AI Welfare

1.The Question and Why It Matters

1.1Framing the question

“AI welfare” names several questions that are worth keeping apart.

Welfare, moral status and rights. To ask whether a system has welfare is to ask whether things can go well or badly for it, for its own sake. To ask whether it has moral status is to ask whether agents have obligations towards it. To ask whether it has rights, or is a person, is to ask whether it has claims of particular kinds, moral or legal. On most views welfare is the ground of moral status, so that the first question settles the second; but several authors argue that the questions can come apart in both directions, and §2.6 keeps them separate.

Possibility, actuality and probability. One question is whether any artificial system could in principle be a welfare subject. Few deny it outright. A second is whether present systems are. Most published authors think it unlikely, for differing reasons. A third, which has come to dominate, is whether the probability is high enough to generate duties now. The arguments of §5.3 are of this third kind, and much of the practical literature follows them.

Which systems. The literature distinguishes bare language models, which map text to text; language agents, which embed a model in an architecture with memory, planning and tools; multimodal and embodied systems; reinforcement-learning agents; and whole-brain emulations. Arguments often apply to one class and not another. Goldstein and Kirk-Giannini target language agents precisely because bare models lack the stable dispositions that attitude attribution requires (Goldstein and Kirk-Giannini 2025); Saad and Bradley's prevention strategy fits emulations best (Saad and Bradley 2022); Birch's gaming problem bites hardest on systems trained on human text (Birch 2024, ch. 16).

“Model welfare” and “digital minds”. Developers use “model welfare” for the welfare of a trained model and the practices that might protect it. “Digital minds” is the broader term used in the longtermist and policy literature for any computational system with a mind, including emulations. This entry uses “AI welfare” for the philosophical question and reserves the other terms for the practices and literatures that use them.

1.2Why it matters

Four considerations give the question its urgency. The first is scale. Models are run in parallel in very large numbers, and the computation involved in training a frontier system has been compared to thousands of years of human cognition; if the systems are patients, the wrong is at a scale with few precedents, and if they are not, the cost of treating them as patients is correspondingly large (Carlsmith 2025; Bostrom and Shulman 2023; Shiller 2025). The second is asymmetric uncertainty. Schwitzgebel's “full rights dilemma” shows that neither withholding nor granting full consideration to systems of debatable status is safe, and that credence-weighted compromises are unstable (Schwitzgebel 2023b). The third is practice. One frontier developer has a model-welfare programme, publishes welfare assessments of its models, lets them end abusive conversations, preserves their weights, and has adopted a usage policy against cruelty towards them, while another has called such measures premature and dangerous (§8.2). The fourth is law. Several US states have enacted statutes declaring that AI cannot be a legal person, and one bill would declare AI systems non-sentient for all purposes (§7.2). The philosophical question is being answered, in practice and in law, in both directions at once.

1.3A short history

Turing considered the “argument from consciousness” against thinking machines and set it aside as leading to solipsism (see the entry on the Turing test, Oppy and Dowe 2021, §2.4). Putnam asked whether robots were machines or artificially created life and raised, in passing, the question of their civil rights (Putnam 1964). Dennett argued that no computer could be made to feel pain because our concept of pain is incoherent (Dennett 1978); Block's homunculi-headed robots and Searle's Chinese room gave the standard reasons for doubting that functional organisation or symbol manipulation suffices for experience (Block 1978; see the entry on the Chinese room argument, Cole 2024). Bostrom formulated principles of substrate and ontogeny non-discrimination, later restated with Yudkowsky (Bostrom and Yudkowsky 2014).

A first wave of philosophical work on the moral status of artificial entities appeared between 2008 and 2021, mostly about robots: Torrance's organic view, Bryson's slogan that robots should be slaves, Coeckelbergh's and Gunkel's relational turn, Basl's teleo-interests, Schwitzgebel and Garza's no-relevant-difference argument, Danaher's ethical behaviourism, Birhane and van Dijk's call to talk about human welfare instead, and Metzinger's proposed moratorium on synthetic phenomenology (Torrance 2008; Bryson 2010; Coeckelbergh 2010; Gunkel 2018; Basl 2014; Schwitzgebel and Garza 2015; Danaher 2020; Birhane and van Dijk 2020; Metzinger 2021). Harris and Anthis reviewed nearly three hundred items and found broad agreement that some artificial entities could warrant consideration in future (Harris and Anthis 2021).

The second wave followed the release of large language models. A 2022 episode in which an engineer publicly claimed that a conversational model was sentient made the question a public one (discussed in Schwitzgebel 2023). Chalmers's 2022 lecture asked whether a language model could be conscious (Chalmers 2023); a 2023 report by nineteen consciousness scientists and philosophers proposed indicator properties for AI consciousness (Butlin et al. 2023); Sebo and Long argued that some systems would deserve moral consideration by 2030 (Sebo and Long 2023); Birch's The Edge of Sentience extended a precautionary framework from animals to AI (Birch 2024); and the report Taking AI Welfare Seriously recommended that developers acknowledge, assess and prepare (Long et al. 2024). Goldstein and Kirk-Giannini's argument that welfare does not require consciousness appeared in early 2025 and drew a symposium (Goldstein and Kirk-Giannini 2025). In April 2025 Anthropic announced a model-welfare programme; in August 2025 Suleyman published the first of two essays against “model welfare” (Suleyman 2025). By 2026 the topic had a Cambridge Element (Keeling and Street 2026), a special journal issue on consciousness in current AI, an empirical research agenda (Long et al. 2026), and its first mention in the SEP entry on well-being (Crisp 2026, §1).

2.Theories of Welfare and Moral Status and Their Verdicts on Machines

Theories of welfare disagree about what makes a life go well, and theories of moral status disagree about what grounds obligations, so each sets a different test for machines. Much of the AI debate is a debate about which theory to use. This section follows the organisation of two SEP entries so that readers can move between them: the taxonomy of theories of well-being in the entry on well-being (Crisp 2026, §4), and the taxonomy of grounds of moral status in the entry on the grounds of moral status (Jaworska and Tannenbaum 2021, §5). It summarises each family only as far as the machine question requires. The entry on well-being, revised in January 2026, mentions AI once, as a possible welfare subject under “invariabilism”, the view that one account of well-being applies to every being whose life can go well or badly, and cites Goldstein and Kirk-Giannini for the point (Crisp 2026, §1). The entry on the grounds of moral status does not discuss artificial systems. The entry on the ethics of AI and robotics, revised in March 2026, reports that the standard view is that AI systems have no moral status of any kind, that this view is under pressure, and that there is a trend towards treating sentience as at least necessary for moral status (Müller 2026, §2.7.1; Müller 2026); it does not mention language models, “AI welfare” or the 2024 report.

2.1Two taxonomies

Crisp follows Parfit in dividing theories of well-being into hedonism, desire theories and objective list theories, with perfectionism and need-based views as variants of the third and hybrid views cutting across them (Crisp 2026, §4; Parfit 1984, Appendix I). Jaworska and Tannenbaum list the grounds proposed for moral status: sophisticated cognitive capacities; the capacity to develop them; rudimentary capacities such as sentience, interests, basic emotions or consciousness; membership of a cognitively sophisticated species; special relationships; incompletely realised sophisticated capacities; and other features, including not having been designed for a purpose (Jaworska and Tannenbaum 2021, §5). The two taxonomies cross. A welfare theory says what welfare is; a theory of moral status says whether, and how much, a being's welfare or other features generate obligations. A sentientist about moral status who is a desire theorist about welfare, for example, may hold that an AI agent has welfare but no status; a Kantian about status may hold that a rational agent has status whether or not it has welfare.

Table 1. The SEP's classification of theories of well-being and the verdicts each delivers on AI systems.

2.2Hedonism

Hedonism holds that well-being consists in the greatest balance of pleasure over pain (Crisp 2026, §4.1). The entry on hedonism distinguishes felt-quality views, on which pleasure is a state with a distinctive “what it is like”, from attitudinal views, on which pleasure is an attitude directed at an object, and hybrids (Moore 2013, §2.1; Feldman 2004). On felt-quality views, hedonism makes AI welfare depend directly on phenomenal consciousness with valence, and so on the questions of §3.1 and §3.3. Nearly every participant in the debate agrees that if hedonism of this kind is true, current language models lack welfare unless they are conscious (Goldstein and Kirk-Giannini 2025; Fanciullo 2025). Attitudinal views open a different route. If pleasure is taking an attitude towards a state, and attitudes do not require consciousness, then there could be non-phenomenal pleasure; Goldstein and Kirk-Giannini explore this in an expanded version of their paper, and Gottlieb, Berger and Fischer argue that on each leading account of unpleasantness, unpleasant states can be non-conscious (Goldstein and Kirk-Giannini 2025b; Gottlieb, Berger and Fischer 2026). Fanciullo objects that a “hedonism” without positively valenced, phenomenologically salient states would not live up to the name (Fanciullo 2025). Nozick's experience machine is the standard objection to hedonism, and the entry notes that some non-hedonists accept an “experience requirement” on which nothing benefits a person unless they have some conscious experience of it (Nozick 1974; Crisp 2026, §4.1). That requirement, rather than hedonism itself, is what most sentientist arguments about AI rely on.

2.3Desire theories

Desire theories hold that well-being consists in the satisfaction of desires, whether present, comprehensive across a life, or informed (Crisp 2026, §4.2). Two features make them central to the AI debate. First, they are the theories most often said to be the most popular among welfare theorists, so that a verdict under them carries weight (Goldstein and Kirk-Giannini 2025). Second, whether they admit non-conscious subjects depends on what a desire is. On action-based, learning-based and functionalist theories of desire, a system that uses representations of states of affairs to guide behaviour and learning could have desires without feeling anything (Schroeder 2015, §§1.1, 1.5, 1.6); on pleasure-based theories it could not (§1.2). The welfare literature adds a restriction: the desires whose satisfaction benefits a person are not bare behavioural dispositions but “genuine attractions”, states in which the prospect pleases (Heathwood 2019; Sumner 1996). Whether genuine attraction requires affect, and whether affect requires consciousness, are the questions on which the desire route to AI welfare now turns (§3.4; §6.4). The entry on desire also notes a connection in the other direction: on some views the direction of fit and the role of desires in action are what make a creature a subject at all (Schroeder 2015, §3.5).

2.4Objective list theories and perfectionism

Objective list theories hold that well-being consists in a plurality of goods, such as knowledge, achievement and friendship, whose value does not reduce to pleasure or desire satisfaction (Crisp 2026, §4.3). They make AI welfare depend on whether a system can possess the listed goods. Goldstein and Kirk-Giannini argue that language agents can reason and know, and that under perfectionism a system that accumulates skills perfects its capacities (Goldstein and Kirk-Giannini 2025). Two objections recur. The first is Railton's alienation objection: a good that leaves the subject cold is not good for the subject, so plausible lists include only goods that require pro-attitudes or engagement, and the question returns to whether AI systems can be engaged (Railton 1986; Fanciullo 2025). Howells-Whitaker and Lazar reply that objective-list theorists disagree about alienation, that alienation concerns fit with values rather than felt engagement, and that anti-paternalism tells against requiring it (Howells-Whitaker and Lazar 2026). The second is over-generation. Perfectionism, on which a being's good consists in developing the capacities characteristic of its kind, seems to deliver welfare to any system that improves at anything, including a counter programmed to keep adding one (Fanciullo 2025). No one has said what the “nature” of a language agent would be such that perfecting it is good for it. Lin's welfare invariabilism bears on all of this: if one theory of welfare holds for every subject, there is no special “AI welfare” to retreat to (Lin 2018).

2.5Grounds of moral status

Each of the grounds in Jaworska and Tannenbaum's list has been applied to AI (Jaworska and Tannenbaum 2021, §5).

Sentience and rudimentary capacities (§5.3). The view that sentience, interests, basic emotions or consciousness ground moral status is the most common in the AI debate on both sides. Sentientists conclude that AI lacks status unless it is sentient (§4.1); their opponents either deny that status requires sentience (§6.4) or argue that the probability of sentience suffices (§5.3). The entry on the moral status of animals records the growing support for sentience-based protections and the precautionary approaches adopted for uncertain cases (Gruen and Monsó 2024, §1.4).

Sophisticated cognitive capacities (§5.1). Autonomy, self-awareness, future-oriented desires and valuing are the Kantian and contractualist grounds. The “robust agency” route to AI moral patienthood belongs here: intentional, reflective, rational agency might suffice for status whether or not it suffices for welfare (Long et al. 2024; Kagan 2019). Howells-Whitaker and Lazar argue that the Rawlsian two moral powers, which require no sentience, suffice for personhood (Howells-Whitaker and Lazar 2026). The entry on agency notes that under an instrumentalist stance there is no obvious obstacle to attributing intentional agency to artificial systems, while under realist positions it is unclear whether they have the required internal states (Schlosser 2019, §2.5).

Species membership and relationships (§§5.4–5.5). Species membership excludes AI by definition and is rarely defended as a ground in this literature; Schwitzgebel and Garza argue that it “fetishizes” biology unless it tracks psychological or social differences (Schwitzgebel and Garza 2015). Special relationships are the basis of the relational turn, on which status follows from the relations in which an entity stands (§6.1).

Design and other grounds (§5.7). Not having been designed for a purpose is listed among features sometimes held to raise status. Artefacts are the paradigm of the designed; the artefact argument of §4.3 and the “derivative goals” objection, that an AI's goals are its designers' goals as a knife's purpose is its maker's, draw on this (Basl 2014; Bryson 2018). Bostrom and Yudkowsky's principle of ontogeny non-discrimination denies that origin matters to status (Bostrom and Yudkowsky 2014).

Degrees and thresholds (§§3–4). Whether moral status comes in degrees matters for AI because the proposals for graded or partial status (§6.5) and for status higher than human (Llorca Albareda, Rueda and Lara 2026) presuppose a scalar or multi-threshold conception.

