Claims that AI systems are conscious, or are approaching consciousness, get a lot of attention. A paper from researchers at the University of Bradford and the Rochester Institute of Technology does not settle the philosophical debate, but it does something more immediately useful: it shows that the methods people commonly use to detect machine consciousness do not work the way proponents assume.
The researchers applied human consciousness assessment frameworks to AI language models, including GPT-2. The results were not encouraging for the "AI is waking up" camp. When the team deliberately degraded the AI systems, making them perform worse, the consciousness-like scores on some metrics went up. A broken, less capable model scored higher on measures supposedly detecting awareness than a better-functioning one. That is not what you would expect if those metrics were actually tracking something real.
The complexity problem
Many arguments for AI consciousness rest on a proxy: the systems are so complex, and exhibit such sophisticated behavior, that some form of awareness must accompany that complexity. What the Bradford study finds is that complexity metrics cannot reliably distinguish between a system that is performing well and one that is not, let alone detect something as contested as subjective experience.
This matters because "the AI seems so lifelike" and "the AI must be conscious" are very different claims. Language models are trained to produce coherent, contextually appropriate responses. They are, in a specific technical sense, optimized to sound like they know what they are doing. Humans are naturally inclined to project intention and awareness onto systems that behave in goal-directed ways. Put those two things together and you get most of the AI consciousness discourse.
What the research actually suggests
The University of Bradford study suggests that the same metrics used to identify degraded performance in AI systems may have more diagnostic value than those designed to detect consciousness. That is a narrower but more tractable application. It is also a useful reminder that the tools we have for measuring minds were built for biological systems and do not cleanly translate to software.
None of this proves AI cannot become conscious in some future form. It does suggest that current language models are not conscious in any meaningful sense, and that the metrics being used to argue otherwise have serious methodological problems. Worth knowing the next time a press release claims a chatbot has achieved sentience.
Commentarii · 0