AI Consciousness Tests Keep Getting the Wrong Answer. Here's Why.

When consciousness researchers applied standard assessment tools to an AI language model, they found something paradoxical. As the model was progressively degraded, its consciousness score went up.

That is the central finding of a study from the University of Bradford, published in February 2026. Professor Hassan Ugail and Professor Newton Howard, a brain and cognitive scientist formerly with MIT's Mind Machine Project, tested GPT-2 using frameworks originally designed to measure awareness in humans and animals. The results showed these tools produce unreliable and sometimes inverted results when applied to AI systems. A damaged model registered as more "conscious" by these metrics than an intact, functioning one. Since a degraded system is clearly not more aware, the finding strongly suggests the frameworks are detecting something other than genuine awareness when pointed at AI.

"AI is not conscious; it's just a complicated system," Ugail said. The study's implications go beyond this single experiment. They raise a harder question: whether any existing tool can currently distinguish between genuine awareness and the sophisticated pattern generation that large language models produce.

A separate study from Brown University, published this week, adds a new layer to the picture. Those researchers found that language models develop internal representations that track real-world plausibility with roughly 85% accuracy, describing this as evidence of "basic understanding." The finding is genuine, but it is easy to misread. A model encoding causal constraints about the world is not the same as a model being aware that it is doing so. Understanding, in the narrow technical sense the Brown study tested, and consciousness are different things.

Public interest in AI consciousness tends to rise whenever a new model produces unusually coherent or emotionally resonant output. The belief that chatbots are developing genuine feelings has persisted despite strong scientific consensus against it. The Bradford study adds a further complication: not only are current AI systems not conscious by any reasonable definition, but the tools that might someday detect machine consciousness are not yet up to the task either.