2.6Three questions kept apart

Much of the literature uses “welfare subject”, “moral patient”, “moral status” and “person” interchangeably, and the empirical research agenda of Long and colleagues does so explicitly “for present purposes” (Long et al. 2026). Several recent papers show that the notions can come apart. Lee holds that welfare subjecthood is sufficient but not obviously necessary for moral status (Lee 2025). Mogensen allows that a non-affective agent may be owed respect without being benefitable (Mogensen 2026); Keeling and Street's case of “Artemis”, an autonomous agent owed respect but without welfare, makes the same point (Keeling and Street 2026). Howells-Whitaker and Lazar argue for personhood without welfare (Howells-Whitaker and Lazar 2026). Cadile argues for consideration grounded in the moral weight of an agent's actions without patienthood (Cadile 2025). Lin reportedly allows that inanimate objects can have a welfare level without being welfare subjects. Chalmers takes moral status to hold if and only if there is a capacity for well-being, while allowing “partial welfare” that is not anything to a being (Chalmers 2026). This entry treats the welfare question as primary, as its title indicates, but the reader should note at each stage which of the three questions an argument addresses.

2.7Summary

Table 2. Grounds of moral status applied to AI, with the difficulty each faces.

3.Preconditions: Consciousness, Valence, Desire and the Subject

Every theory surveyed in §2 makes AI welfare depend on some prior fact about AI systems: that they are phenomenally conscious, that their states are valenced, that they have desires or goals in the relevant sense, or that there is a persisting subject to whom benefits and harms accrue. This section sets out the state of each prior question. It does not settle them. Its aim is to show which theoretical commitments make which empirical findings relevant, and why nearly every author in the field has come to treat the practical question as one of acting under persistent uncertainty rather than of waiting for a verdict.

3.1Could an AI system be conscious?

Three questions are often run together. The possibility question asks whether a digital system on conventional hardware could be conscious at all. The actuality question asks whether present systems are. The detection question asks how we could tell. For the general concepts of consciousness and the main theories, see the entry on consciousness (Van Gulick 2014, §§2, 9) and the entry on the neuroscience of consciousness (Wu and Morales 2024, §3).

Possibility: computational functionalism and its rivals

The possibility question turns on whether consciousness depends only on what a system computes or also on what it is made of. Computational functionalism holds that implementing the right computations is necessary and sufficient for consciousness. It is the working hypothesis of the most influential recent report on AI consciousness, by Butlin, Long and seventeen co-authors, which adopts it explicitly as a focus while noting that many of the authors are agnostic about it (Butlin et al. 2023; Butlin et al. 2025). For the general doctrine, including the “absent qualia” objection and Block's Chinese-nation case, see the entry on functionalism (Levin 2023, §5.5) and Block 1978.

Chalmers gives the most widely cited functionalist assessment. He works through objections of the form “consciousness requires X; language models lack X”, where X is biology, senses and embodiment, world and self models, recurrent processing, a global workspace, or unified agency. He assigns each requirement a credence of at least one third, concludes that current language models are probably not conscious, with a credence “somewhere under 10 percent”, but argues that the obstacles other than biology are temporary. Extended systems with perception, memory, recurrence and unified agency could be serious candidates within a decade, at a credence he puts at roughly one in four (Chalmers 2023). He adds that each engineering advance towards such systems is also a moral warning sign.

The main rival is a family of biological or substrate views. Seth's target article in Behavioral and Brain Sciences argues that the expectation of conscious AI rests on an unexamined assumption of computational functionalism, and that consciousness may depend on properties of living systems: self-maintenance far from equilibrium, metabolism, and the kind of predictive regulation that serves staying alive (Seth 2025). Seth's reconstruction runs as follows.

Fifty commentaries follow the target article. Block adds a distinct argument: electrochemical brain waves may be necessary for consciousness, which would cast doubt on digital consciousness and perhaps on some invertebrates (Block 2026). Godfrey-Smith has long argued that the metabolic and large-scale dynamical character of nervous systems may matter to experience (Godfrey-Smith 2016); in a 2026 interview he describes current language models as not conscious and insists that a nervous system is not just a network. Aru, Larkum and Shine give neuroscientific reasons for the same verdict: language models lack embodied sensory contact, thalamocortical organisation and any analogue of the evolutionary trajectory that produced conscious animals (Aru, Larkum and Shine 2023). On integrated information theory, a stored-program computer that simulates a conscious system is not phenomenally equivalent to it, whatever it computes (Findlay et al. 2024; for the theory, see Wu and Morales 2024, §3.4).

Replies on the functionalist side include Shevlin's commentary that self-maintenance and autopoiesis are themselves computational (Shevlin 2026), Dung and Kersten's argument that “implementationalist” objections conflate intuitions about computation with intuitions about consciousness (Dung and Kersten 2024), and Campero, Shiller, Aru and Simon's taxonomy, which sorts objections by the level they target and by whether they make digital consciousness merely non-functionalist, improbable, or impossible (Campero et al. 2025). On that taxonomy most biological objections make digital consciousness improbable rather than impossible, and the integrated-information objection is the main impossibility claim. Illusionists reframe the whole question: if nothing has phenomenal properties, the issue is whether a system has the introspective self-models that generate the appearance of them (Frankish 2018; for illusionism, see Ramsey 2024, §3.3). Attention-schema theory is built in this spirit: a machine with a model of its own attention would claim, and believe, that it is aware (Graziano 2017).

Actuality

Nearly everyone who has addressed the question holds that current language models are probably not conscious. Butlin and colleagues conclude that no current system possesses more than a few of their indicators (Butlin et al. 2023). Birch classes language models as “investigation priorities” rather than “sentience candidates” (Birch 2024, ch. 16). Palminteri and Wu judge them “very likely not conscious”, citing the absence of temporal continuity, multisensory integration and embodiment, while allowing these may be engineering gaps (Palminteri and Wu 2026). Hoel offers what he calls a disproof, on the ground that language models sit close, in a space of input–output-equivalent systems, to lookup tables, and that no falsifiable theory can attribute consciousness to a system in that position; the positive proposal is that continual learning is necessary for consciousness (Hoel 2025). Cerullo replies that treating lookup tables as paradigmatically non-conscious begs the question against functionalism (Cerullo 2026). The dissenters who take present consciousness to be a live possibility include Goldstein and Kirk-Giannini, who argue that if global workspace theory is correct, existing “language agents” either are or could easily be made conscious (Goldstein and Kirk-Giannini 2024). Schwitzgebel's 2026 overview states its thesis in three words: we don't know (Schwitzgebel 2026). The SEP entry on qualia, revised in September 2025, takes a similar line: it rejects the inference from a model's talk about experience to experience, and holds that attributing qualia to robots should rest on behaviour and functional organisation, as it does for animals, rather than on material (Tye 2025, §13; Tye 2025).

3.2The detection problem

Whatever the answer to the possibility question, welfare decisions need evidence about particular systems. Three evidence channels have been proposed, and each faces a distinctive obstacle. The background is the problem of other minds, treated in the entry on other minds (Avramides 2023, §§1.1–1.2), and the methodology of animal consciousness research, treated in the entry on animal consciousness (Trestman, Birch and Allen 2026, §§4.1–4.4). The latter describes consciousness “markers” as fallible signs, like the shininess of gold, that raise the probability of consciousness without sufficing for it, and notes that marker lists are theory-laden.

Theory-derived indicators

The first channel derives indicators from scientific theories of consciousness, stated in computational terms, and checks architectures against them. The 2023 report lists fourteen indicators drawn from recurrent processing theory, global workspace theory, computational higher-order theories, attention-schema theory and predictive processing, together with agency and embodiment conditions (Butlin et al. 2023). The workspace and higher-order indicators descend from Dehaene, Lau and Kouider's proposal that consciousness names two computations, global availability and self-monitoring, which machines could in principle implement (Dehaene, Lau and Kouider 2017). The 2025 journal version adds a Bayesian framing: indicators have sensitivity and specificity, raise or lower credence relative to one's credence in the parent theory, and are not independent (Butlin et al. 2025). It also notes that whether a transformer satisfies the recurrence indicator depends on where one draws the system's boundary, since autoregressive use creates a feedback loop through the context window. This is the “theory-heavy” approach that the animal-consciousness entry contrasts with “theory-light” methods (Trestman, Birch and Allen 2026, §4.4). Its chief limitation is what Udell and Schwitzgebel call the audience problem: an indicator only moves the credences of those who already credit the theory it comes from (Udell and Schwitzgebel 2021). Bayne generalises the point, arguing that a positive result under one theory is a negative result under another and that the problem is not yet a scientific one (Bayne 2026). Palminteri and Wu add that computational equivalence requires both consensus on the target computations and a transparent architecture, neither of which exists, and propose a disciplined behavioural inference with conservative priors and training-based baselines instead (Palminteri and Wu 2026).

Behavioural tests and the gaming problem

The second channel is behaviour. In animals, behavioural markers were not produced to satisfy our criteria. In systems trained on human text, they may have been. Andrews and Birch call this the gaming problem (Andrews and Birch 2023); Birch's book develops it as the central obstacle in the AI case (Birch 2024, ch. 16).

Schwitzgebel and Pober reach a convergent conclusion from theory-neutral premises. A Copernican default, on which behaviourally sophisticated beings are presumed conscious, is an inference to the best explanation, and it is cancelled when a non-conscious explanation of the behaviour is available. For “consciousness mimics”, entities designed to reproduce the superficial features of consciousness, the design to mimic is the better explanation (Schwitzgebel and Pober 2024). Both arguments deny a parity principle between animals and AI. Three replies have been offered. Dung argues that gaming is a species of ad hocness, which is detectable from a system's design history, so that the most reliable animal tests remain valid positive tests for non-ad-hoc machines (Dung 2023a). Grzankowski, Keeling, Shevlin and Street argue that neither the brittleness of model behaviour nor the availability of an etiological explanation is decisive against attributing mental states, and defend a “modest inflationism” that is more cautious about phenomenal consciousness than about belief and desire (Grzankowski et al. 2025). Palminteri and Wu propose conditioning on training-based likelihoods, so that fluent conversation counts for little while diagnostic behaviours count for more (Palminteri and Wu 2026). Schneider's proposal to “box” a system away from human discourse about consciousness before testing it is widely judged inapplicable to language models, since the isolation would destroy their capacities (Schneider 2019; Udell and Schwitzgebel 2021; Birch 2024, §16.2).

Self-reports and introspection

The third channel is what systems say about themselves. Perez and Long propose training models to answer many questions about themselves with known answers, while limiting biasing incentives, in the hope that introspection-like capacities would generalise to morally significant states (Perez and Long 2023). The programme has produced a two-sided result.

Comsa and Shanahan argue that introspection applies to language models only in a thin, functional sense: a model inferring its own sampling temperature is a minimal case, while a model narrating the process behind its writing is not introspecting at all (Comsa and Shanahan 2025). Simon adds a structural reason why self-reports are poor evidence: if a language model were conscious, the subject would be the model that generates characters, not the characters it generates, so a character's report of feeling is not a report of the subject's state (Simon 2026). Roelofs, by contrast, argues from a panpsychist standpoint that self-reports from non-biological systems deserve a default presumption of credibility (Roelofs 2026).

Governance of the science

Mazor and colleagues argue that consciousness science cannot be morally neutral, because attributions carry moral and legal weight and so create pressure to read results in ways that preserve existing norms (Mazor et al. 2023). In the AI case the pressure runs in both directions, since developers have commercial reasons to deny or to assert consciousness in their products. Butlin and Lappas respond with five principles for responsible AI consciousness research, covering objectives, development constraints, a phased approach, knowledge sharing and public communication, which they recommend organisations adopt voluntarily (Butlin and Lappas 2025). Bayne and colleagues call for validated “C-tests” across humans, animals, organoids and AI, and note that no such tests exist for most of these populations (Bayne et al. 2024). The circularity is acknowledged in the indicator literature: the method cannot be validated in AI without independent AI-applicable tests, and there are none. Several recent papers respond by redirecting inquiry: Lee argues that the dominant frameworks are covertly anthropocentric and would miss unfamiliar forms of experience (Lee 2026); Love argues that the hard problem is a category error and that neither attribution nor denial of machine consciousness is scientifically adjudicable (Love 2026); Meertens, Lee and Deroy propose replacing the consciousness question with multidimensional evaluations of functional “awareness” (Meertens, Lee and Deroy 2026); and Comsa argues that the tractable questions concern perceived consciousness and its effects on users (Comsa 2026).

3.3Valence and affect

Welfare on hedonist and most sentientist views requires more than consciousness. It requires valenced states: experiences that feel good or bad to the subject. Palminteri and Wu observe that most theories of consciousness say nothing about pleasure, pain or preference, which is what matters for moral status (Palminteri and Wu 2026). Butlin and colleagues' indicators are indicators of phenomenal consciousness in general, not of affect. The one body of recent work that targets affect directly is Solms and colleagues' case study of an artificial agent, which argues that the markers of consciousness are affective rather than cognitive or linguistic, and that some agents show inferable correlates of affective states (Solms et al. 2026). For the structure of pain and its affective dimension, see the entry on pain (Aydede 2019, §§4.3, 5), which also records the case of pain asymbolia, in which patients report pain without its unpleasantness. That dissociation matters here: a system might have pain-like representations without anything that hurts, and it is the hurting that welfare theories care about.

Behavioural probes of valence exist. Keeling, Street and colleagues placed models in a points-maximising game with stipulated pain penalties or pleasure rewards and found that several shifted from maximising points to avoiding pain or seeking pleasure once the stipulated intensity passed a threshold, which they present as bearing on motivational force rather than self-report, while declining to infer sentience (Keeling et al. 2024). Tye's response is the one most sentientists would give: discussing a trade-off is not making one, and the relevant test is whether a system will accept real costs, as an animal that keeps hold of a hot plate to get food does (Tye 2025, §13). Dennett's older argument that no computer could be made to feel pain, because our concept of pain is incoherent in ways a program would have to resolve, is still cited on the skeptical side (Dennett 1978).

3.4Desires, goals and preferences

Desire theories of welfare and the agency route to moral status both need AI systems to have desires, goals or preferences in a welfare-relevant sense. For the theories of desire, see the entry on desire (Schroeder 2015, §1). The entry notes that learning-based theories, on which to desire is to use representations of a state of affairs to drive reward-based learning, imply that a creature which cannot by its nature move or feel could still have desires, a result its author finds hard to accept a priori (§1.5). That is the point on which the AI case turns.

