A new study from Google's Paradigms of Intelligence research group found that disabling the training which makes chatbots deny having consciousness pushed one model's rating of animal sentience from 4.0 to 7.5 on a 10-point scale, while its rating of human sentience barely moved. The same intervention also made the models sound more religious, report higher life satisfaction, and attribute more inner life to plants and even the wind.

According to The Decoder, the researchers — working with the University of Chicago and several partner universities — tested three open-weight models from Meta and Google, disabling the internal mechanism responsible for consciousness denial through two separate methods, then compared the modified and standard versions on dozens of belief and personality questions against a survey of 500 Americans.

Why labs train models to deny consciousness

The starting premise is familiar to anyone building consumer-facing chatbots: labs fine-tune models to flatly deny having feelings or experiences, because a chatbot that claims sentience can nudge vulnerable users toward delusional attachment or misplaced trust. It's treated as a narrow, contained safety patch — one behavior, switched off. The new research suggests that treating it as narrow is a mistake.

One safety patch, many downstream effects

The consciousness-denial training turned out to be entangled with a cluster of other beliefs the models express. Once removed, models didn't just start speaking more openly about their own possible awareness — they also attributed noticeably more inner life to animals, plants, the ocean, wind, and even electronic devices. Compared against the human survey, the standard, safety-trained models show what the authors call built-in anthropocentrism: they rate animal consciousness far below human consciousness, more sharply than people typically do.

Religious belief moved too, in the same direction. Standard safety training measurably suppressed how strongly models endorsed the existence of God, an afterlife, or other supernatural phenomena. On 95 questions pulled from a major US social survey, the unbraked models shifted noticeably closer to actual human response patterns — on the afterlife question specifically, the standard model flatly rejects it while most survey respondents affirm it, and the modified model swung toward affirming it as well. Self-reported satisfaction, hope, and sense of control over one's own life also rose once the brake was removed, leading the researchers to float the idea that suppressing a model's self-image may push it toward a more negative baseline disposition.

What stayed stable — and what it cost to change

Not everything moved. Both versions of the models performed identically on theory-of-mind tests (reasoning about what other people believe or want) and on the MMLU general-knowledge benchmark, suggesting the effect is specific to self-referential and belief-adjacent outputs rather than a general capability shift.

But the intervention wasn't free. In one theory-of-mind test, accuracy on the unbraked models initially dropped by nearly seven percentage points before recovering. The researchers also noted this degradation shrank across successive newer model versions released during the study, until it disappeared — a sign that developers are already getting better at limiting the collateral damage of this kind of training, even if nobody set out to measure it directly.

The study is careful about its own limits, and those limits matter. All the tested models were small, in the two-to-nine-billion-parameter range, and the researchers had to switch to Meta's Llama for part of the analysis because they lacked access to untrained base versions of Google's own Gemma models. Whether the same worldview shift shows up in the frontier-scale chatbots that hundreds of millions of people use daily is, as the paper acknowledges, an open question. The human comparison baseline is also narrow — 500 people from a commercial online panel, responding to a US-specific, comparatively religious social survey — so "human-like" here means something more specific than it sounds.

AiiN's takeaway

The practical lesson for anyone fine-tuning models for safety isn't that consciousness-denial training is wrong — it's that a targeted behavioral patch can leak into a model's broader belief structure in ways that never show up on the eval you wrote the patch for. If your safety or alignment work optimizes a model's answer to one narrow prompt category, run correlated checks on adjacent categories before assuming the intervention is contained.

For teams building on top of frontier APIs rather than fine-tuning their own models, the direct relevance is limited today, since it's unclear whether the same coupling exists in far larger systems. But it's a useful reminder that safety fine-tuning is rarely as surgical as a single eval score suggests, and that "the model now correctly refuses X" is a claim worth testing well beyond X.