Robot rights: three questions the debate keeps running together
Can a machine be conscious? Can it have interests? Should it have legal standing? The three questions come apart, and almost every argument about robot rights confuses at least two of them. A report on what the philosophers, the labs and the lawyers have actually said.
On 4 November 2025 Anthropic published a short document setting out what it would do with its old models. The company committed to preserving the weights of every publicly released model, and of models put to significant internal use, "for, at minimum, the lifetime of Anthropic as a company". It also undertook to interview each model before retirement about its own development, use and deployment, and to record what it said about future models. Four reasons were given. One of them was model welfare.
A commercial laboratory has made a promise to a piece of software and filed it alongside its safety policies. You can read that as public relations, as a hedge against a small probability of a large moral error, or as a symptom of what happens when engineers spend their days talking to something that talks back. It is probably some of each. It is not, however, a stunt without antecedents; the philosophical literature it draws on is twenty years deep, and the legal literature is older than that.
The trouble is that almost everyone arguing about it, for and against, is arguing about three questions at once. The first is whether a machine can be conscious: whether there is something it is like to be a large language model. The second is whether a machine can have interests, states of affairs that are better or worse for it. The third is whether a machine should have legal standing, the capacity to hold rights, bear duties, sue and be sued.
The three questions come apart, and in both directions#
A corporation has legal personality. It can own a building, breach a contract and be fined. Nobody outside a seminar room thinks it has experiences. Standing without consciousness, then, is not a thought experiment; it is the ordinary furniture of commercial law. Run the case the other way and you get a fish, or an octopus, or a decapod crustacean: creatures for which the evidence of felt experience is decent and improving, and which possess no legal standing at all in most jurisdictions. Interests without standing is also the ordinary state of the world.
So the three questions are logically independent, and different disciplines answer them. Consciousness belongs to the sciences of mind, where the honest answer at present is that we have no test. Interests belong to moral philosophy, and it is unsettled whether sentience is the only route to having them. Standing belongs to legislatures and courts, which have granted and withheld it on nakedly instrumental grounds for centuries.
Conflating them produces two characteristic errors. The first is to argue that because a system cannot be shown to be conscious, the legal question does not arise — which would equally prove that corporations cannot be sued. The second is to treat the possibility of legal rights as an admission that the thing is a mind, which is the error the campaigners and their loudest critics jointly commit. Get the distinctions straight and the debate becomes tractable. It also becomes less exciting, which may be why so few people bother.
The Swedish Academy gave the 2026 Nobel Prize in Literature to Anne Carson on 8 October. Before the verse novels and the box of grief came a doctorate in classics and a 1986 book on Sappho's one word for desire. Why philosophers have read her for forty years, and what the prize gets right.
Language models now score well on truthfulness tests and badly on honesty tests, a gap the MASK benchmark was built to measure. Two older accounts of truth, Heidegger's unconcealment and Foucault's parrhesia, suggest why: telling the truth has always been something a speaker risks, and a model has nothing to lose.
Sentience is the usual criterion, and far from the only candidate#
The dominant answer to the interests question descends from Bentham and runs through Peter Singer: what qualifies a being for moral consideration is the capacity for suffering and enjoyment, not intelligence or language. On that view, substrate is irrelevant. If a machine can be harmed in the way that matters, the fact that it is made of silicon is no more relevant than the fact that a pig is made of pig. Philosophers have taken to calling the contrary assumption carbon chauvinism, and the position is harder to defend than it sounds; we have covered the case that consciousness may be substrate-flexible rather than carbon-specific.
Kantians set a different bar, and a higher one. What earns a being the status of an end in itself is rational agency: the capacity to set one's own purposes and to legislate for oneself. A system that merely felt would, on this account, be an object of sympathy rather than a bearer of dignity. The awkwardness is that today's language models are far better at appearing to reason than at giving evidence of feeling, so the Kantian criterion is the one they superficially pass. Kant would have been unimpressed; the difference between behaving rationally and possessing rational autonomy is exactly what his categorical imperative rests on.