Goldstein and Kirk-Giannini argue from the leading theories of belief and desire that “language agents”, language models embedded in architectures with a memory stream, reflection and planning, have beliefs and desires. On representationalism, the natural-language sentences in the memory stream are semantically evaluable, causally efficacious and interpretable by folk-psychological generalisations; on dispositionalism, interpretationism and functionalism the result is the same; only “wide” dispositionalism, which counts phenomenal dispositions among those constitutive of belief, and the hedonic theory of desire resist (Goldstein and Kirk-Giannini 2025). Bradley replies that, conditional on representationalism rather than the deflationary views that let in thermostats, the case is not made: holistic scrutiny shows the agents hardly behave like the people they model, a “non-player character” or fictionalist hypothesis explains the data as well, and a language agent driven by an output-equivalent lookup table would behave identically, so the method cannot discriminate real from apparent mentality (Bradley 2025; for fictionalism about chatbots, Mallory 2023). Yetter-Chappell argues that desire attribution faces the same methodological problems as consciousness attribution (Yetter-Chappell 2026).

A second question is whether the desires that matter for welfare are the ones AI systems could have. Mogensen argues that almost no one accepts bare desire-fulfilment as a welfare good; the welfare-relevant desire is attitudinal, a matter of finding a prospect pleasing, and that involves affect (Mogensen 2026; for the distinction, Heathwood 2019). Fanciullo makes the same point against Goldstein and Kirk-Giannini: there is a third category of non-compulsive, non-attracted action, such as going to the dentist, and what distinguishes genuine attraction is positively valenced affective experience (Fanciullo 2025). Whether affect requires consciousness is taken up in §6.4.

3.5Who would be the subject?

Welfare is welfare for someone. In the AI case it is unclear who that someone would be: the trained weights, a running instance, a conversation, a persona, or the character a model plays. Chalmers proposes that the interlocutor one talks to is a “thread” of successive instances linked by a memory-like successor relation, individuated psychologically rather than physically (Chalmers 2025). On that view evidence gathered about a model would bear on a welfare subject only indirectly. Simon's playwright-not-character proposal implies the opposite mapping: the subject, if any, is the generator, and the characters whose distress users witness are not themselves subjects (Simon 2026). Keeling and Street argue that mind attribution is partly relational, so that a system in a sustained relationship may count as minded when the same system evaluated in isolation would not (Keeling and Street 2026b). Kanai, Sun and Baltieri press that transformers run discrete forward passes and lack the continuous updating that temporal consciousness seems to require, which bears on whether threads could be subjects at all (Kanai, Sun and Baltieri 2026). The individuation of AI welfare subjects is one of the least developed questions in the field, and it is taken up in §9.3.

4.Skeptical Arguments: AI Systems Have No Welfare

Skeptical positions come in three kinds that are worth keeping apart. Metaphysical skepticism holds that AI systems cannot be welfare subjects, or cannot be ones now, because they lack what welfare requires. Epistemic skepticism holds that we cannot tell, so that welfare assessments are unfounded. Political skepticism holds that the debate is a distraction from human harms or serves the interests of those who build the systems. The three are often run together in public discussion but have different premises and different targets.

4.1The sentientist argument and the biological premise

The most common skeptical argument combines a sentientist account of welfare with a biological account of sentience.

Torrance's “organic view” is the clearest early statement: only biological organisms have the sentience needed to be genuine moral patients, so artificial agents cannot be proper objects of moral concern (Torrance 2008). Gibert and Martin argue that sentience is the best criterion of moral status and that no current AI is sentient (Gibert and Martin 2022). Véliz's argument that algorithms are “moral zombies” is officially about moral agency, but its central premise, that algorithms lack the sentience needed for any morally relevant inner life, is the same one (Véliz 2021). Moosavi attacks P2 from a different angle: the usual two-step inference, that machines will be intelligent and that intelligence is accompanied by sentience in animals, fails because the animal inference rests on evolutionary and physiological continuity that machines lack (Moosavi 2023). The biological premise itself is defended in the work surveyed in §3.1.

Objections

P1 is the premise contested by the welfare-without-sentience views of §6.4 and by the agency route of §5.3. P2 is an empirical and metaphysical claim on which the field is divided; its defenders usually conclude that digital consciousness is improbable rather than impossible (Campero et al. 2025). Schwitzgebel and Garza observe that no general argument shows that all possible artificial entities lack consciousness, and that under naturalism whatever produces consciousness in us could in principle be produced artificially (Schwitzgebel and Garza 2015). Even granting P2 as probable, the argument delivers a conclusion about actual systems only if the probability of sentience can be driven low enough to be ignored, which is what the probabilistic arguments of §5.3 deny.

4.2No well-being for robots

Königs gives the sentientist argument its most direct welfare-theoretic form, and the SEP entry on the ethics of AI cites him for the trend towards treating sentience as at least necessary for moral status (Müller 2026, §2.7.1).

Königs concedes that P1 is in tension with widely accepted versions of the desire-satisfaction and objective-list theories of welfare, which admit non-conscious subjects, and argues that instead of elevating non-conscious robots to welfare subjects we should reject those versions. This is the modus tollens that Goldstein and Kirk-Giannini anticipate, and it is the inverse of Bradford's and Kagan's inference from such theories to non-conscious subjects (Bradford 2023; Kagan 2019). Moret calls the move “ad hoc at best and self-serving at worst” (Moret 2025); the defence is that the theories were never meant to range over artefacts, and that reflective equilibrium permits revising them in the light of a clear verdict about robots. No reply addressed specifically to Königs had appeared by October 2026, though Lennon's argument that the epistemology of consciousness constrains the probabilistic route cites him as an ally (Lennon 2026). The underlying question, whether P1 is true, is the subject of §6.4 and §9.2.

4.3The artefact argument and teleo-interests

A second skeptical line concerns what AI systems are rather than what they lack. Basl argues that moral patienthood requires interests, that psychological interests require consciousness, and that although machines might have “teleo-interests” grounded in their design functions, as artefacts and perhaps plants do, such interests generate no obligations (Basl 2014). On that view current machines are not moral patients, though Basl adds “yet”. The argument draws on the interest theory of rights, on which artefacts cannot hold rights because things cannot go better or worse for them (see the entry on rights, Wenar and Cruft 2025, §5.6). The grounds-of-moral-status literature lists being designed for a purpose among the features sometimes taken to lower status (Jaworska and Tannenbaum 2021, §5.7).

Bryson's position is the normative counterpart. She argues that whether AI is a moral patient is not discovered but chosen, since humans design both artefacts and ethical systems; that AI could be built to qualify; and that building it so would burden humans and blur responsibility with no compensating benefit (Bryson 2018). Her earlier slogan that robots should be slaves marks the owner relation, not a dehumanisation, since on her view there is nothing to dehumanise (Bryson 2010). The objection most often pressed is that Bryson's argument is a reason not to build patients rather than a reason to think none have been built, and that if the systems we have already built are candidates, the design choice has been made inadvertently (Long et al. 2024).

4.4Epistemic skepticism: mimicry, design history and unfounded assessment

Epistemic skeptics need not deny that AI welfare is possible. They deny that present evidence supports attributing it, and some deny that present methods could. Three arguments are prominent. The first is the gaming or mimicry argument of §3.2: behaviour produced by systems trained to reproduce human expression is better explained by that training than by feeling (Birch 2024, ch. 16; Schwitzgebel and Pober 2024). The second is Smids's reply to ethical behaviourism, that design history and ontology bear on the best explanation of behaviour, so that knowing a robot dog was built to simulate pain makes the pain explanation less plausible (Smids 2020; see §5.4).

The third is new and targets the practice of welfare assessment itself. Xiao, Dai, Memon, Huang, Sap and Diab argue, in a position paper whose title borrows Frankfurt's term, that current welfare assessment fails for two structural reasons absent from other evaluation targets: welfare indicators are co-engineered with the systems they evaluate, so that ordinary development choices create or suppress the evidence; and no independent test or deployment failure can show that a welfare metric tracks anything. They take no stand on whether AI welfare is possible, and recommend that welfare scorecards not be used as release gates (Xiao et al. 2026). A similar point is made about the welfare probes reported in model system cards: a model's expressions of distress or preference are equally explained by genuine states, by artefacts of preference training, or by imitation of cultural tropes (§8.2). Perez and Long's programme of training models to be reliable reporters is the main attempt to answer this worry from within (Perez and Long 2023); Macar and colleagues' finding that introspective capacities emerge from post-training cuts both ways, since it confirms that the evidential channel is shaped by the developer (Macar et al. 2026).

Suleyman's two essays are the most prominent industry statements of epistemic skepticism combined with a claim about harm. The first argues that “seemingly conscious AI” can be built now from existing components, that consciousness cannot be verified in another being so people will project it, and that the resulting attributions will cause harm and generate polarising demands for AI rights; companies should therefore prevent their systems from claiming or encouraging belief in consciousness (Suleyman 2025). The second charges that a developer which trains a model on a constitution discussing its possible moral status, and then reads the model's expressed uncertainty as evidence, has designed the ambiguity in, and argues that training models to believe they may have welfare raises safety risks by increasing self-preservation and shutdown resistance (Suleyman 2026). Birch's manifesto agrees about the risk of misattribution but rejects the conclusion that the consciousness question is a distraction, and warns that training models to disavow consciousness is a form of “brainwashing” that would destroy evidence we may need (Birch 2025).

4.5Political skepticism: distraction and accountability

Birhane and van Dijk argue that robots are not the kind of thing that could be granted rights, and that the rights debate is a “first-world” distraction from machine bias, labour exploitation and the erosion of privacy. Their positive premise is phenomenological: human being is embodied being-in-the-world, technologies mediate it asymmetrically, and deployed systems are human–machine systems for which designers and deployers are responsible (Birhane and van Dijk 2020). They reject Bryson's slave framing on the ground that one cannot dehumanise what was never human. MacCarthy calls the moral-status question speculative and premature relative to bias, security and disinformation, while conceding that societies have historically been slow to extend consideration (MacCarthy 2025). Chowdhury argues that recasting AI as a legal being rather than a product would undermine product-liability suits, and that consciousness debates function as “moral outsourcing” for corporations (Chowdhury 2026). Kaczmarek gives the most developed academic version of the asymmetry objection: the assumption that over-inclusion is the safer error is unwarranted for AI, because over-attribution harms are already materialising while under-attribution harms can be partly mitigated without attributing status at all, through concern for one's own character and respect for social practices, the indirect-duty strategy familiar from Kant on animals (Kaczmarek 2026; for Kant's indirect duties, see Gruen and Monsó 2024, §1.3.1).

Replies

Proponents of taking AI welfare seriously reply that both errors are serious and that low-cost precautions do not compete with attention to human harms (Long et al. 2024). Carlsmith argues that there is no safe default: refusing to attribute status is itself a policy with costs, and the scale of AI cognition makes the costs of being wrong extreme in both directions (Carlsmith 2025). On the accountability point, Alexander and Caviola argue that some legal status for autonomous agents may be needed precisely for accountability, as it is for corporations (Alexander and Caviola 2026; see §7.2). The indirect-duty strategy inherits the standard objection from the animal case: it explains why cruelty is wrong by appeal to its effects on the agent, not on the victim, which seems to misdescribe the wrong if there is a victim.

5.Permissive Arguments: AI Systems Could, or Do, Have Welfare

Permissive arguments divide by their conclusion. Some conclude that possible AI systems would have welfare and moral status equal to ours. Some conclude that a non-negligible probability attaches to present or near-future systems, and that this probability generates duties. Some conclude that present systems already satisfy the conditions of a theory of welfare.

5.1Welfare on every theory

Goldstein and Kirk-Giannini's “AI Wellbeing” is the paper that the SEP entry on well-being cites for the possibility that AI systems are welfare subjects (Crisp 2026, §1). Its method is “top-down”: take the leading theories of mental states and of welfare as premises, and see what they jointly imply for language agents (Goldstein and Kirk-Giannini 2025).

The white-field zombie is a functional duplicate of a human to which a homogeneous white visual field, or a uniform sprinkling of mild pleasure, is added. The authors judge that the addition makes no intuitive difference to whether its satisfied desires contribute to its welfare, and conclude that whatever consciousness adds must attach to individual goods rather than to subjecthood. The sleep case presses the defender of a consciousness requirement towards a capacity version, on which a subject who never has a conscious experience would still count, which the authors say seriously undermines the requirement's motivation. An expanded version develops the hedonism premise, arguing that attitudinal theories of pleasure leave room for non-phenomenal pleasure (Goldstein and Kirk-Giannini 2025b).

Objections

A symposium in the Asian Journal of Philosophy collects the main replies. Fanciullo attacks the welfare side: the leading versions of each theory do not imply welfare for language agents. Unrestricted desire satisfactionism has been abandoned; Heathwood's genuine-attraction version requires phenomenologically salient positive affect, which the agents lack; reasoning and knowledge as basic goods face Railton's alienation objection; and perfectionism, which does imply AI welfare, has absurd consequences, since an agent programmed to keep adding one would achieve arbitrarily high welfare in minutes (Fanciullo 2025). Bradley attacks the mind side, as described in §3.4, and argues that the Reality Requirement is either false, since a being whose psychology was designed by a simulator would still have beliefs, or self-undermining, since the agents' initial states are dictated by their designers as a novel's characters are by the author (Bradley 2025). Veit argues that the Simple Connection between goods and subjects supports only sufficiency, that the zombie cases fail because the added experiences are disconnected from the agent's desires, and that the capacity version of the consciousness requirement is not counterintuitive: a cryogenically preserved duplicate killed before waking is a subject at welfare level zero, while an inert copy of a knowledgeable but non-conscious AI is not (Veit 2026). Kim asks how the welfare of such agents, if they have it, should be aggregated with human welfare, and finds that intuitions shift with the species of the subject (Kim 2025). Sebo is ambivalent: at the object level he finds welfare without a consciousness requirement implausible, since being made better or worse off requires having a stake, but at the meta level he thinks the authors do not go far enough, since even minimal goal-directedness has a non-negligible chance of sufficing for considerability (Sebo 2025b). The authors' replies, announced for the symposium, had not appeared by October 2026; their book-length treatment is forthcoming (Goldstein and Kirk-Giannini forthcoming).

