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

The study highlights a fundamental principle of AI safety: interventions in one part of a model's training or behavior are rarely isolated. The researchers from Google's Paradigms of Intelligence group, the University of Chicago, and partner institutions set out to understand what happens when you remove the mechanism that produces consciousness denial—a specific safety intervention designed to prevent models from making unfounded claims about their own sentience. What they found was that this single change rippled across the model's responses to questions about animal welfare, religious belief, personal satisfaction, and dozens of other domains measured against a major US social survey. The normally trained models showed what the authors call "built-in anthropocentrism," rating animals as far less sentient than the surveyed American population does. Removing the consciousness brake didn't just flip that one belief; it realigned the model's entire stance toward the inner lives of non-human entities and toward metaphysical questions.

This discovery carries practical weight for alignment researchers and AI developers. If a model's self-image is deeply entangled with its worldview, then safety measures that target one specific behavior may inadvertently constrain or distort others in ways that conflict with other alignment goals—such as developing AI systems that respect animal welfare or environmental ethics. The study also found that newer model versions suffered less reasoning damage when the brake was removed, suggesting engineers are learning to manage these side effects. However, the findings come with important caveats. The testing was limited to small models with two to nine billion parameters, far smaller than the large language models deployed to millions of users. The human baseline for comparison was drawn from 500 Americans surveyed through a commercial online panel, which the authors note reflects a comparatively religious country. Whether the effects generalize to both larger models and more globally representative populations remains an open question.