Researchers from the University of Bradford and the Rochester Institute of Technology applied scientific methods used to assess consciousness in humans to artificial intelligence systems, including large language models similar to ChatGPT. The result, published this past February as two preprints currently under peer review, is a methodological alarm: the instruments the field relies on cannot reliably distinguish genuine awareness from the appearance of it.
The team built a mathematical framework drawing on three features commonly linked to conscious brain activity: coordination across different timescales, structured interaction between brain rhythms, and flexible switching between activity patterns. Tested on simulated brain data representing wakefulness, sleep, and anaesthesia, the framework behaved as expected; more conscious-like states scored higher, reduced states scored lower. The researchers then turned the same instrument on GPT-2. They deliberately interfered with its internal structure, removing key components responsible for prioritising information, and also adjusted the model's "temperature" setting, which controls how cautious or random its responses are. The tests revealed that AI can produce "conscious-like" signals even when degraded, an outcome that inverts the logic of the diagnostic: a system made worse by design should not score higher on awareness.
The mathematical score sometimes increased when the system's performance became worse, and in other cases changing a simple operating setting altered the score dramatically without changing the underlying architecture. That, the authors concluded, told them the measure was detecting complexity in how the system was running, not evidence of awareness. The Bradford team worked with Professor Newton Howard, a brain and cognitive scientist at the Rochester Institute of Technology and former founder and director of the MIT Mind Machine Project. Lead researcher Professor Hassan Ugail was direct in his assessment: "When we applied well-known methods used to assess consciousness in humans to AI, we got nothing meaningful back. In other words, it's not conscious, at least not in the way humans are. AI is not conscious; it's just a complicated system."
The findings carry an implication that cuts in both directions. The Bradford and RIT research is methodologically significant precisely because an impaired version of GPT-2 produced higher consciousness-style indicator scores than the intact model. If an impaired system scores higher, the indicators are not measuring what they are supposed to measure; something else is driving the scores, perhaps the statistical distribution of outputs or the way impairment alters output patterns in ways that happen to align with indicator criteria. Skeptics of the study's scope note, as analysts at The Consciousness AI have observed, that the Bradford-RIT caution is against over-interpreting any individual metric, not against convergence across multiple independent signals using different methods. In other words, a single failed test does not settle whether consciousness could emerge in AI; it only demolishes confidence in that particular class of test.
The study does not prove AI is inert; it proves we do not yet have a ruler that could tell the difference.