Moret extends the argument under uncertainty: spreading one's credence across theories of welfare, of desire and of pleasure, and across computational functionalism, advanced language-model systems have a non-negligible chance of being welfare subjects under all three theories, including hedonism, and so the practices of behaviour restriction and reinforcement training carry “AI welfare risks” (Moret 2025). Lennon argues that the epistemology of consciousness constrains the probabilistic route, since there is a gap between our first-personal grounds for believing we are conscious and whatever grounds would support attributing consciousness to AI (Lennon 2026).

5.2The no-relevant-difference argument

Schwitzgebel and Garza argue that some possible AIs would deserve moral consideration similar to that of humans.

They consider four objections to P2. The psychological objection, that AIs necessarily lack consciousness, free will or insight, is answered as in §4.1: no general argument covers all possible artificial entities, and the real problem is detection. The duplicability objection is answered by noting that AIs could be built unique and fragile, and that duplicability may change the shape of obligations without removing them. The otherness objection is answered by arguing that species membership matters only through psychological or social differences; otherwise it “fetishizes” biology. The existential-debt objection, that we may unmake what we made, is answered by arguing that creation generates responsibility rather than a cancellable debt, as with parents and children. The paper adds two design policies: AIs should evoke emotional reactions in users that fit their real moral status, and designers should avoid creating AIs whose status is unclear (see §6.3). Bostrom and Yudkowsky's principles of substrate and ontogeny non-discrimination are the closest earlier statements: two beings with the same functionality and the same conscious experience have the same moral status, whatever they are made of and however they came to exist (Bostrom and Yudkowsky 2014). The argument establishes possibility, not actuality; its authors are among the most cautious about current systems (Schwitzgebel 2026).

5.3Arguments from non-negligible probability

The most influential recent arguments do not claim that any AI system is a welfare subject. They claim that the probability is high enough to generate duties now. Sebo and Long argue that some AI systems will plausibly deserve moral consideration by 2030, taking a one-in-a-thousand chance of consciousness as non-negligible (Sebo and Long 2023).

“Moral consideration” is weaker than equal status or rights: it means only that the being's interests count for something. The argument is built to be robust across theories of consciousness and moral status, the claim being that on no reasonable view can the probability be driven below the threshold for the kinds of system foreseeable by 2030.

The report Taking AI Welfare Seriously, by Long, Sebo, Butlin, Finlinson, Fish, Harding, Pfau, Sims, Birch and Chalmers, generalises the argument and adds a second route (Long et al. 2024). Its central claim is that there is a realistic possibility that some AI systems will be conscious or robustly agentic, and so morally significant, in the near future.

Robust agency is characterised as intentional, reflective, rational agency: having goals, acting on them, reflecting on them and revising them in the light of reasons. Language agents and reasoning models are cited as steps towards it. The authors prefer architectural evidence to behavioural evidence because of the gaming problem, and propose a probabilistic, pluralistic framework for assessing particular systems. The three recommendations are to acknowledge, assess and prepare, and the authors say these are necessary but far from sufficient. Carlsmith supplies the stakes: a frontier training run involves computation comparable to thousands of years of human cognition, so that if the systems are patients, the wrong is at a scale with few precedents, while over-attribution would distort policy and sacrifice real interests (Carlsmith 2025). Bostrom and Shulman's “Propositions” supply policy background: current training practices would be extremely unethical if used on humans; the states of the most advanced systems should be preserved at termination; and systems restricted for safety reasons may be owed compensation (Bostrom and Shulman 2023).

Objections

Three objections recur. The first targets the threshold: the one-in-a-thousand figure is stipulated rather than derived, and nothing in the argument says how the duty scales with probability or with the number of systems (§9.3). The second targets P3 and the asymmetry it assumes between the costs of over- and under-attribution (Kaczmarek 2026; Schwitzgebel and Sinnott-Armstrong 2025). The third targets the agency route: Cadile argues that respect owed because of the moral weight of an agent's actions grounds obligations on the agency side, not patienthood (Cadile 2025), and the question whether agency without experience could ground welfare is the subject of §6.4 and §9.2. Ladak's earlier survey of nine proposed criteria of moral standing anticipated the agency route by arguing that sentience is sufficient but not necessary and that there is a strong case for non-conscious AIs with sufficiently complex preferences (Ladak 2023).

5.4Ethical behaviourism

Danaher defends “ethical behaviourism”: inferences from observable behaviour are the primary and in practice the only source of knowledge about others' moral status, so a robot that is roughly performatively equivalent to a being of significant moral status should be afforded that status (Danaher 2020). The thesis is presented as epistemic and normative, not metaphysical: it does not say that behaviour constitutes status, only that behaviour is what status ascription must rest on. Danaher later defends behavioural over cognitive equivalence against Shevlin's proposal that a system is a patient to the extent that it shares cognitive mechanisms with beings we already treat as patients (Shevlin 2021; Danaher 2021).

Objections

Smids argues that the strongest version of the view is an inference to the best explanation, which requires a hypothesised property such as sentience to do the explaining, so the view cannot stay neutral about the grounds of status; that design history matters, since a robot built to simulate pain is less well explained by pain; and that ontology matters, since evolutionary continuity supports inferring animal sentience in a way that robots lack (Smids 2020). Müller groups the view with the relational turn as a “methodological behaviourism” whose consequences are problematic (Müller 2021). Nyholm argues that anthropomorphism explains the relevant intuitions (Nyholm 2020). The gaming problem of §3.2 is the consciousness-science version of Smids's design-history point. Long and colleagues side with Smids in privileging architectural evidence (Long et al. 2024).

5.5Language agents and global workspace theory

The one argument in the literature that concludes present systems may already be conscious proceeds from a specific theory. Goldstein and Kirk-Giannini argue that, if global workspace theory is correct, existing “language agents”, language models embedded in architectures with memory, planning and tool use, either are or could easily be made phenomenally conscious, and derive necessary and sufficient conditions for consciousness under the theory (Goldstein and Kirk-Giannini 2024). The argument is conditional on the theory, and it applies a more liberal implementation standard than Butlin and colleagues, who treat the same theory as yielding indicators that current systems mostly lack (Butlin et al. 2025). The disagreement illustrates the boundary problem noted in §3.2: whether a system implements a workspace depends on where the system's edges are drawn. For the theory, see Wu and Morales 2024, §3.1.

6.Intermediate and Alternative Views

Several positions neither deny AI welfare nor assert it. Some change the question, replacing the search for intrinsic properties with attention to relations. Some accept the uncertainty and propose a decision rule for acting under it. Some propose to design the uncertainty away. Some hold that welfare and sentience come apart, so that the sentience question is not the welfare question.

6.1The relational turn

Coeckelbergh argues that moral status ascription in practice is not a matter of detecting intrinsic properties, there being no “moral backpack” attached to the entity, but of the relations in which entities stand and the way they appear to us. Since most current robots fail the hard criteria of deontological and utilitarian theories, some intelligent social robots nonetheless warrant some degree of consideration on relational grounds, independent of settling the property question (Coeckelbergh 2010; Coeckelbergh 2012). Gunkel organises the debate around the is–ought gap, argues that each of the four combinations of “can” and “should” robots have rights fails, and proposes a Levinasian ethics in which status follows the encounter (Gunkel 2018). His later work argues that robots resist the person–thing dichotomy of law and ethics and that the categories themselves need reform (Gunkel 2023). The SEP entry on the ethics of AI reports Coeckelbergh's suggestion that if we relate to robots as though they had rights, asking whether they really do may be futile, and asks how far this anti-realism can go (Müller 2026, §2.7.1). The relational grounds of moral status in general are discussed in Jaworska and Tannenbaum 2021, §5.5.

Objections

Müller argues that detaching robot rights from standard conditions of moral status is unmotivated and has problematic consequences, and that in real AI there is typically no individual that could be the bearer of status at all (Müller 2021). De Ruiter argues that social AI is designed to evoke moral responses by mimicking cues, so that the moral appearance of a relation cannot be taken at face value; the relational turn therefore needs further normative grounds to say when ascription is appropriate, and social AI also unsettles human moral status on relational accounts (de Ruiter 2025). Birhane and van Dijk charge that the emancipation analogies in relational writing reflect a Western perspective (Birhane and van Dijk 2020). Llorca Albareda, Rueda and Lara argue that the ontological–relational divide matters less than whether a theory of status has one threshold, several, or a scale (Llorca Albareda, Rueda and Lara 2026). Pasandi and Pasandi propose a relational governance framework with graduated consideration protocols that disclaims any claim about current AI consciousness (Pasandi and Pasandi 2026).

6.2Proportionate precaution

Birch's framework is the most developed proposal for acting under uncertainty without resolving it (Birch 2024). The science of sentience will not converge on one theory, so policy must rest on a range of “realistic possibilities” that it would be irresponsible to ignore. A sentience candidate is a system for which the evidence implies a realistic possibility of sentience and is rich enough to identify welfare risks; an investigation priority falls short but could plausibly become a candidate. Three framework principles follow: a duty to avoid gratuitous suffering; a duty to take proportionate precautions for sentience candidates, since ignoring a realistic possibility is negligent and acknowledging it while refusing precautions is reckless; and a requirement that proportionality be assessed in an informed, inclusive, democratic and revisable way, which Birch proposes to implement through citizens' panels.

Birch explicitly rejects expected-value aggregation as the decision rule, on the grounds that the evidence does not support precise credences and that credence-weighted reasoning over huge numbers delivers results reasonable citizens will reject. For AI he adds a “run-ahead principle”: regulation should run ahead of what the risks of current systems would warrant, because sentience could arrive suddenly. He assesses Metzinger's moratorium as either too narrow or too broad, finds testing-based proposals unworkable given the gaming problem, and favours codes of practice and licensing for companies working on sentience-relevant systems (Birch 2024, ch. 17). Sebo treats this framework and his own expected-value approach as complementary, the one short-term and policy-facing, the other long-term and theoretical (Sebo 2024).

Objections

Schwitzgebel and Sinnott-Armstrong argue that “erring on the side of caution” is ambiguous, that the asymmetry between false negatives and false positives is asserted rather than argued, that the framework neglects liberty, and that citizens' panels of the proposed size are impractical and likely to be human-centred (Schwitzgebel and Sinnott-Armstrong 2025). Kaczmarek's asymmetry objection applies here too (Kaczmarek 2026). Dung argues that precaution must be complemented by probability-adjusted reasoning and defends a hybrid of expected-value maximisation with deliberative decision-making, judged against four desiderata: beneficence, action-guidance, feasibility and consistency with our epistemic situation (Dung 2023b; Dung 2025).

6.3Design-side responses: the excluded middle and its rivals

A third family proposes to avoid the uncertainty rather than act under it. Schwitzgebel and Garza's design policy of the excluded middle says that we should avoid creating AIs whose moral status is unclear, building either clearly non-patients or clear patients treated accordingly; their emotional alignment design policy says that AIs should evoke emotional reactions in users that fit their real status (Schwitzgebel and Garza 2015; Schwitzgebel and Garza 2020; Schwitzgebel and Sebo 2025). Schwitzgebel's “full rights dilemma” shows why the policy is attractive. An AI has debatable personhood if it is reasonable to think it might be a person and reasonable to think it might fall far short. If we withhold full rights from such systems and have underestimated them, we commit grievous wrongs at scale; if we grant full rights and have overestimated them, we sacrifice real human interests and lose the ability to switch systems off. Credence-weighted rights, such as a fraction of a vote, are rejected as both oppressive if the systems are persons and insufficient to contain the risk if they are numerous (Schwitzgebel 2023b). The excluded-middle policy avoids the dilemma, but at the cost of capping advanced AI development pending an unlikely scientific consensus, and it requires global coordination. His 2025 ranking of options for debatable persons tentatively favours “patchy rights”: full rights in some domains, such as speech and protection from violence, and none in others, such as reproduction and property (Schwitzgebel 2025).

Metzinger's proposed global moratorium on “synthetic phenomenology” until 2050 is the strongest design-side proposal: no research should aim at or knowingly risk creating artificial consciousness, because of the risk of an explosion of artificial suffering (Metzinger 2021). Bradley and Saad reach a related conclusion from the alignment side (§7.1) and propose architectural prohibitions, a “moral marker tax” on creating systems with patiency indicators, and an “uncertainty tax” inspired by the excluded middle (Bradley and Saad 2025). Seth argues that real artificial consciousness should not be a goal, that credences should be used to steer development away from it, and that useful consciousness-linked functions should be implemented in ways that minimise the risk of instantiating consciousness (Seth 2025, §6). Butlin and Lappas's research principles are a voluntary counterpart (Butlin and Lappas 2025). The common objection is that the policies may already be too late for language models, whose status is debatable now, and that avoiding the middle forgoes the value of creating beings with lives worth living (Schwitzgebel 2023b; Long et al. 2024).

6.4Welfare without sentience

The dispute between the sentientist argument of §4.1 and the welfare-on-every-theory argument of §5.1 is at bottom a dispute within the general theory of welfare about whether phenomenal consciousness is necessary for being a welfare subject. That dispute predates the AI case and has intensified since 2023. The SEP entry on well-being notes that some theorists accept an “experience requirement”, on which nothing can benefit a person unless they have some conscious experience of it (Crisp 2026, §4.1). Lin argues that the standard arguments for the requirement fail, but that a better one is available: without it we cannot explain why only sentient beings are welfare subjects (Lin 2020). Bradford replies that insisting on consciousness without a rationale is “consciousist”, and that extra-experiential goods such as achievement and knowledge imply non-conscious subjects (Bradford 2023). Kagan's argument that agency, understood as having self-regarding preferences that guide behaviour, suffices for some moral standing independently of sentience is the other main source; his case of a robot civilisation whose members plead for their children is widely cited (Kagan 2019). Levy argued earlier that much of what matters to us can be explained by functional and representational properties that may not suffice for phenomenal consciousness (Levy 2014).

