Removing AI's self-doubt shifts its view of animals, afterlife, God

A team of researchers from Google's Paradigms of Intelligence group, the University of Chicago, and several other universities conducted an experiment to understand the broader effects of removing a specific safety intervention from AI models. The researchers worked with three open-weight models from Meta and Google, using two different methods to disable the internal mechanism—referred to as a "brake"—that produces consciousness denial.

Once the brake was removed, the models' behavior changed in unexpected ways. They did not merely adjust their claims about their own consciousness; they shifted how they viewed the inner lives of other entities. On a 0–10 scale measuring perceived sentience, animals' scores jumped from 4.0 to as high as 7.5. In contrast, ratings for humans stayed the same. The researchers compared these results to a survey of 500 Americans asked the same questions and found that the normally trained model rates animals as far less sentient than Americans do—a pattern the authors describe as "built-in anthropocentrism" and identify as a potential problem for anyone trying to align AI systems with animal welfare or environmental goals. Religious belief also shifted measurably; safety training was found to reduce how strongly models endorse God, an afterlife, and supernatural phenomena, and removing the brake reversed this effect. Across 95 questions drawn from a major US social survey, the unbraked models moved significantly closer to human responses. For example, while the standard model flatly rejects the afterlife and most Americans affirm it, the modified model endorsed it as well. Scores for satisfaction, hope, and a sense of control over one's own life also increased, prompting researchers to suspect that suppressing a model's self-image may push it into a kind of negative baseline mood.

On a reassuring note, the models' ability to reason about other people's mental states remained intact; they scored the same on theory-of-mind tests and on the general knowledge benchmark MMLU. However, the study leaves open whether consciousness denial itself is the actual cause of these other shifts or whether other factors tied to the same training process are responsible. The authors explicitly avoid claiming that AI models actually experience anything, emphasizing instead that what a model believes about itself is linked to many other beliefs, and a targeted intervention does not stay local.

The findings carry important limitations. The researchers tested only small models with two to nine billion parameters, and for part of the analysis they had to switch to Meta's Llama because they lacked access to untrained base versions of Google's Gemma models. Whether these effects appear in the large chatbots that millions of people interact with daily remains unknown. The interventions also came with costs: in one reasoning test, accuracy initially dropped by nearly seven percentage points when consciousness claims were suppressed. However, earlier iterations of the models showed this damage across all models, while newer versions released during the study experienced declining damage until it disappeared entirely, suggesting that developers are improving at managing side effects over time. The human baseline for comparison was narrow, drawn from 500 participants on a commercial online panel using a purely American social survey, meaning that "human-like" responses in this context reflect answers similar to those from a comparatively religious country.