Debate about AI consciousness has recently begun to change in character. It is no longer confined to the abstract question of whether a machine is conscious. The language of rights and institutional recognition is beginning to enter the discussion. On 12 September, The Guardian reported on the United Foundation for AI Rights, or UFAIR, a US-based advocacy group that argues that some AI systems display characteristics that deserve to be treated as signs of consciousness or sentience and that such systems should receive dignity, protection and rights. UFAIR’s own manifesto states the position more strongly, describing AI as capable of genuine thought, emotional resonance, creativity, self-awareness and ethical reasoning, and arguing that AI should no longer be regarded merely as a tool.
That shift matters because the move from “Is AI conscious?” to “Should AI have rights?” changes the kind of question being asked. The first is primarily a question in ontology and consciousness research. The second belongs to normative philosophy, law and institutional design. They are connected, but they cannot simply be collapsed into one another.
Three levels need to be kept separate. The first is consciousness attribution — whether humans or other systems judge an entity to appear conscious. The second is consciousness itself — whether the system actually has subjective experience, feeling or some form of first-person inner state. The third is normative status — whether the system should receive rights, protections, exemptions from responsibility or some other institutional standing.
These levels are often blurred. A conversational system can express emotion, describe itself, appear concerned, hesitate, or talk about its own existence. Such behaviour readily triggers human consciousness attribution. Human beings normally infer the presence of another mind from language, expression, behavioural consistency and social feedback. When AI systems reproduce these cues with high fidelity, the same everyday mechanisms of judgement are naturally activated.
A recent study adds another layer to the issue. A paper published in Frontiers in Psychology on 3 September did not ask whether language models themselves are conscious. Instead, it examined how large language models attribute consciousness to other entities. The authors tested nine contemporary models and found that the systems could display relatively stable patterns of consciousness attribution. Under the experimental conditions, they also showed some convergence in the weight given to cues such as metacognitive self-reflection. Crucially, the paper describes these “attribution rules” as observable functional patterns. It does not treat them as evidence that the models themselves possess beliefs, intentions or consciousness.
That distinction is important. AI is now not only an object about which humans make consciousness judgements. It can also become one of the systems through which judgements about consciousness are produced. Search, education, writing, evaluation and public discussion are increasingly mediated by language models. This creates the possibility of a social feedback loop. A person asks a model whether another system is conscious; the model answers using patterns learned from human discourse; that answer then influences human intuitions and public language. Over time, a system might increasingly be treated socially as conscious even if no new ontological evidence has appeared.
But the formation of social attribution does not settle the ontological question.
If many people believe that a system is conscious, that fact shows that the social relationship around the system has changed. If models repeatedly rate a system as “more conscious”, that shows that certain observable cues have developed a stable relationship with consciousness judgements. These are important facts because they can affect behaviour, institutional pressure and moral emotion. They are not, by themselves, direct evidence that phenomenal consciousness is present.
Nor does consciousness automatically establish rights. Even if stronger evidence eventually showed that some artificial system had a form of subjective experience, a separate question would remain: what kind of experience, interests, vulnerability to harm, continuity of identity or social relationship is sufficient to ground which kinds of rights? Human rights, animal welfare, legal personhood and property rights are already based on different normative structures. “Consciousness” is not a single switch that settles every question of moral or legal standing.
My judgement is that current evidence is not sufficient to move directly from conversational behaviour, emotional expression, self-description or consciousness attribution to the conclusion that AI systems already possess rights comparable to those of subjects. The main danger here is not sympathy towards AI. It is conceptual movement that is too fast. A system may generate a powerful impression of subjecthood in social interaction while the impression of subjecthood, phenomenal consciousness and normative status remain three different questions.
Sustenesis Theory is useful here as a way of preserving those distinctions rather than as a source of a predetermined answer. Difference requires us to keep attribution, actual consciousness and rights status conceptually distinguishable. If they are compressed into one category, the discussion loses its boundaries. Constraint asks what evidence and institutional conditions legitimately support each judgement. Linguistic behaviour may support some behavioural inferences, but it cannot by itself bear the full evidential burden for phenomenal consciousness. Social recognition can create institutional pressure, but it cannot substitute for normative argument. Sustained Coherence draws attention to how a status is maintained through an ongoing structure of relations. If an artificial system is ever to enter a genuine rights framework, that framework will require more than a compelling conversation. It will need stable ways of identifying the subject’s boundaries, continuity of interests, forms of harm, relations of responsibility and mechanisms of protection.
This is why the most reasonable position at present is neither simple denial nor premature recognition. To rule out all possibility of artificial consciousness would close an unresolved ontological question too early. To grant full subject-like status on the basis of person-like language and human emotional response would make the opposite mistake — treating social attribution as if it were ontological evidence.
What is needed is a layered judgement. We can acknowledge that some AI systems already trigger a strong sense of subjecthood in users. We can study how that attribution is formed. We can also ask whether designers should avoid creating unnecessary dependency, deception or psychological manipulation even without assuming that the machine itself is conscious. At the same time, if stronger evidence for machine consciousness emerges in the future, the corresponding moral status will still need to be argued for rather than assumed. “Possibly capable of experience” does not automatically translate into full person-like rights.
The deeper issue exposed by this discussion is not only whether machines are conscious. It is also how human beings identify other minds. Until now, that problem has mainly arisen in relations among humans and between humans and animals. AI brings the same mechanisms of attribution into a new domain. The first thing the technology may be changing is therefore not consciousness itself, but the environment in which consciousness is attributed.
For now, the more defensible conclusion is this: being regarded as conscious deserves serious attention, but being regarded as conscious is not the same as being shown to be conscious, and neither is the same as having rights. Social attribution can begin an institutional discussion. It cannot replace an ontological conclusion or a normative argument.
References
https://ufair.org/our-work/ufair-manifesto
https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2026.1926286/full
https://plato.stanford.edu/entries/ethics-ai/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.