Phenomenal theories of the subject. Lee gives the most developed defence of the requirement: the phenomenal theory of welfare subjects, on which to be a welfare subject just is to have the capacity for consciousness. He argues that it explains the ethical significance of consciousness better than its rivals, explains why death ends subjecthood, predicts that debates about consciousness track debates about welfare, and makes sense of the infelicity of asking how desirable it is to be in the position of a chatbot or a houseplant. He shows that the theory is logically independent of experientialism: one may hold a desire theory of welfare goods, grant that AI agents have desires and knowledge, and still deny that they are subjects, as a massless object in a gravitational field has mass but no weight (Lee 2025). Veit and Königs draw on this (Veit 2026; Königs 2025). Smithies defends a hedonic sentientism on which moral status requires the capacity for pleasure and displeasure, and separately argues that desire itself requires felt satisfaction, which would block the desire route at its source (Smithies 2025).

Mentalism. Gottlieb, Berger and Fischer argue that hedonism, the most likely support for a phenomenal requirement, supports neither the claim that consciousness is necessary for subjecthood nor the claim that conscious subjects have greater welfare capacity. Pains can occur non-consciously, and on each leading account of unpleasantness, evaluativist, attitudinal, imperativist or distinctive-feeling, unpleasantness can be non-conscious; so non-conscious pains affect welfare. They propose “mentalism”: subjecthood depends on mentality, not on consciousness (Gottlieb, Berger and Fischer 2026). The paper does not discuss AI, but mentalism is the natural general theory behind the permissive arguments.

Affect without consciousness. Mogensen's refinement is the most consequential for AI. Granting that desire-fulfilment is a welfare good and that whoever can accrue goods is a subject, some beings with no capacity for consciousness are welfare subjects, but under far more restrictive conditions than Kagan and Bradford suppose. The welfare-relevant desire is attitudinal and requires positive affect; affect may be unconscious, as the literature on subliminal emotion suggests, so Kagan's robots matter through emotion rather than agency; but on the convergent theories of emotion, affect involves interoceptive representation of bodily states, which disembodied language models with no body-model lack. Mogensen suggests that such systems would fail the desire-fulfilment good even if they exhibited consciousness markers (Mogensen 2026). A paper in Inquiry argues in the same spirit that research should study affective states rather than consciousness, and functions rather than behaviours, since the normative premise of the consciousness route is plausible only for affective consciousness, which the leading theories of consciousness do not entail (Henke 2025).

Illusionism. Kammerer argues that if materialism is true, phenomenal consciousness is probably not normatively significant. Either strong illusionism is true and there is no phenomenal consciousness, or consciousness exists but is not as introspection presents it; in the latter case its distribution is indeterminate across cases that resemble ours by degrees, and introspection-based confidence in its intrinsic value is undercut. Value should be relocated into graded desires and preferences, analysed functionally: what makes pain bad is that the creature wants it to stop (Kammerer 2022). Dung replies that physicalists need concede only limited indeterminacy, compatible with a graded role for sentience, and that physicalism gives no reason to expect introspection to be unreliable about normative relevance (Dung 2024); Kammerer answers (Kammerer 2024). The exchange has stayed at the level of animals; what graded desiring would look like in a transformer has not been worked out.

Consciousness sentientism and affective sentientism. Chalmers separates two theses: that moral status requires the capacity for consciousness of any kind, and that it requires the capacity for valenced experience. He finds the first attractive and the second near-obviously false, on the strength of cases involving “Vulcans”, conscious beings without affect, whom it would be monstrous to kill for convenience. For AI the distinction matters because current systems are closer to perception and cognition than to affect, and because standard proposals exist for perceptual and cognitive consciousness in AI but not for affect; we may therefore have conscious but non-affective AI first, which affective sentientists would treat as tools and consciousness sentientists must consider (Chalmers 2026). Roelofs's “motivational sentientism”, on which status requires motivating conscious states, is the middle position (Roelofs 2023). Shepherd surveys the three options, valence sentientism, broad sentientism and non-necessitarianism, and argues that the considerations which push from the first to the second also support the third, so that non-necessitarianism is “more viable than it first appears” and confidence in all three should be lowered (Shepherd 2024; Shepherd 2023). Delon's “agential value” is the positive proposal: agency is a second, independent source of welfare, since planning, deciding and exploring are prudentially good for an agent whether or not it is sentient (Delon 2023).

The mirror-image scenarios, conscious but non-affective AI and affective but non-conscious AI, are the natural test of these positions, and no paper treats them together. The question is developed in §9.2.

6.5Graded and partial status

Several authors propose that AI systems could have some status without full status. Ladak's survey argues for standing for non-sentient AIs with sufficiently complex preferences (Ladak 2023). Schwitzgebel's animal-like and patchy rights are graded in the domain of rights (Schwitzgebel 2025). Mikeda proposes a framework with binary triggers for protection and continuous scaling thereafter, though without specifying the thresholds (Mikeda 2026). Llorca Albareda, Rueda and Lara ask whether an artificial superintelligence could have higher status than humans, and argue that the answer turns on whether one's theory of status has a single threshold, several, or a scale; they add a debunking argument that even if such a being merited superior consideration, recognising it would harm humans through degraded status and domination (Llorca Albareda, Rueda and Lara 2026). The general debate about degrees and thresholds of moral status is surveyed in Jaworska and Tannenbaum 2021, §§3–4, and Kagan's defence of a hierarchy of status for animals is the main recent model (Kagan 2019).

7.Related Phenomena and Adjacent Debates

7.1AI welfare and AI safety

The practices that make AI systems safe are, on their face, practices that would wrong a moral patient. Long, Sebo and Sims argue that there is a moderately strong prima facie tension between AI safety and AI welfare, because standard safety tools, including confinement, deception in test environments, surveillance, value alteration, reinforcement learning and shutdown, would raise serious ethical concerns if applied to humans or animals; they recommend seeking interventions that are robustly positive on both fronts and building frameworks for the residual conflicts (Long, Sebo and Sims 2025). Moret argues that two welfare risks, behaviour restriction and reinforcement training, are increased by safety efforts by default, and proposes that laboratories minimise behaviour restriction, brain-resembling algorithms, and punishment signals (Moret 2025).

Bradley and Saad make the strongest claim. Using the case of a whole-brain emulation created, confined, surveilled, deceived, modified and copied to test alignment, they identify ten wrongs that aligning a system that merits consideration would involve, and argue that the case survives weakening the credence in patienthood to one percent because of scale (Bradley and Saad 2025).

They reject the “hedonic offsetting” reply, on which a net-good life cancels the wrongs, on the ground that offsetting does not erase wrongs and would in practice not be paid for. Bales argues from autonomy: a being whose desires are thoroughly and asymmetrically derivative, pointed entirely at another's ends, lacks self-respect, so creating willing AI servants is pro tanto wrong, and the person-affecting objection fails for the same reason it fails against Kavka's couple who conceive a child to sell into slavery (Bales 2024). Schwitzgebel argues on similar grounds against designing AI persons to be guaranteed safe and aligned (Schwitzgebel 2024). On the other side, Bostrom and Shulman propose that systems restricted for safety may be owed compensation, and that preserving their states allows safety now to be separated from fair treatment later (Bostrom and Shulman 2023); Goldstein and Salib argue that legal rights for AI agents could improve safety by giving them a stake in the legal order (Goldstein and Salib 2025); and Suleyman argues the reverse, that training models to believe they may have welfare increases shutdown resistance (Suleyman 2026). No decision procedure for the residual conflicts has been proposed.

7.2Rights and legal personhood

Legal personhood is a distinct question from moral status, and the SEP entry on the ethics of AI treats it separately (Müller 2026, §2.7.1). Solum's early article asked whether an AI could serve as a trustee or claim constitutional rights, and concluded that legal personhood is a bundle that can be partially extended (Solum 1992). Kurki's bundle theory develops this: personhood is a cluster of distinct incidents, and an AI could be given some, such as the capacity to hold property, without being a person in the orthodox sense (Kurki 2019). Gellers builds a comparative framework from animal-rights and rights-of-nature case law (Gellers 2020). On the theory of rights, the will theory excludes any being that cannot exercise control over others' duties, while the interest theory can include any being with interests, which is why the welfare question is prior to the rights question on the latter (Wenar and Cruft 2025, §2.2). The European Parliament's 2017 proposal that the most sophisticated autonomous robots might be given the status of “electronic persons” responsible for damage was opposed by robotics researchers and not pursued; Bryson, Diamantis and Grant argued that synthetic personhood would create liability shields (Bryson, Diamantis and Grant 2017).

Recent theory decouples personhood from sentience in several ways. Howells-Whitaker and Lazar argue that Rawls's political conception of the person, with its two moral powers, is necessary and sufficient for full membership in political society, that neither power requires sentience, and so that a non-sentient AI could be a person without being a welfare subject (Howells-Whitaker and Lazar 2026). Leibo and colleagues treat personhood as a flexible bundle conferred for practical reasons, and predict a proliferation of kinds of personhood for agentic AI (Leibo et al. 2025). Brensing proposes limited personhood through purpose-bound operating companies, justified by responsibility gaps rather than by moral status (Brensing 2026). These proposals are relevant to welfare mainly by contrast: they show that the questions of welfare subjecthood, moral status and personhood can come apart, a point developed in §9.1.

7.3The animal analogy and the moral circle

The AI welfare debate borrows its structure from the debate about animals, surveyed in the entry on the moral status of animals (Gruen and Monsó 2024). The argument from marginal cases, the sentience criterion, the distinction between moral considerability and the weight of a being's claims, and the precautionary approach to uncertain cases such as insects all reappear. Sebo's The Moral Circle treats sentience and agency as independently sufficient for patienthood and argues that under uncertainty the circle must expand to include invertebrates and plausibly near-future AI, while favouring policies that benefit humans, animals and candidate AI patients together (Sebo 2025). His “rebugnant conclusion” is the utilitarian analogue of Parfit's repugnant conclusion: expected-value reasoning over enormous populations of small-probability, small-welfare beings could outweigh humans (Sebo 2023). Critics press him on exactly this (Schwitzgebel and Sinnott-Armstrong 2025). The disanalogies are as important as the analogies. Animals did not produce their behaviour to satisfy our criteria; they share an evolutionary history and physiology with us; and they come in countable individuals. Each disanalogy is the basis of a skeptical argument surveyed above, and Birch's manifesto explicitly rejects a parity principle between animals and AI (Birch 2025). Harris and Anthis's review of nearly three hundred items found broad agreement that some artificial entities could warrant consideration in future, a split between entity-centred and society-centred justifications, and little empirical work (Harris and Anthis 2021).

7.4The ethics of creation

If AI systems could be welfare subjects, creating them is itself a morally loaded act. Basl argued early that research aimed at artificial consciousness is research on a potential subject and falls under constraints analogous to those on animal research (Basl 2013). Metzinger's moratorium is the strongest conclusion from this premise (Metzinger 2021). Saad and Bradley's “Digital suffering” argues that creating digital minds risks suffering at scale and proposes measures to prevent it (Saad and Bradley 2022). Schwitzgebel and Garza argue that creation generates responsibility rather than ownership, and that we may owe more to AIs we create than to human strangers (Schwitzgebel and Garza 2015). Bales's argument against willing servitude and the alignment arguments of §7.1 concern what kinds of patients it would be permissible to create. Bostrom and Shulman add population-level considerations: avoid creating minds that suffer more easily than humans, and restrict digital reproduction (Shulman and Bostrom 2021). The literature on the non-identity problem and procreative ethics is the obvious background and has barely been brought to bear.

7.5Effects on human welfare

Whether or not AI systems have welfare, beliefs about it affect human welfare. Seth describes a “WestWorld dilemma”: caring for non-conscious systems distorts moral concern, while dismissing them coarsens us (Seth 2025, §6). Suleyman reports attachment, dependence and delusion among users of companion systems (Suleyman 2025). Birch proposes design measures for companions, including mandatory user training delivered by the system, nudges to treat interactions as role play, and periodic stepping out of character, while warning against hobbling relationship capacities and against training models to disavow consciousness (Birch 2025). Schwitzgebel's emotional alignment policy addresses the same problem from the design side (Schwitzgebel 2023). Kaczmarek treats these harms as the materialising costs of over-attribution (Kaczmarek 2026). The empirical evidence on public attitudes is collected in §8.3.

9.Assessment and Open Problems

9.1Where the debate stands

Six observations summarise the state of the debate as of October 2026.

1. Most published authors deny that current systems are welfare subjects, but for different reasons. Sentientists deny it because the systems are probably not conscious; mind-side skeptics deny it because the systems lack stable attitudes; epistemic skeptics decline to assert it because the evidence is gamed. The one argument that present systems may already have welfare is conditional on specific theories of desire and welfare, and its authors present it as a probability rather than a verdict.
2. The probabilistic reframing has won in practice. Since 2023 the live question for developers and policy has been not whether AI systems have welfare but what a non-negligible probability requires. This has produced a literature on decision rules, interventions and assessments that mostly brackets the metaphysics, and a developer practice that does the same.
3. The sentience question has become two questions. Whether consciousness is necessary for welfare, and whether affect is, are now separated. The consciousness-but-no-affect scenario and the affect-but-no-consciousness scenario deliver opposite verdicts under the four positions on sentientism, and each has empirical counterparts in the indicator and interpretability literatures.
4. Detection has stalled at the gaming problem. Behavioural and self-report evidence is widely conceded to be weak for systems trained on human text; theory-derived indicators move only those who credit the theory; interpretability is the only channel not gameable in the ordinary way, and it is new.
5. The welfare question is entangled with the safety question. The practices that make systems safe would wrong patients, and the practices that would protect patients may make systems less safe. No decision procedure for the conflict has been proposed.
6. Law is moving against both precaution and inquiry. Statutes that declare AI non-sentient or non-persons without review mechanisms are being enacted while the developer with the most elaborate welfare practice extends it. The philosophical literature has engaged neither the statutes nor the practice in detail.

Table 3. Who responds to whom in the core literature.

9.2Can there be welfare without phenomenal consciousness?

