In two new studies, 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 findings, published as preprints and currently under peer review, challenge growing claims that AI systems are on the verge of becoming self-aware, and reveal why measuring consciousness in machines is far more difficult than many people assume.
The framework, developed by Professor Hassan Ugail of Bradford's Centre for Visual Computing and Intelligent Systems and Professor Newton Howard of RIT's School of Individualised Study, characterises consciousness-related dynamics through three properties: hierarchical integration, cross-frequency complexity, and metastability. The researchers first tested the framework on simulated brain data representing conditions such as wakefulness, sleep, and anaesthesia; results aligned with expectations, with more conscious-like states scoring higher and reduced states scoring lower. Adapting neuroscience concepts of temporal integration and metastability to transformer models, they then computed a composite dynamical metric from activation time-series of GPT-2-medium across five conditions, including structured reasoning, forced repetition, high-temperature noisy sampling, attention-head pruning, and weight-noise injection.
The central, troubling result was that 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, perhaps how impairment alters output patterns that happen to align with indicator criteria. Professor Ugail described this as revealing a fundamental misunderstanding about machine intelligence: "These kinds of measures are very good at detecting complex activity," he said. "But complexity is not the same thing as consciousness." In artificial systems, unlike in the human brain, the same mathematical patterns can be dialled up or down simply by changing settings, making them unreliable as any kind of test for awareness.
The work does not leave the field empty-handed. What these methods can do, the researchers say, is help engineers understand when an AI system is functioning well or when it is beginning to break down, a utility they argue could prove valuable for AI safety, reliability, and future regulation. Skeptics of the Bradford-RIT conclusions point to a distinct body of competing evidence: the most technically significant counterclaim to emerge in 2026 concerns what researchers have described as functional introspective awareness in current large language models, with evidence suggesting that current LLMs possess some degree of functional introspective awareness of their own internal states. The Bradford-RIT caution, however, is against over-interpreting any individual metric, and both positions are right: the appropriate response is to treat convergence of multiple indicators as raising the probability estimate above zero without treating any specific finding as decisive.
The most honest read of the Bradford-RIT data is that the field's go-to consciousness yardsticks are broken, and fixing the ruler has to come before any verdict on the mind.