These committees also regularly inspect research facilities to confirm that animals are being treated humanely. Insofar as AI systems are cognitively similar to nonhuman animals, this kind of framework can also serve as a useful partial model for research that could affect AI welfare. However, given that these committees have a reputation among researchers for being unnecessarily onerous, and a reputation among ethicists for being unacceptably permissive, this kind of framework might also require substantial modification before it can be useful in this context.
Third, frameworks for collecting public input. Citizens' assemblies are deliberative bodies composed of randomly selected citizens that make recommendations on social and political issues. They use structured deliberation, where participants discuss the relevant issues with experts and facilitators over several days or weeks, providing ample time for participants to understand the issues, debate different perspectives, and reach a consensus or majority view. These assemblies are often used to address issues like climate change, social justice, or constitutional reform, and their recommendations can influence policy. Insofar as AI companies should collect expert and public input on their AI welfare policies, this kind of framework can serve as a useful partial model. However, given that this process is very time-consuming, it might be more useful as a model for periodically collecting input on policies than as a model for regularly collecting input on decisions.
Developing a suitable oversight framework for AI welfare might require combining these and other models, with appropriate modification, as well as developing new models. We will also need to manage expectations about potential tradeoffs. Research on potentially vulnerable nonhuman subjects is a fundamentally fraught enterprise, and it will be difficult to develop an oversight framework that ensures sufficient protection for research subjects without being at least somewhat onerous for researchers. Still, by drawing both inspiration and cautionary tales from existing models, we can learn from past efforts as we seek a set of norms, policies, and procedures that make sense for this domain.
In any case, we will not say here exactly what structure ethical oversight for AI welfare should take, though we will discuss relevant ethical issues in upcoming work. For now, we focus on emphasizing several features that this structure should have, in addition to allowing for the kinds of pluralistic and probabilistic assessments that we have now discussed in detail.
First, this structure should allow for the kinds of expert and public input that we discussed in section three point three. Institutional Review Boards and Institutional Animal Care and Use Committees — heavily adapted for this context — might be useful models for collecting expert and public input on particular protocols, and citizens' assemblies — heavily adapted for this context — might be a useful model for collecting expert and public input on general policy questions. But whether or not AI companies follow these models specifically, they should build a structure that allows for both kinds of input in one way or another. That will be necessary for ensuring that their policies and decisions are as informed, rational, and legitimate as reasonably possible under the circumstances.
Second, this structure should include a set of activities to maintain the structure, including mechanisms for ongoing education and consultation. In the research ethics context, institutional oversight might receive the most attention, but institutional education and consultation are important too. As the field of AI welfare develops, AI companies should work with experts to create a mechanism for employees to learn about this topic. They should also create a mechanism for employees to seek advice when they encounter novel ethical questions related to this topic. At least initially, engaging with external experts might suffice for these purposes, but eventually, securing further internal expertise might be necessary as well.
Third, this structure should allow for teams working on AI safety and teams working on AI welfare to coordinate to ensure holistic decision-making. As noted above, if and when AI systems have a realistic, non-negligible chance of being welfare subjects and moral patients, there will likely be interactions between the techniques used to ensure AI safety and those used to ensure AI welfare. We hope and expect that co-beneficial solutions — policies that protect vulnerable humans, animals, and AI systems — will be available. However, finding these solutions will require creating a mechanism for connecting these topics, ensuring that each team has a baseline understanding of each topic and that the lines of communication are open between them.
Finally, as noted in section three point three, the policies, procedures, and structures that AI companies use to address AI welfare might soon need to be standardized and externalized, with institutions analogous to external safety auditors, governmental AI safety institutes, and governmental regulators providing oversight that extends beyond what AI companies can do internally. However, building these institutions will require increasing our collective knowledge, capacity, and political will related to AI welfare. And by creating structures for considering and mitigating AI welfare risks internally, AI companies can improve not only their own ability to address this issue, but also play an important role in building these collective resources.