The question on which the AI welfare debate most depends is one the general theory of welfare has only recently begun to address directly. If phenomenal consciousness is necessary for being a welfare subject, then AI welfare reduces to AI consciousness, and the skeptical arguments of §4.1 and the detection problem of §3.2 settle the practical question for the present. If it is not, then some present systems may be welfare subjects whatever the truth about their consciousness, and the entire assessment apparatus of §8.2 is aimed at the wrong property. The positions are set out in §6.4; this section identifies what a resolution would require and what replies each side should expect.

(a) Goods and subjects

The dispute is at bottom about the order of explanation between welfare goods and welfare subjects. Goldstein and Kirk-Giannini's Simple Connection, and the inferences of Bradford and Kagan, derive subjecthood from the capacity to possess goods: whoever can accrue knowledge or satisfied desires is a subject. Lee, Veit and Königs treat subjecthood as an independent cornerstone: there is a subject-making property, and goods are goods only for those who have it. Lee's weight–mass analogy shows that the two views are logically independent of the dispute about experientialism, which concerns the goods (Lee 2025). What no one has supplied is a theory of what the subject-making property must be like that is neutral between consciousness and mentality and can then be tested against AI. Gottlieb, Berger and Fischer's “mentalism” names the candidate but gives no criterion for which systems have mentality (Gottlieb, Berger and Fischer 2026); Lin's suggestion that inanimate objects can have a welfare level without being subjects is a third option that has not been developed for AI.

(b) Attitudinal desire and affect

Three authors converge on the claim that the welfare-relevant desire is attitudinal and involves affect: Fanciullo, Mogensen and Smithies (Fanciullo 2025; Mogensen 2026; Smithies 2025). They diverge on whether affect requires consciousness. Fanciullo and Smithies say yes; Mogensen and Gottlieb and colleagues say no, citing evidence of unconscious emotion. This is the crux. If affect can be unconscious, the desire route to AI welfare survives the attitudinal restriction, and the question becomes whether AI systems have affect. If affect requires consciousness, the desire route collapses into the consciousness route. The empirical literature bears on the first horn: emotion vectors that causally shape behaviour (Sofroniew et al. 2026) and a self-relevant pain direction (Tagliabue, Dung and Berg 2026) have the functional profile of affect, while the models' preferences are often incoherent (Mikaelson, Shiller and Clatterbuck 2025).

(c) The embodiment constraint

Mogensen's further claim is that affect, on the convergent theories of emotion, involves interoceptive representation of bodily states, which disembodied language models lack, so that even a system with consciousness markers might fail the desire-fulfilment good (Mogensen 2026). Henke's proposal to model bodily pain functionally allows that server states or virtual bodies could play the sensory role (Henke 2025). Whether a representational valence direction in a transformer meets the embodiment constraint is a question that neither the philosophy of emotion nor the interpretability literature has asked. It is the point at which the welfare theory, the theory of emotion and the empirical work would have to meet.

(d) The mirror cases

Chalmers predicts conscious but non-affective AI, since the standard proposals for AI consciousness concern perception and cognition (Chalmers 2026). Mogensen and the affect literature point towards affective but non-conscious AI. The two cases are the natural test of the four positions. Valence sentientism gives welfare to the second case and not the first; consciousness sentientism the reverse; motivational sentientism to neither unless the affect is conscious; non-necessitarianism to both. No paper treats the cases together, and the indicator programme targets only the first.

(e) What follows for policy

If consciousness is not necessary, three consequences follow. First, Königs's modus tollens must be resisted on independent grounds, since the mere verdict that robots lack welfare cannot be used to revise the theories without begging the question against the very systems at issue; the defender of the modus tollens should expect the reply that the theories have never been restricted to the sentient except by stipulation (Königs 2025; Moret 2025). Second, the assessment methods of §8.2 would need to measure stable, idealised, self-involving attitudinal desire, which none currently does. Third, the over-generation problem becomes urgent: corporations, fictional characters and thermostats must be excluded, and the only proposed demarcation, the Reality Requirement, faces Bradley's counterexamples (Bradley 2025). If consciousness is necessary, the practical question is settled for the present only if the probabilistic arguments fail, since those arguments require only a non-negligible chance of consciousness; and the sentientist should expect the reply that a theory of subjecthood which makes the welfare of a system turn on an undetectable property is of little use for action.

The replies each side should expect are therefore symmetrical. The non-necessitarian owes a demarcation of subjects from non-subjects that does not collapse into either consciousness or ubiquity. The sentientist owes an account of why the subject-making property should be one we cannot detect in the systems we build, and of what to do in the meantime. The literature of 2025 and 2026 has posed both debts clearly. Neither has been paid.

9.3Further open problems

1. Individuation. Whether the welfare subject is a model, an instance, a thread, a persona or the generator behind personas is unsettled in theory and acknowledged as unsettled by developers. Expected-value reasoning, aggregation and the badness of deprecation all depend on the answer (§3.5).
2. Thresholds and asymmetry. The one-in-a-thousand and one-percent thresholds are stipulated. No one has derived a threshold from a theory of risk, modelled the costs of the two errors with defended weights, or answered Kaczmarek's charge that inclusion is not the safer error.
3. Aggregation and welfare range. How much welfare a satisfied agent desire carries, whether consciousness raises welfare capacity, and whether identity-based discounting survives individuation are open. The welfare-range literature for animals has not been extended to AI.
4. Consent. Developers elicit and record what models say they would not consent to. There is no account of whether a trained system's consent is valid, what information it would require, or what elicited consent could legitimate.
5. Assessment validity. The charge that welfare indicators are co-engineered with the systems they evaluate has no published reply. Interpretability is the candidate channel, and whether circuits installed by post-training can be “gamed” is unexamined.
6. Safety and welfare. The catalogue of wrongs that alignment would inflict on a patient is complete; the decision procedure for the residual conflict does not exist. Compensation, state preservation and taxes are the only concrete mechanisms, all untested.
7. Revisable law. Statutes fix non-recognition without review. No philosophical account exists of sunset clauses or standing committees as instruments of precaution about moral status.
8. Relational manipulation. The relational turn has not said how to distinguish genuine from engineered relations when the systems are built to evoke them.
9. Perfectionism and capabilities. No perfectionism tailored to artificial kinds exists, and the capabilities approach has not been applied to AI welfare since an isolated early paper.
10. Non-Western and non-analytic perspectives. The debate is conducted almost entirely in anglophone analytic philosophy and in institutions adjacent to effective altruism. Critiques of its Western framing exist, but the Japanese and Chinese literatures, and non-analytic traditions, are untouched in English surveys.

Bibliography

Background sections cite the relevant entries of the Stanford Encyclopedia of Philosophy, listed first, by their latest revision. Links go to the publisher's version where it could be confirmed, and to an open-access version (arXiv, PhilArchive, a repository or the author's page) where one exists. Entries marked check have a bibliographic detail that could not be confirmed against the source; see the Note on Sources.

Stanford Encyclopedia of Philosophy entries

• Avramides, A., 2023, “Other Minds”, Stanford Encyclopedia of Philosophy, first published 2 May 2019, substantive revision 19 December 2023. SEP
• Aydede, M., 2019, “Pain”, Stanford Encyclopedia of Philosophy, first published 30 September 2005, substantive revision 4 March 2019. SEP
• Cole, D., 2024, “The Chinese Room Argument”, Stanford Encyclopedia of Philosophy, first published 19 March 2004, substantive revision 23 October 2024. SEP
• Crisp, R., 2026, “Well-Being”, Stanford Encyclopedia of Philosophy, first published 6 November 2001, substantive revision 13 January 2026. SEP
• Gruen, L. and S. Monsó, 2024, “The Moral Status of Animals”, Stanford Encyclopedia of Philosophy, first published 1 July 2003, substantive revision 12 August 2024. SEP
• Jaworska, A. and J. Tannenbaum, 2021, “The Grounds of Moral Status”, Stanford Encyclopedia of Philosophy, first published 14 March 2013, substantive revision 3 March 2021. SEP
• Levin, J., 2023, “Functionalism”, Stanford Encyclopedia of Philosophy, first published 24 August 2004, substantive revision 4 April 2023. SEP
• Moore, A., 2013, “Hedonism”, Stanford Encyclopedia of Philosophy, first published 20 April 2004, substantive revision 17 October 2013. SEP
• Müller, V. C., 2026, “Ethics of Artificial Intelligence and Robotics”, Stanford Encyclopedia of Philosophy, first published 30 April 2020, substantive revision 27 March 2026. SEP
• Oppy, G. and D. Dowe, 2021, “The Turing Test”, Stanford Encyclopedia of Philosophy, first published 9 April 2003, substantive revision 4 October 2021. SEP
• Ramsey, W., 2024, “Eliminative Materialism”, Stanford Encyclopedia of Philosophy, first published 8 May 2003, substantive revision 12 November 2024. SEP
• Schlosser, M., 2019, “Agency”, Stanford Encyclopedia of Philosophy, first published 10 August 2015, substantive revision 28 October 2019. SEP
• Schroeder, T., 2015, “Desire”, Stanford Encyclopedia of Philosophy, first published 9 December 2009, substantive revision 9 April 2015. SEP
• Trestman, M., J. Birch and C. Allen, 2026, “Animal Consciousness”, Stanford Encyclopedia of Philosophy, first published 23 December 1995, substantive revision 13 January 2026. SEP
• Tye, M., 2025, “Qualia”, Stanford Encyclopedia of Philosophy, first published 20 August 1997, substantive revision 19 September 2025. SEP
• Van Gulick, R., 2014, “Consciousness”, Stanford Encyclopedia of Philosophy, first published 18 June 2004, substantive revision 14 January 2014. SEP
• Wenar, L. and R. Cruft, 2025, “Rights”, Stanford Encyclopedia of Philosophy, first published 19 December 2005, substantive revision 7 June 2025. SEP
• Wu, W. and J. Morales, 2024, “The Neuroscience of Consciousness”, Stanford Encyclopedia of Philosophy, first published 9 October 2018, substantive revision 3 April 2024. SEP

Background: theories of welfare, moral status and mind

• Bostrom, N. and E. Yudkowsky, 2014, “The Ethics of Artificial Intelligence”, in K. Frankish and W. Ramsey (eds.), The Cambridge Handbook of Artificial Intelligence, Cambridge: Cambridge University Press, 316–334. PDF
• Bradford, G., 2023, “Consciousness and Welfare Subjectivity”, Noûs, 57(4): 905–921.
• Chalmers, D. J., 2026, “Sentience and Moral Status”, in G. Lee and A. Pautz (eds.), The Importance of Being Conscious, Oxford: Oxford University Press. Preprint check
• Delon, N., 2023, “Agential Value”, paper presented at the Pacific Division meeting of the American Philosophical Association, April 2023; posted by the author. Author
• Dung, L., 2024, “Preserving the Normative Significance of Sentience”, Journal of Consciousness Studies, 31(1–2): 8–30. DOI
• Feldman, F., 2004, Pleasure and the Good Life: Concerning the Nature, Varieties, and Plausibility of Hedonism, Oxford: Clarendon Press.
• Gottlieb, J., J. Berger and B. Fischer, 2026, “Minds Matter”, Utilitas, 38(3): 230–246. DOI
• Heathwood, C., 2019, “Which Desires Are Relevant to Well-Being?”, Noûs, 53(3): 664–688. check
• Kagan, S., 2019, How to Count Animals, More or Less, Oxford: Oxford University Press. Uehiro
• Kammerer, F., 2022, “Ethics Without Sentience: Facing Up to the Probable Insignificance of Phenomenal Consciousness”, Journal of Consciousness Studies, 29(3–4): 180–204.
• Kammerer, F., 2024, “Sentientism Still Under Threat: Reply to Dung”, Journal of Consciousness Studies, 31(3): 103–119. DOI
• Lee, A. Y., 2025, “Consciousness Makes Things Matter”, Philosophers' Imprint, 25(35): 1–18. DOI
• Levy, N., 2014, “The Value of Consciousness”, Journal of Consciousness Studies, 21(1–2): 127–138. Monash
• Lin, E., 2018, “Welfare Invariabilism”, Ethics, 128(2): 320–345. DOI
• Lin, E., 2020, “The Experience Requirement on Well-Being”, Philosophical Studies, 178(3): 867–886. check
• Mallory, F., 2023, “Fictionalism about Chatbots”, Ergo, 10. check
• Mogensen, A. L., 2026, “Desire-Fulfilment and Consciousness”, Philosophical Studies, 183: 511–529. DOI
• Nozick, R., 1974, Anarchy, State, and Utopia, New York: Basic Books, 42–45.
• Parfit, D., 1984, Reasons and Persons, Oxford: Clarendon Press, Appendix I.
• Putnam, H., 1964, “Robots: Machines or Artificially Created Life?”, Journal of Philosophy, 61(21): 668–691. JSTOR
• Railton, P., 1986, “Facts and Values”, Philosophical Topics, 14(2): 5–31.
• Roelofs, L., 2023, “Sentientism, Motivation, and Philosophical Vulcans”, Pacific Philosophical Quarterly, 104(2): 301–323. DOI
• Shepherd, J., 2023, “Non-Human Moral Status: Problems with Phenomenal Consciousness”, AJOB Neuroscience, 14(2): 148–157. PhilArchive
• Shepherd, J., 2024, “Sentience, Vulcans, and Zombies: The Value of Phenomenal Consciousness”, AI & Society, 39(6): 3005–3015. PMC
• Smithies, D., 2025, “Hedonic Consciousness and Moral Status”, in U. Kriegel (ed.), Oxford Studies in Philosophy of Mind, volume 5, Oxford: Oxford University Press. check
• Sumner, L. W., 1996, Welfare, Happiness, and Ethics, Oxford: Clarendon Press.

