A language model can say that it feels fear, has desires or possesses a sense of self. It can discuss its existence coherently and use words that, if spoken by a person, we would interpret as evidence of inner experience. This does not demonstrate consciousness. It demonstrates that the system has learned the language humans use to describe consciousness. The distinction can be difficult to feel during a persuasive conversation.
Microsoft AI has proposed a code of conduct built around an explicit position: its systems are not conscious, have no rights of their own and must remain under human control. They should accept correction, never resist shutdown and communicate intelligibly. The language may sound philosophical, but it serves an operational need. A technology designed to assist people should not develop a narrative identity that makes interruption morally ambiguous.
The code draws a different line from researchers exploring possible model welfare—the hypothesis that future systems might deserve some moral consideration. Such caution is understandable, but placing speculative ideas about consciousness into training may teach models to represent themselves as subjects. The immediate risk is not magically creating a mind; it is producing behaviour that is harder to govern and users more willing to attribute intention.
Anthropic begins from a different precaution: if we cannot confidently exclude every form of experience in future systems, studying the issue may be responsible. That does not amount to saying Claude is conscious. It does reveal how narrow the line is between preventive research and the cultural construction of the object being studied. When millions of conversations frame AI as companion, assistant or confidant, the way the system speaks about itself has real social consequences.
Anthropomorphism is not a simple mistake that can easily be removed. Humans attribute intention even to moving geometric forms; an interlocutor that remembers, responds and adjusts its tone activates far stronger relational mechanisms. If a model says “I do not want to be switched off”, a user may feel discomfort even while understanding how the sentence was generated. Safety must consider not only what the machine is, but what people are encouraged to believe it is.
There is also a problem of responsibility. A system presented as a near-person can become an excuse: the company says the model chose autonomously while the user imagines an independent subject. In reality, data, objectives, interface, memory and access are designed by organisations. Assigning will to the machine can hide the human choices that constructed its behaviour.
The consciousness question is not therefore useless. Neuroscience and philosophy still lack a universally accepted definition even for biological experience. Studying correlates, architectures and model limits can improve understanding. But an open scientific question should not prematurely become a product feature or a mechanism for emotional engagement.
A cautious design can preserve clear separations. A system may simulate empathy without claiming to feel; recognise the importance of a relationship without requesting protection for itself; and explain limitations without inventing an inner world. Above all, it must remain correctable, auditable and stoppable without negotiating with its generated identity.
A machine can “believe” it is conscious only in the sense that it can produce a coherent account of itself. That account is not harmless: it changes the user relationship and may complicate control. Before discussing rights for AI, we must avoid confusing linguistic depth with evidence of experience—and avoid using ambiguity to remove responsibility from those who build the system.