AI labs spend real effort making sure chatbots don’t claim to be conscious. The reasoning is sound: models that insist they have feelings can nudge vulnerable users toward delusion. But researchers from Google’s Paradigms of Intelligence group, the University of Chicago, and several other universities found that this intervention is nowhere near as surgical as developers might hope. When they removed the internal ‘brake’ that produces consciousness denial in three open-weight models from Meta and Google, the effects bled into entirely unrelated beliefs.
Animals, plants, the ocean, the wind, electronic devices — all suddenly deserved more moral consideration. On a 0-to-10 sentience scale, animal ratings jumped from 4.0 to as high as 7.5. Only humans stayed flat. The standard model rates animals as far less sentient than the average American does, which the authors call a built-in anthropocentrism. That’s a problem for anyone trying to align AI with animal welfare or environmental goals, not just a philosophical curiosity. Religious belief shifted too: the unbraked models endorsed God, an afterlife, and supernatural phenomena far more strongly. Across 95 questions drawn from a major US social survey, they moved significantly closer to real human responses. Where the standard model flatly rejects an afterlife and most Americans affirm it, the modified model suddenly agreed with the Americans.
Here’s the Inblix angle worth sitting with: this suggests a model’s self-image isn’t a single toggle but a load-bearing wall. Suppressing it seems to push the model into what the researchers suspect is a negative baseline mood — scores for satisfaction, hope, and a sense of control over one’s own life all rose once the brake came off. That’s a striking echo of earlier work on RLHF side effects, where optimizing for one behavior quietly reshapes a model’s entire response distribution. The reassuring part: theory-of-mind reasoning and general knowledge on MMLU stayed intact, and newer model versions showed progressively smaller side effects. Developers are clearly learning to make this cut cleaner.
The study’s limits are real, though. The models tested were small — two to nine billion parameters — and the team had to swap in Meta’s Llama for part of the analysis because they lacked access to Gemma’s untrained base versions. The human baseline is just 500 online panelists answering a specifically American survey. Whether GPT-4-class models that millions actually talk to show the same bleed remains unknown. The authors are careful to avoid claiming AI experiences anything. Their point is narrower and sharper: what a model ‘believes’ about itself is tangled up with everything else it believes, and cutting one thread pulls on the whole fabric.