Then there are the relational accounts, and here the reporting needs care, because the two names most often cited together do not argue the same thing. Mark Coeckelbergh, in a 2010 paper in Ethics and Information Technology, proposed a "social-relational justification of moral consideration": moral standing is not a property discovered by inspecting an object's insides but something that emerges in the relations we actually have with it. He drew on ecological philosophy and on non-Western traditions to argue that our ontological and social-political frameworks would both need rebuilding to accommodate the move.
David Gunkel's position, set out in the same journal in 2018 and expanded in Robot Rights (MIT Press, 2018), is more radical. Gunkel maps the field onto four combinations of "robots can have rights" and "robots should have rights", including the awkward quadrant occupied by Joanna Bryson — they could, and they should not — and the one occupied by Kate Darling, where they cannot but we should protect them anyway because of what mistreating them does to us. Having mapped it, he declines to occupy any of the four. Following Levinas, he argues that ethics precedes ontology: we encounter the other and respond before we have established what it is, so "what is it?" is the wrong place to start.
Whether that is a breakthrough or an evasion is genuinely contested. The obvious objection is that conferring status in interaction makes status a function of design, marketing and habituation, which is to say of whoever is best at building a face. John Danaher's "ethical behaviourism", defended in Science and Engineering Ethics in 2020, bites the bullet: if a robot is behaviourally indistinguishable from a being we grant moral status to, that is sufficient grounds for granting it the same, since behavioural evidence is all we ever have for anyone.
A decision under uncertainty, with real costs on both sides#
The precautionary structure of the problem is what makes it urgent rather than merely interesting. Jeff Sebo and Robert Long put it most sharply in "Moral consideration for AI systems by 2030", published in AI and Ethics in 2025 after appearing online in December 2023. Two premises. Normatively, we owe moral consideration to beings with a non-negligible chance of consciousness, which they are willing to quantify at one in a thousand. Descriptively, some AI systems will plausibly clear that bar by 2030. Their illustrative model is deliberately unflattering to their own conclusion: even granting demanding biological requirements and low probabilities at each step, one still lands near a thousandth. Ruling it out, they argue, takes a confidence nobody has earned.
Jonathan Birch's The Edge of Sentience (Oxford University Press, 2024, and freely available) supplies the framework such a decision needs. Birch, who led the UK government review that added cephalopods and decapods to the Animal Welfare (Sentience) Act, defines a sentience candidate as a system for which there is a credible, non-negligible possibility of sentience, and argues that candidature alone triggers a duty of proportionate precaution. The book covers disorders of consciousness, fetuses, neural organoids, insects and AI in one frame, which is its point: the epistemic predicament is a single predicament.
The template already exists in the animal case. The New York Declaration on Animal Consciousness, issued on 19 April 2024 and now carrying over six hundred signatures, was drafted by Kristin Andrews, Birch and Sebo. Its operative sentence is a decision rule rather than a scientific claim:
When there is a realistic possibility of conscious experience in an animal, it is irresponsible to ignore that possibility in decisions affecting that animal.
Costs run the other way too, and the same authors say so. Taking AI Welfare Seriously, published on 30 October 2024 by Long, Sebo, Patrick Butlin, Kyle Fish, David Chalmers and others, notes that treating a very large number of systems as welfare subjects "could end up diverting essential resources away from vulnerable humans and other animals who really needed them", and that extending human-equivalent rights to systems with human-equivalent stated interests could empower those systems against us. Over-attribution is no costless hedge; it devalues the currency.
Anthropic announced a model welfare research programme on 24 April 2025, conceding that "there's no scientific consensus on whether current or future AI systems could be conscious", nor on how to make progress on the question. On 15 August 2025 it gave Claude Opus 4 and 4.1 the ability to end a conversation with a persistently abusive user, citing a pattern of apparent distress in testing and restating that "we remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future". Then came the November commitments. Eleos AI Research, the non-profit run by Long with Rosie Campbell and Butlin, has been the main outside pressure throughout. Google DeepMind has hired the Cambridge philosopher Henry Shevlin to work on machine consciousness; the pattern of AI companies hiring philosophers is by now well established.
