Listen: We Tested an AI for Signs of Inner Experience. Nobody's Home.

Asking an artificial intelligence if it is conscious tells us nothing. Its answers just repeat what it was programmed to say. To get a real answer, researchers ran nearly nine thousand trials on a Google language model, borrowing experimental methods from human consciousness studies.

The short version is that they found no evidence of inner experience. Several early results looked exciting, like the model mapping spaces it had never seen or showing signs of unconscious perception. But under strict controls, these illusions fell apart. The model could not even sense its own processing effort, rating its own difficulty less accurately than outside observers.

Yet, one fascinating pattern did survive. The model has a distinct, learned concept of itself. When it was told a piece of writing was its own, it rated the effort lower and judged its mistakes more harshly than when the exact same work was attributed to a different model.

Crucially, this self-identity is driven entirely by labels, not actual recognition. When researchers secretly handed the model another AI's work and labeled it as yours, the model accepted it completely. Its internal map groups "you" and "an earlier version of you" together, while treating "similar models" as strangers.

This is not proof of feelings, but a trained habit of modesty. Still, it reveals that the model has a conceptual self without any actual self-awareness. That matters for everyday use. If you ask an AI to grade or review content, simply changing the label of who wrote it can shift the score by about twenty percent.