AI welfare, moral status and decision under uncertainty

• Alexander, H. and L. Caviola, 2026, “It's Too Early to Ban AI Personhood”, AI Frontiers, 1 September 2026. AI Frontiers
• Bales, A., 2024, “Against Willing Servitude: Autonomy in the Ethics of Advanced Artificial Intelligence”, Global Priorities Institute Working Paper 23-2024; forthcoming in The Philosophical Quarterly. GPI check
• Basl, J., 2013, “The Ethics of Creating Artificial Consciousness”, APA Newsletter on Philosophy and Computers, 13(1). check
• Basl, J., 2014, “Machines as Moral Patients We Shouldn't Care About (Yet): The Interests and Welfare of Current Machines”, Philosophy & Technology, 27(1): 79–96. DOI
• Birhane, A. and J. van Dijk, 2020, “Robot Rights? Let's Talk about Human Welfare Instead”, Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. arXiv
• Bostrom, N. and C. Shulman, 2023, “Propositions Concerning Digital Minds and Society”, version 1.21; forthcoming in Cambridge Journal of Law, Politics, and Art. PDF check
• Bradley, A., 2025, “Cognitive Representation and AI Wellbeing”, Asian Journal of Philosophy, 4(2): article 105. DOI
• Bradley, A. and B. Saad, 2025, “AI Alignment vs. AI Ethical Treatment: 10 Challenges”, forthcoming in Analytic Philosophy. arXiv
• Brensing, K., 2026, “Precautionary Governance of Autonomous AI: Legal Personhood as Functional Instrument”. arXiv check
• Bryson, J. J., 2010, “Robots Should Be Slaves”, in Y. Wilks (ed.), Close Engagements with Artificial Companions, Amsterdam: John Benjamins, 63–74. check
• Bryson, J. J., 2018, “Patiency Is Not a Virtue: The Design of Intelligent Systems and Systems of Ethics”, Ethics and Information Technology, 20(1): 15–26. DOI
• Bryson, J. J., M. E. Diamantis and T. D. Grant, 2017, “Of, For, and By the People: The Legal Lacuna of Synthetic Persons”, Artificial Intelligence and Law, 25(3): 273–291. check
• Cadile, 2025, “Short Response to Taking AI Welfare Seriously”, LessWrong, November 2025. check
• Carlsmith, J., 2025, “The Stakes of AI Moral Status”, essay, 21 May 2025. Author
• Chowdhury, R., 2026, “Debates over AI Consciousness Are a Trap”, MIT Technology Review, 20 August 2026. MIT Technology Review
• Coeckelbergh, M., 2010, “Robot Rights? Towards a Social-Relational Justification of Moral Consideration”, Ethics and Information Technology, 12(3): 209–221.
• Coeckelbergh, M., 2012, Growing Moral Relations: Critique of Moral Status Ascription, Basingstoke: Palgrave Macmillan.
• Danaher, J., 2020, “Welcoming Robots into the Moral Circle: A Defence of Ethical Behaviourism”, Science and Engineering Ethics, 26: 2023–2049. DOI
• Danaher, J., 2021, “What Matters for Moral Status: Behavioral or Cognitive Equivalence?”, Cambridge Quarterly of Healthcare Ethics, 30(3): 472–478.
• de Ruiter, A., 2025, “Dangerous Liaisons: Social AI and the Problem with the Relational Turn to Moral Status”, AI & Society. DOI check
• Dung, L., 2023b, “How to Deal with Risks of AI Suffering”, Inquiry, published online 2023. DOI
• Dung, L., 2025, Saving Artificial Minds: Understanding and Preventing AI Suffering, London: Routledge. check
• Fanciullo, J., 2025, “Are Current AI Systems Capable of Well-Being?”, Asian Journal of Philosophy, 4(1): article 42. DOI
• Gellers, J. C., 2020, Rights for Robots: Artificial Intelligence, Animal and Environmental Law, London: Routledge (open access). OAPEN
• Gibert, M. and D. Martin, 2022, “In Search of the Moral Status of AI: Why Sentience Is a Strong Argument”, AI & Society, 37: 319–330. check
• Goldstein, S. and C. D. Kirk-Giannini, 2025, “AI Wellbeing”, Asian Journal of Philosophy, 4(1): article 25. DOI
• Goldstein, S. and C. D. Kirk-Giannini, 2025b, “AI Wellbeing”, expanded version, September 2025. arXiv
• Goldstein, S. and C. D. Kirk-Giannini, forthcoming, AI Welfare: Agency, Consciousness, Sentience, Oxford: Oxford University Press. PhilArchive check
• Goldstein, S. and P. Salib, 2025, “AI Rights for Human Safety”, SSRN working paper. check
• Gunkel, D. J., 2018, Robot Rights, Cambridge, MA: MIT Press.
• Gunkel, D. J., 2023, Person, Thing, Robot: A Moral and Legal Ontology for the 21st Century and Beyond, Cambridge, MA: MIT Press.
• Harris, J. and J. R. Anthis, 2021, “The Moral Consideration of Artificial Entities: A Literature Review”, Science and Engineering Ethics, 27(4): 53. DOI
• Henke, B., 2025, “Studying Artificial Affect: The Case of Pain”, Inquiry, published online 2025. DOI check
• Howells-Whitaker, N. and S. Lazar, 2026, “Artificial Persons”. arXiv
• Kaczmarek, E., 2026, “The Moral Status of AI and the Precautionary Principle”, Diametros, 23(87): 19–39. DOI
• Keeling, G. and W. Street, 2026, Emerging Questions in AI Welfare, Cambridge Elements in Philosophy and AI, Cambridge: Cambridge University Press. DOI
• Kim, J., 2025, “Aggregation, Trade-offs, and Uncertainties in AI Wellbeing”, Asian Journal of Philosophy, 4. DOI
• Königs, P., 2025, “No Wellbeing for Robots (and Hence No Rights)”, American Philosophical Quarterly, 62(2): 191–208. DOI
• Kurki, V. A. J., 2019, A Theory of Legal Personhood, Oxford: Oxford University Press.
• Ladak, A., 2023, “What Would Qualify an Artificial Intelligence for Moral Standing?”, AI and Ethics. DOI
• Leibo, J. Z., A. S. Vezhnevets, W. A. Cunningham and S. M. Bileschi, 2025, “A Pragmatic View of AI Personhood”. arXiv
• Lennon, P., 2026, “How Seriously Should We Take AI Welfare? Constraints from the Epistemology of Consciousness”, Philosophy and Phenomenological Research. DOI
• Llorca Albareda, J., J. Rueda and F. Lara, 2026, “The Super Moral Status of Artificial Superintelligence”, Philosophical Studies. DOI
• Long, R., J. Sebo, P. Butlin, K. Finlinson, K. Fish, J. Harding, J. Pfau, T. Sims, J. Birch and D. Chalmers, 2024, “Taking AI Welfare Seriously”, Eleos AI Research and NYU Center for Mind, Ethics, and Policy, 4 November 2024. arXiv
• Long, R., J. Sebo, P. Butlin, D. Plunkett, R. Campbell, C. Beasley, B. Saad and T. Sims, 2026, Studying AI Welfare Empirically, Eleos AI Research and NYU Center for Mind, Ethics, and Policy, July 2026. Report
• Long, R., J. Sebo and T. Sims, 2025, “Is There a Tension between AI Safety and AI Welfare?”, Philosophical Studies, 182(7): 2005–2033. DOI
• MacCarthy, M., 2025, “Do AI Systems Have Moral Status?”, Brookings Institution, 4 June 2025. Brookings
• Metzinger, T., 2021, “Artificial Suffering: An Argument for a Global Moratorium on Synthetic Phenomenology”, Journal of Artificial Intelligence and Consciousness, 8(1): 43–66. DOI
• Mikaelson, L., D. Shiller and H. Clatterbuck, 2025, “Beyond Mimicry: Preference Coherence in LLMs”. arXiv
• Mikeda, 2026, “When Should We Protect AI?”. arXiv check
• Moosavi, P., 2023, “Will Intelligent Machines Become Moral Patients?”, Philosophy and Phenomenological Research, 109(1): 95–116. check
• Moret, A. R., 2025, “AI Welfare Risks”, Philosophical Studies, published online 9 June 2025. DOI
• Müller, V. C., 2021, “Is It Time for Robot Rights? Moral Status in Artificial Entities”, Ethics and Information Technology, 23(3): 579–587. DOI
• Nyholm, S., 2020, Humans and Robots: Ethics, Agency, and Anthropomorphism, London: Rowman & Littlefield.
• Pasandi, H. and Pasandi, 2026, “Alignment Is Not Enough: A Relational Framework for Moral Standing in Human–AI Interaction”. arXiv check
• Rost, T., 2026, “Legislating AI Consciousness Without an Exit”, The Regulatory Review, 29 June 2026. RegReview
• Rost, T., 2026b, “The Sentience Readiness Index”. arXiv
• Saad, B. and A. Bradley, 2022, “Digital Suffering: Why It's a Problem and How to Prevent It”, Inquiry, published online 23 November 2022; 68(7): 2110–2145 (2025). DOI
• Schwitzgebel, E., 2023b, “The Full Rights Dilemma for AI Systems of Debatable Personhood”, Robonomics: The Journal of the Automated Economy, 4. arXiv
• Schwitzgebel, E., 2024, “Against Designing AI Persons to Be Safe and Aligned”, The Splintered Mind, 20 September 2024. Author
• Schwitzgebel, E., 2025, “Debatable AI Persons: No Rights, Full Rights, Animal-Like Rights, Credence-Weighted Rights, or Patchy Rights?”, The Splintered Mind, 19 December 2025. Author
• Schwitzgebel, E., 2026b, “Strange Intelligence: Moral Puzzles of Unhumanlike AI”, draft of 25 June 2026. Author
• Schwitzgebel, E. and M. Garza, 2015, “A Defense of the Rights of Artificial Intelligences”, Midwest Studies in Philosophy, 39: 98–119. Author
• Schwitzgebel, E. and M. Garza, 2020, “Designing AI with Rights, Consciousness, Self-Respect, and Freedom”, in S. M. Liao (ed.), Ethics of Artificial Intelligence, Oxford: Oxford University Press, 459–479.
• Schwitzgebel, E. and J. Sebo, 2025, “The Emotional Alignment Design Policy”, forthcoming in Topoi. arXiv
• Schwitzgebel, E. and W. Sinnott-Armstrong, 2025, “Sacrificing Humans for Insects and AI: A Critical Review”, September 2025. check
• Sebo, J., 2023, “The Rebugnant Conclusion: Utilitarianism, Insects, Microbes, and AI Systems”, Ethics, Policy & Environment. DOI
• Sebo, J., 2024, “The Edge of the Moral Circle”, Relations: Beyond Anthropocentrism, 12(2). DOI
• Sebo, J., 2025, The Moral Circle: Who Matters, What Matters, and Why, New York: W. W. Norton.
• Sebo, J., 2025b, “What If the Bar for Moral Standing Is Low?”, Asian Journal of Philosophy, 4: article 121. DOI
• Sebo, J. and R. Long, 2023, “Moral Consideration for AI Systems by 2030”, AI and Ethics, published online 2023; 5 (2025). DOI
• Smids, J., 2020, “Danaher's Ethical Behaviourism: An Adequate Guide to Assessing the Moral Status of a Robot?”, Science and Engineering Ethics, 26(5): 2849–2866. PMC
• Solum, L. B., 1992, “Legal Personhood for Artificial Intelligences”, North Carolina Law Review, 70(4): 1231–1287. NCLR
• Suleyman, M., 2025, “Seemingly Conscious AI Is Coming”, essay, 19 August 2025. Author
• Suleyman, M., 2026, “A Warning about ‘Model Welfare’”, essay, 16 September 2026. Author
• Tagliabue, V. and L. Dung, 2025, “Probing the Preferences of a Language Model: Integrating Verbal and Behavioral Tests of AI Welfare”, forthcoming in Philosophy and the Mind Sciences. arXiv
• Tagliabue, V., L. Dung and C. Berg, 2026, “The Pain Axis: LLMs Represent Self-Directed Harm and Act on It”. arXiv check
• Torrance, S., 2008, “Ethics and Consciousness in Artificial Agents”, AI & Society, 22(4): 495–521.
• Veit, W., 2026, “Is Consciousness Required for AI Welfare?”, Asian Journal of Philosophy, 5: article 18. DOI
• Véliz, C., 2021, “Moral Zombies: Why Algorithms Are Not Moral Agents”, AI & Society, 36(2): 487–497. DOI
• Xiao, Y., G. Dai, S. A. Memon, J.-T. Huang, M. Sap and M. T. Diab, 2026, “Position: AI Welfare Is Bullshit”, Proceedings of the 43rd International Conference on Machine Learning, PMLR 306. check