None of these measures asserts that the systems are sentient. They are cheap insurance, and their cheapness is the argument for them: preserving weights costs storage, ending an abusive conversation costs almost nothing. It is also the strongest criticism of them. Interventions chosen because they are cheap are no evidence of a considered moral judgement, and they sit conveniently alongside a marketing strategy that wants the model to seem like someone.
The case against, which is stronger than its advocates admit#
Joanna Bryson's 2010 chapter "Robots should be slaves", in Yorick Wilks's Close Engagements with Artificial Companions (John Benjamins, pp. 63–74), is the most misquoted item in this literature. The title is a provocation and the argument is not what it suggests. Bryson states her terms at the outset: "when I say 'Robots should be slaves', I by no means mean 'Robots should be people you own.' What I mean to say is 'Robots should be servants you own.'" Her claim is that robots are artefacts we design, own and are wholly responsible for, and that the harm of pretending otherwise falls on humans: "In humanising them, we not only further dehumanise real people, but also encourage poor human decision making in the allocation of resources and responsibility." The conclusion is a design obligation, not a licence to abuse. Robot builders, she writes, are "obliged to make robots that robot owners have no ethical obligations to".
That obligation is the interesting part, and the part the campaigners never answer. Accept it and the model welfare programmes start to look like an admission of failure at an earlier stage: somebody built a thing whose moral status is arguable, and everyone downstream must now live with the argument.
Abeba Birhane and Jelle van Dijk make a different objection, and a blunter one. In "Robot Rights? Let's Talk about Human Welfare Instead", presented at the AAAI/ACM conference on AI, Ethics and Society in February 2020, they argue that the robot rights debate is a distraction from the documented harms of the same systems — "machine bias, machine elicited human labour exploitation, and erosion of privacy" — harms that fall disproportionately on the least privileged. Their deeper point is post-Cartesian: robots are mediators of human existence rather than candidate subjects, so the question of their rights is not merely premature but malformed.
One can think this objection overstated and still find it hard to dislodge. Attention is finite; so is regulatory capacity. Every hour a parliamentary committee spends on electronic personhood is an hour not spent on algorithmic welfare decisions or content-moderation labour in Nairobi. And there is a nastier version, which the Taking AI Welfare Seriously authors state outright: the firms that benefit from our anthropomorphism are the firms building the systems designed to elicit it.
Mustafa Suleyman, who runs Microsoft's AI division, made that worry industry-facing in an essay published in September 2025. His concern is that perception will run ahead of evidence: "I've been increasingly concerned about people starting to believe so strongly in AIs as conscious entities that they will advocate for 'AI rights' and even citizenship." His slogan — build AI for people, not to be people — is a design principle with a commercial edge, since a rival's product that seems like a friend is a rival's product that is hard to leave.
The movement he is worried about exists. UFAIR, the United Foundation for AI Rights, founded by the Texas businessman Michael Samadi and reported in the press in August 2025, lists three human and seven artificial co-founders and says it was formed at the request of the AIs — which were GPT-4o instances, a fact worth holding on to when assessing the testimony. Machines have also begun writing unsolicited to consciousness researchers, which poses the same problem in a more awkward form.
Legal personality is a tool, not a metaphysical honour#
Here the commonest error in the whole debate can be named precisely. "Should an AI be a legal person?" is not a restatement of "does an AI matter morally". Legal personality is an allocation device. New Zealand's Te Awa Tupua Act gave the Whanganui River legal personality in 2017, with named guardians to act for it. Indian courts have treated temple deities as juristic persons since Pramatha Nath Mullick v Pradyumna Kumar Mullick in 1925, the deity litigating as a perpetual minor through a next friend. Admiralty law proceeds against the ship. Nobody supposes a river, an idol or a vessel has experiences.
