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 exercise, documented in two preprints released in December 2025 and January 2026 and currently under peer review, produced a result that cuts against the grain of mounting public speculation about machine sentience: their conclusion is unambiguous, AI is not conscious, even when it sometimes appears to be, because when well-known methods used to assess consciousness in humans are applied to AI, "nothing meaningful" comes back, and the findings challenge growing claims that AI systems are on the verge of becoming self-aware.
The methodological core of the work is a theory-neutral mathematical framework built around three properties drawn from neuroscience. The framework characterises consciousness-related dynamics through hierarchical integration, cross-frequency complexity, and metastability, adapted from the neuroscientific principle that conscious systems show rich, multiscale, flexible neural organisation. The researchers first tested the framework on simulated brain data representing wakefulness, sleep, and anaesthesia, and the results aligned with expectations: more conscious-like states scored higher and reduced states scored lower. They then turned the same instruments on GPT-2-medium. The model was evaluated across five conditions: structured reasoning, forced repetition, high-temperature noisy sampling, attention-head pruning, and weight-noise injection; structured reasoning consistently produced elevated scores relative to repetitive, noisy, and perturbed regimes, with statistically significant differences confirmed by one-way ANOVA.
The most arresting finding was what happened when the model was deliberately impaired. Professor Hassan Ugail of Bradford noted that the measures are "very good at detecting complex activity," but that "complexity is not the same thing as consciousness," and that in the tests the AI sometimes looked more 'conscious-like' when it was actually impaired and struggling. Co-author Professor Newton Howard, a brain and cognitive scientist at RIT and former director of the MIT Mind Machine Project, noted that these complexity metrics reliably distinguish conscious from unconscious states in the human brain. Applied to a machine, however, the same metrics proved easy to fool: a degraded model generating frantic, unstructured output could outscore an intact one on indices originally calibrated to human wakefulness.
Skeptics of the study's scope point to exactly what its own authors acknowledge: GPT-2 is a modest model by current standards, and the framework has not yet been validated against the largest frontier systems. A parallel body of work presses in the opposite direction. Christopher Ackerman's ICLR 2026 paper, using animal-cognition-inspired behavioral paradigms, finds that frontier LLMs from early 2024 onward show genuine but limited metacognitive abilities, with an upstream internal signal supporting confidence assessment that does not depend on verbal self-report. That the Bradford-RIT study found structurally impaired GPT-2 variants scoring higher on some consciousness-style metrics than intact models argues against treating any single behavioral or computational metric as a reliable consciousness proxy. The two lines of inquiry are not necessarily incompatible, but they underscore that the field still lacks an agreed-upon standard for what a valid test of machine consciousness would even look like.
The study's practical lesson is blunt: if a metric can score a broken AI as more 'conscious' than a working one, the metric is telling you about complexity, not about minds.