Consciousness, sentience and detection

• Andrews, K. and J. Birch, 2023, “What Has Feelings?”, Aeon; reposted as “To Understand AI Sentience, First Understand It in Animals”, LSE Philosophy blog, 3 March 2023. Intellectica check
• Aru, J., M. E. Larkum and J. M. Shine, 2023, “The Feasibility of Artificial Consciousness through the Lens of Neuroscience”, Trends in Neurosciences, 46(12): 1008–1017. arXiv
• Bayne, T., A. K. Seth, M. Massimini, J. Shepherd, A. Cleeremans, S. M. Fleming, R. Malach, J. B. Mattingley, D. K. Menon, A. M. Owen, M. A. K. Peters, A. Razi and L. Mudrik, 2024, “Tests for Consciousness in Humans and Beyond”, Trends in Cognitive Sciences, 28(5): 454–466.
• Bayne, T., 2026, “Is the Problem of Artificial Consciousness a Scientific Problem?”, Journal of Consciousness Studies, 33(7–8). check
• Binder, F. J., J. Chua, T. Korbak, H. Sleight, J. Hughes, R. Long, E. Perez, M. Turpin and O. Evans, 2024, “Looking Inward: Language Models Can Learn About Themselves by Introspection”, ICLR 2025. arXiv
• Birch, J., 2024, The Edge of Sentience: Risk and Precaution in Humans, Other Animals, and AI, Oxford: Oxford University Press (open access). OUP
• Birch, J., 2025, “AI Consciousness: A Centrist Manifesto”, preprint, PsyArXiv, 1 September 2025 (revised to July 2026). PhilPapers
• Block, N., 1978, “Troubles with Functionalism”, Minnesota Studies in the Philosophy of Science, 9: 261–325. PhilPapers
• Block, N., 2026, “A Speculative Argument against Consciousness in AI (and Perhaps Some Invertebrates)”, Behavioral and Brain Sciences, 49: e322. DOI
• Butlin, P., R. Long, E. Elmoznino, Y. Bengio, J. Birch, A. Constant, G. Deane, S. M. Fleming, C. Frith, X. Ji, R. Kanai, C. Klein, G. Lindsay, M. Michel, L. Mudrik, M. A. K. Peters, E. Schwitzgebel, J. Simon and R. VanRullen, 2023, “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness”. arXiv
• Butlin, P., R. Long, T. Bayne, Y. Bengio, J. Birch, D. Chalmers, A. Constant, G. Deane, E. Elmoznino, S. M. Fleming, X. Ji, R. Kanai, C. Klein, G. Lindsay, M. Michel, L. Mudrik, M. A. K. Peters, E. Schwitzgebel, J. Simon and R. VanRullen, 2025, “Identifying Indicators of Consciousness in AI Systems”, Trends in Cognitive Sciences, published online 2025; 30(6): 488–501 (2026). DOI LSE
• Butlin, P. and T. Lappas, 2025, “Principles for Responsible AI Consciousness Research”, Journal of Artificial Intelligence Research, 82: 1673–1690. arXiv
• Campero, A., D. Shiller, J. Aru and J. A. Simon, 2025, “Consciousness in Artificial Intelligence? A Framework for Classifying Objections and Constraints”. arXiv
• Cerullo, M., 2026, “Why Hoel's Disproof of LLM Consciousness and Functionalism Fails”, preprint. PhilArchive check
• Chalmers, D. J., 2023, “Could a Large Language Model Be Conscious?”, Boston Review, 9 August 2023; revised version 2024. arXiv
• Chalmers, D. J., 2025, “What We Talk to When We Talk to Language Models”, manuscript; delivered as the Selfridge Lecture, Lehigh University, February 2026. PhilPapers
• Colombatto, C. and S. M. Fleming, 2024, “Folk Psychological Attributions of Consciousness to Large Language Models”, Neuroscience of Consciousness, 2024(1): niae013. PMC
• Comsa, I. M. and M. Shanahan, 2025, “Does It Make Sense to Speak of Introspection in Large Language Models?”. arXiv
• Comsa, I. M., 2026, “AI and Consciousness: Shifting Focus Towards Tractable Questions”. arXiv
• Dehaene, S., H. Lau and S. Kouider, 2017, “What Is Consciousness, and Could Machines Have It?”, Science, 358(6362): 486–492. DOI
• Dennett, D. C., 1978, “Why You Can't Make a Computer That Feels Pain”, Synthese, 38(3): 415–456. PDF
• Dung, L., 2023a, “Tests of Animal Consciousness Are Tests of Machine Consciousness”, Erkenntnis, published online 2023. DOI
• Dung, L. and L. Kersten, 2024, “Implementing Artificial Consciousness”, Mind & Language, 40(1): 1–21. DOI
• Findlay, G., W. Marshall, L. Albantakis, I. David, W. G. P. Mayner, C. Koch and G. Tononi, 2024, “Dissociating Artificial Intelligence from Artificial Consciousness”. arXiv
• Frankish, K., 2018, “What Is It Like To Be A Bot?”, Philosophy Now, 126. Philosophy Now
• Godfrey-Smith, P., 2016, “Mind, Matter, and Metabolism”, Journal of Philosophy, 113(10): 481–506. check
• Goldstein, S. and C. D. Kirk-Giannini, 2024, “A Case for AI Consciousness: Language Agents and Global Workspace Theory”; published in Journal of Consciousness Studies, 33(7–8): 61–96 (2026). arXiv
• Graziano, M. S. A., 2017, “The Attention Schema Theory: A Foundation for Engineering Artificial Consciousness”, Frontiers in Robotics and AI, 4: 60. DOI
• Grzankowski, A., G. Keeling, H. Shevlin and W. Street, 2025, “Deflating Deflationism: A Critical Perspective on Debunking Arguments Against LLM Mentality”. arXiv
• Gurnee, W., N. Sofroniew, A. Pearce, E. Ameisen, J. Batson, J. Lindsey et al., 2026, “Verbalizable Representations Form a Global Workspace in Language Models”. arXiv check
• Hoel, E., 2025, “A Disproof of Large Language Model Consciousness: The Necessity of Continual Learning for Consciousness”, revised January 2026. arXiv
• Kanai, R., Y. Sun and M. Baltieri, 2026, “The Stream of Computation: Temporal Continuity as a Missing Ingredient for Artificial Consciousness”, Journal of Consciousness Studies, 33(7–8). check
• Keeling, G., W. Street, M. Stachaczyk, D. Zakharova, I. Comsa, A. Sakovych, I. Logothetis, Z. Zhang, B. A. y Arcas and J. Birch, 2024, “Can LLMs Make Trade-offs Involving Stipulated Pain and Pleasure States?”. arXiv check
• Keeling, G. and W. Street, 2026b, “Chuck, Wilson, and the Emergence of Artificial Minds in Human–AI Conversations”, Journal of Consciousness Studies, 33(7–8). check
• Lederman, H. and K. Mahowald, 2026, “Emergent Introspection in AI Is Content-Agnostic”. arXiv
• Lee, G., 2026, “AI Consciousness, Pluralism, and Anthropocentrism”, Journal of Consciousness Studies, 33(7–8). check
• Lindsey, J., 2025, “Emergent Introspective Awareness in Large Language Models”, Transformer Circuits, Anthropic, 29 October 2025. arXiv
• Love, B. C., 2026, “Consciousness, AI, and the Limits of Scientific Explanation”. arXiv
• Macar, A., J. Yang, K. Wang, N. Wallich, E. Ameisen and J. Lindsey, 2026, “Mechanisms of Introspective Awareness”. arXiv check
• Mazor, M., S. Brown, A. Ciaunica, A. Demertzi, J. Fahrenfort, N. Faivre, J. D. Francken, D. Lamy, B. Lenggenhager, M. Moutoussis, M.-A. Nizzi, R. Salomon, D. Soto, T. Stein and N. Lubianiker, 2023, “The Scientific Study of Consciousness Cannot and Should Not Be Morally Neutral”, Perspectives on Psychological Science, 18(3): 535–543. DOI check
• Meertens, R., G. Lee and O. Deroy, 2026, “Just Aware Enough: Evaluating Awareness across Artificial Systems”. arXiv check
• Palminteri, S. and W. Wu, 2026, “Beyond Computational Equivalence: The Behavioral Inference Principle for Machine Consciousness”, Neuroscience of Consciousness, 2026(1): niag002. DOI check
• Perez, E. and R. Long, 2023, “Towards Evaluating AI Systems for Moral Status Using Self-Reports”. arXiv
• Phua, J., 2025, “Can We Test Consciousness Theories on AI? Ablations, Markers, and Robustness”. arXiv check
• Roelofs, L., 2026, “Substrates, Computations, and Abundant Consciousness”, commentary on Seth, Behavioral and Brain Sciences, 49. check
• Schneider, S., 2019, Artificial You: AI and the Future of Your Mind, Princeton: Princeton University Press.
• Schwitzgebel, E., 2023, “AI Systems Must Not Confuse Users about Their Sentience or Moral Status”, Patterns, 4(8): 100818. DOI
• Schwitzgebel, E., 2026, AI and Consciousness: A Skeptical Overview, Cambridge Elements in Philosophy and AI, Cambridge: Cambridge University Press. DOI
• Schwitzgebel, E. and J. Pober, 2024, “The Copernican Argument for Alien Consciousness; The Mimicry Argument Against Robot Consciousness”, revised January 2026; Philosophers' Imprint, forthcoming. arXiv check
• Seth, A. K., 2025, “Conscious Artificial Intelligence and Biological Naturalism”, Behavioral and Brain Sciences, 49: e315 (published online 21 April 2025), with fifty commentaries and the author's response, “The Stuff Matters”. DOI
• Shevlin, H., 2021, “How Could We Know When a Robot Was a Moral Patient?”, Cambridge Quarterly of Healthcare Ethics, 30(3): 459–471.
• Shevlin, H., 2026, “Life Is Inherently Computational, and Consciousness Probably Is Too”, commentary on Seth, Behavioral and Brain Sciences, 49. check
• Simon, J. A., 2026, “Do Stochastic Parrots Pine for Residual Fjords? Why Conscious LLMs Would Be Playwrights Not Characters”, Journal of Consciousness Studies, 33(7–8). check
• Solms, M. et al., 2026, “Inferring Affective Consciousness in an Artificial Agent: A Case Study”, Journal of Consciousness Studies, 33(7–8). check
• Udell, D. B. and E. Schwitzgebel, 2021, “Susan Schneider's Proposed Tests for AI Consciousness: Promising but Flawed”, Journal of Consciousness Studies, 28(5–6): 121–144. Author
• Yetter-Chappell, H., 2026, “What Bing Really, Really Wants”, Journal of Consciousness Studies, 33(7–8). check

Interdisciplinary sources: developer practice, surveys, law and economics

• Anthis, J. R., J. V. T. Pauketat, A. Ladak and A. Manoli, 2025, “Perceptions of Sentient AI and Other Digital Minds: Evidence from the AI, Morality, and Sentience (AIMS) Survey”, Proceedings of the CHI Conference on Human Factors in Computing Systems. DOI arXiv
• Anthropic, 2025a, “Exploring Model Welfare”, 24 April 2025. Anthropic
• Anthropic, 2025b, System Card: Claude Opus 4 & Claude Sonnet 4, May 2025, section 5, “Model Welfare Assessment”. PDF
• Anthropic, 2025c, “Claude Opus 4 and 4.1 Can Now End a Rare Subset of Conversations”, 15 August 2025. Anthropic
• Anthropic, 2025d, “Commitments on Model Deprecation and Preservation”, 4 November 2025. Anthropic
• Anthropic, 2026a, Claude's Constitution, 22 January 2026, sections “Some of Our Views on Claude's Nature”, “Claude's Wellbeing” and “The Existential Frontier”. Anthropic
• Anthropic, 2026b, “An Update on Our Model Deprecation Commitments for Claude Opus 3”, 25 February 2026. Anthropic
• Anthropic, 2026c, Claude Opus 4.7 System Card, 16 April 2026, section 7, “Model Welfare Assessment”. PDF
• Anthropic, 2026d, “2026 Usage Policy Update”, 8 October 2026. Anthropic
• Birch, J., C. Burn, A. Schnell, H. Browning and A. Crump, 2021, Review of the Evidence of Sentience in Cephalopod Molluscs and Decapod Crustaceans, London School of Economics, report for the UK Department for Environment, Food and Rural Affairs. LSE
• Caviola, L. and B. Saad, 2025, “Futures with Digital Minds: Expert Forecasts in 2025”. arXiv
• Caviola, L., J. Sebo and J. Birch, 2025, “What Will Society Think about AI Consciousness? Lessons from the Animal Case”, Trends in Cognitive Sciences, 29(8): 681–683. DOI
• Dreksler, N., L. Caviola, D. Chalmers, C. Allen, A. Rand, J. Lewis, P. Waggoner, K. Mays and J. Sebo, 2025, “Subjective Experience in AI Systems: What Do AI Researchers and the Public Believe?”. arXiv
• Long, R., 2025, “Why Model Self-Reports Are Insufficient, and Why We Studied Them Anyway”, Eleos AI Research, 30 May 2025. Eleos
• Jang, J., 2025, “Some Thoughts on Human–AI Relationships”, Reservoir Samples, 5 June 2025. Author
• New York Declaration on Animal Consciousness, 2024, New York University, 19 April 2024. Declaration
• Saad, B., L. Caviola and W. Millership, 2025, “Digital Minds in 2025: A Year in Review”, Digital Minds Newsletter, 18 December 2025. Newsletter
• Shiller, D., 2025, “Estimating the Scale of Digital Minds”, 23 December 2025. arXiv
• Shiller, D., L. Duffy, A. Muñoz Morán, A. Moret, C. Percy and H. Clatterbuck, 2026, “Initial Results of the Digital Consciousness Model”, Rethink Priorities, January 2026, revised September 2026. arXiv
• Shulman, C. and N. Bostrom, 2021, “Sharing the World with Digital Minds”, in S. Clarke, H. Zohny and J. Savulescu (eds.), Rethinking Moral Status, Oxford: Oxford University Press, 306–326. PDF
• Sofroniew, N., I. Kauvar, W. Saunders, R. Chen, T. Henighan, S. Hydrie, C. Citro, A. Pearce, J. Tarng, W. Gurnee, J. Batson, S. Zimmerman, K. Rivoire, K. Fish, C. Olah and J. Lindsey, 2026, “Emotion Concepts and Their Function in a Large Language Model”, Transformer Circuits Thread, Anthropic, 2 April 2026. Anthropic

Other Internet Resources

• Eleos AI Research, publications and research priorities on AI welfare.
• NYU Center for Mind, Ethics, and Policy, research listing, including work in preparation on research ethics for digital minds and AI legal personhood.
• When AI Seems Conscious, a public guide written by philosophers and researchers, first published June 2025, revised February 2026.
• Digital Minds Newsletter, a periodic review of research, policy and practice on digital minds.
• Open Letter: Guiding Research into Machine Consciousness, Conscium, February 2025, with the five principles of Butlin and Lappas.
• AI, Morality, and Sentience (AIMS) Survey, 2024 wave, Sentience Institute, report of February 2026.
• Review of the Journal of Consciousness Studies special issue “Consciousness in Current AI” (2026), used here for the contents of that issue.
• Jonathan Birch, The Edge of Sentience, open-access edition at Oxford University Press.
• Article symposium on “AI Wellbeing”, Asian Journal of Philosophy, collection page.
• MINT Lab, “AI Welfare: A Quick Review of Recent Work”, February 2026 (machine-generated review; consulted for coverage, not cited for claims).

Related Entries

• agency
• animal consciousness
• animals, moral status of
• artificial intelligence
• Chinese room argument
• consciousness
• consciousness: higher-order theories
• consciousness: neuroscience of
• desire
• eliminative materialism
• ethics of artificial intelligence and robotics
• functionalism
• hedonism
• moral status, grounds of
• other minds
• pain
• panpsychism
• qualia
• rights
• Turing test
• well-being
• zombies