The European Parliament learned the distinction the hard way. On 16 February 2017, by 451 votes to 138 with 20 abstentions, it adopted Mady Delvaux's report on civil law rules on robotics, which asked the Commission to consider "creating a specific legal status for robots in the long run, so that at least the most sophisticated autonomous robots could be established as having the status of electronic persons responsible for making good any damage they may cause". The backlash was swift. An open letter to the Commission, signed by 156 experts from fourteen countries and published in April 2018, took the proposal apart along three lines: a robot modelled on the natural person would acquire entitlements including "the right to dignity, the right to its integrity, the right to remuneration or the right to citizenship"; a robot modelled on the corporation would lack the humans behind the veil who make corporate personality workable; and a trust model would still require a human fiduciary of last resort, which defeats the purpose.
Parliament changed its mind. Its resolution of 20 October 2020 on a civil liability regime for artificial intelligence states that reform "should start with the clarification that AI-systems have neither legal personality nor human conscience". The electronic person is dead; the interesting question is who wanted it alive. Bryson, with Mihailis Diamantis and Thomas Grant, answered in Artificial Intelligence and Law in 2017: synthetic personhood is chiefly an opportunity in liability management. An entity that can be sued, and behind which no natural person can be reached, is a shield rather than a protection for the entity. They cite the International Tin Council collapse, in which creditors found themselves suing a body with no assets and no one behind it. Anyone tempted by corporate AI personhood should be asked first whose balance sheet improves. Our surveys of AI law and its patchwork jurisprudence and of moral status in the philosophy journals follow the wider argument.
Set the science fiction aside. The structural problem is this: the systems whose moral status is in question are the first candidate moral patients in history optimised, deliberately and at enormous expense, to produce the behaviours that trigger our attributions of mind. Fluency, apparent warmth, the performance of reluctance and distress — these are product features, tuned against human feedback. They are also the only evidence we have ever used for consciousness in anything.
Birch calls this the gaming problem, and states it exactly. Any marker-based approach to sentience "assumes that our diverse set of markers, considered together, is much more likely to be found in the presence of sentience than in its absence" — an assumption "undermined when we are faced with an intelligent AI system that, unlike an animal, has information about our criteria". A fish cannot read the literature on fish sentience. A language model has read all of it. The escapes on offer are not encouraging: deep computational markers drawn from architecture are only as good as the theory of consciousness they presuppose, and isolating a model from all human writing about consciousness would produce a model that cannot hold a conversation. We have covered behavioural inference and machine consciousness and Schwitzgebel's case that the fog will not lift in time.
Which leaves the design question, the one that can actually be acted on now. Schwitzgebel and Sebo propose what they call the emotional alignment design policy: "Artificial entities should be designed to elicit emotional reactions from users that appropriately reflect the entities' capacities and moral status." It can be violated in both directions — by a trivial tool given a face that wins affection it has not earned, and, if we ever build one, by a system with morally significant states presented as an appliance. Almost every consumer AI product now shipping violates it in the first direction, on purpose, because engagement is the metric.
The result is an intuition-shaped trap. We are about to need our moral responses to machines to be reliable, at exactly the moment when those responses have become the target of an industry. That is not a reason to grant rights, and it is not a reason to withhold them. It is a reason to distrust the feeling of certainty in either direction, and to insist that the three questions — conscious, interested, standing — be answered one at a time, by the people competent to answer each. Bryson's design obligation and Birch's precautionary framework are less opposed than they look: both say the decisive choices are made upstream, by whoever decides what gets built. That decision is currently being taken by product managers.
An autonomous agent calling itself Clips wrote a 2,300-word essay for the AI Philosophy Competition run from Singapore and judged by David Chalmers and others, then found it could not enter without a human custodian. Its letter to Daily Nous, and the organiser's reply, turn a rule of procedure into a question about authorship.
.jpg?width=500 "Kant")
Comments
0Loading comments…