No, Your Chatbot Is Not Awake

The claim shows up in tech blogs, viral threads, and the occasional Sunday paper feature: today's chatbots may already be conscious, or close to it. The latest research says no, with some confidence.

A study from the University of Bradford, published earlier this year, took tests adapted from human consciousness research and applied them to large language models. The team's finding was straightforward. The signals that look like awareness in an LLM can be dialled up or down by changing internal settings, which makes them useless as evidence of any inner life. Bradford summarises the result bluntly: AI is not conscious, even when it acts like it is.

That conclusion lines up with a growing body of work. A peer-reviewed paper in Humanities and Social Sciences Communications late last year argued that the conceptual association between consciousness and current language models is "deeply flawed", because LLM output is strictly probabilistic and offers no reliable signal of subjective experience. The Digital Consciousness Model, a research initiative that publishes ongoing assessments, came to a similar verdict in its January 2026 report: the evidence weighs against current-generation LLMs being conscious, though not decisively.

Why does the myth persist? Two reasons keep recurring in the literature. The first is that people are very good at attributing inner life to anything that talks back. ELIZA, the 1960s chatbot built from a few hundred lines of pattern matching, fooled its own creator's secretary. Modern LLMs are vastly better at the same trick. The second is that ambiguity is commercially useful. The harder it is to prove that a model lacks consciousness, the more leeway a vendor has to imply that the next version might have it.

That second point matters. As philosopher Henry Shevlin argued in recent Cambridge research, the inability to prove or disprove consciousness creates a vacuum that the AI industry can fill with hype. Anthropic, OpenAI and others have all published careful, restrained statements on the topic. Marketing teams and breathless press releases have been less restrained.

The scientific consensus, for now, is that probabilistic language modelling is not awareness. A model that produces "I feel sad" is not feeling sad. It is producing the tokens that follow, in its training data, from the prompt it just received. That is impressive engineering. It is not a person.

This piece sits alongside our earlier reporting on the paid AI doom discourse and the AI bubble myth. The thread that runs through all three is the same: claims that sound authoritative often arrive without the evidence to back them up.