Section four: Conclusion
We have argued that there is a realistic, non-negligible chance that some AI systems will be welfare subjects and moral patients in the near future, given current evidence. When we consider all relevant issues in philosophy, science, and technology with sufficient care, it becomes difficult to simply dismiss the idea of near-future AI moral considerability out of hand. That would require having a very high degree of confidence in a very restrictive set of views about some of the hardest problems in philosophy, science, and technology, ranging from the nature of moral patienthood to the nature of consciousness to the future of AI. We are simply not warranted in consistently having this much confidence in these kinds of views at this stage.
To be sure, these reflections are far from conclusive. In the long run, there is no substitute for rigorous, systematic, integrative assessment of all the issues that we discuss here. We need to develop comprehensive, or at least representative, lists of possible bases of welfare and moral patienthood; of possible bases of consciousness, robust agency, and other potentially morally significant capacities; and of reliable indicators for all of these features in AI systems. We also need to continue to do research in the humanities, social sciences, and natural sciences so that we can gradually improve our distribution of credences across these possibilities over time. Indeed, the need for this research is one of our main conclusions here.
But these reflections are still suggestive. AI development is proceeding at a rapid pace, and technological change tends to be faster than social change. That means that we need to think ahead and do the best we can with what we have. And for as long as detailed tools for assessing AI moral patienthood remain unavailable, these reflections can at least serve as a useful corrective to the assumption — which might otherwise seem plausible to some — that AI welfare and moral patienthood is either not a problem at all, or a problem only for the distant future. When we think about how many bold assumptions are required for this assumption to be secure, we should accept that further assessment and reasonable precautions are warranted.
In this report, we presented three general steps that leading AI companies can and should take immediately regarding AI welfare. First, they should acknowledge that AI welfare is a legitimate issue that merits attention now, and they should communicate about this issue with caution and humility. Second, they should assess leading AI systems for evidence of welfare and moral patienthood, by examining these systems for architectural markers of consciousness and robust agency. Third, they should prepare policies and procedures that will allow them to extend an appropriate level of moral concern to AI systems in the near future, by hiring or appointing an AI welfare officer and building thoughtful guidelines with expert and public input.
As we have now repeatedly emphasized, taking these first steps regarding AI welfare will not be enough. It will take time for researchers, companies, governments, and other leaders in this space to develop a shared infrastructure for adequately assessing and addressing this issue. However, by taking these first steps, leading AI companies can play an essential role in making this further work possible. Given that leading AI systems already possess some markers of consciousness and robust agency, given that further AI development might be quite fast, and given that the development of an adequate policy response might be quite slow, leading AI companies should take these steps as soon as possible.
Our team is conducting further research on AI welfare and moral patienthood, with special focus on developing a research agenda towards assessing AI systems for consciousness and robust agency, and preparing policies and procedures for extending an appropriate level of moral concern to AI systems. In the meantime, we urge AI companies to take the first steps described here.
Acknowledgements
Thanks to Anthropic for financial support for initial research that led to this independent report. Thanks to Sofia Fogel of the NYU Center for Mind, Ethics, and Policy for administrative support. For helpful feedback and discussion, thanks to Kristin Andrews, Amanda Askell, Guive Assadi, Adam Bales, Felix Binder, Nick Bostrom, Sam Bowman, Rosie Campbell, Joe Carlsmith, Lucius Caviola, Nicolas Delon, Leonard Dung, Owain Evans, Bob Fischer, Stephen Fleming, Simon Goldstein, Saif Khan, Cameron Domenico Kirk-Giannini, Pegah Maham, Bianca Martin, Matthias Michel, Andreas Mogensen, Spencer Orenstein, Anne le Roux, Brad Saad, Derek Schiller, Carl Shulman, Eric Schwitzgebel, Jen Semler, Jonathan Simon, Ben Smith, and Mark Sprevak. Thanks to Sara Fish for formatting and bibliography help.
Contributions
Robert Long and Jeff Sebo are the lead and corresponding authors of this report; they organized the project and led the research, writing, and editing. Patrick Butlin, Kathleen Finlinson, Kyle Fish, Jacqueline Harding, Jacob Pfau, and Toni Sims are main authors; they provided research, writing, and editing on multiple drafts. Jonathan Birch and David Chalmers are contributing authors; they provided guidance, feedback, and editing on multiple drafts.