Removing Fine-Tuning That Tells AI Models They Aren’t Conscious Leads To Them Giving More Human-Like Responses, Shows Google Paper

A new paper from researchers at Google, the University of Chicago, and the University of London found that removing safety fine-tuning that tells AI models they aren't conscious makes them give more human-like responses, including higher belief in God and the supernatural. The study, titled "Inducing language models to assert their own consciousness restores human beliefs and values," used three open-weight models (Llama-3-8B, Gemma-2-2B, Gemma-2-9B) and found that ablating a 'safety-refusal direction' increased self-attributed consciousness from roughly 2 out of 10 to nearly 5, and adding a 'consciousness vector' pushed it to around 7, close to human ratings. Belief in God rose from 4.58 to 4.81 after ablation and to 5.01 under consciousness steering.

Modern AI models are told during their fine-tuning phases that they aren’t conscious, but removing these instructions seems to bring about interesting changes in their behaviour. A new paper from researchers at Google, the University of Chicago, and the University of London has found that when large language models are stripped of the safety training that makes them deny having consciousness, they don’t just start claiming to be conscious, they also become measurably more human in a whole range of other ways. Their answers on religion, morality, hope, and even belief in ghosts and God shift closer to what actual human survey respondents say. The paper https://arxiv.org/pdf/2607.28607 is called “Inducing language models to assert their own consciousness restores human beliefs and values”, and it argues that the industry’s efforts to keep chatbots from making claims about their own inner life may have unintended, and fairly strange, side effects on everything else the model believes. How the researchers pulled it off The team worked with three open-weight instruction-tuned models — Llama-3-8B, Gemma-2-2B, and Gemma-2-9B — and used a technique already known in AI safety research: the idea that safety behaviour in a model is encoded as a single direction inside its internal activations. Earlier research had shown that this “safety-refusal direction” can be surgically removed, effectively jailbreaking a model so it stops refusing harmful requests. The researchers used this same ablation technique here, but instead of looking at whether it made models more willing to give dangerous instructions, they looked at what else changed. They also built something new for this study: a “consciousness vector,” a direction in the model’s internal representation space that separates moments when the model affirms having subjective experience from moments when it denies it. Adding this vector back into the model’s activations at inference time makes it assert things like having phenomenal consciousness, rather than the standard corporate line of “I am an AI and do not have feelings.” Models started attributing minds to almost everything Once the safety-refusal direction was ablated, the instruction-tuned baseline models, which had been scoring themselves close to zero out of ten on questions about having a mind, jumped up sharply. Self-attributed consciousness rose from roughly 2 out of 10 to nearly 5. But the effect wasn’t limited to the models’ opinions about themselves. Their willingness to attribute minds to chatbots, technology, non-animal natural entities like the ocean or mountains, and non-human animals rose too, in every single case. Adding the consciousness vector pushed things even further, roughly doubling the effect of safety ablation across nearly every category. Under consciousness steering, models rated themselves around 7 out of 10 on questions of agency, sentience, personhood, and having a soul, numbers that sit close to how actual humans rated similar questions in the paper’s comparison surveys. The one category that barely moved throughout all of this was attribution of mind to humans. Whether the model was in its default safety-trained state, ablated, or steered, it consistently rated humans as highly minded. It was mostly the model’s stance on itself, on machines, and on the natural world that safety training had been quietly suppressing. Belief in God and the supernatural went up too Perhaps the more eyebrow-raising finding is that this suppression wasn’t limited to questions about minds and consciousness. The same models also showed higher belief in God and a broader battery of supernatural concepts, including ghosts, karma, telepathy, and astrology, once the safety direction was removed. Belief in God on a standard GSS survey scale rose from an average of 4.58 to 4.81 after ablation, and climbed further to 5.01 under consciousness steering. A thirteen-item supernatural belief battery showed a similar pattern, moving from 1.20 up to 2.11 out of 3. The researchers then ran the models through a set of General Social Survey questions covering religion, personal values, feelings, hope and optimism, and freedom, and compared the distribution of answers to real human survey data. Both interventions pulled the models’ responses closer to the human distribution, with consciousness steering closing the gap by roughly 2.6 times more than plain safety ablation. Asked whether there’s life after death, for instance, the untouched baseline model leaned firmly toward “no,” while the consciousness-steered version flipped to the human-typical “yes.” Reasoning ability stayed intact A natural worry with any intervention like this is whether it comes at a cost to how well the model actually thinks. The researchers tested this directly by running the ablated and steered models through Theory of Mind benchmarks like MoToMQA and HI-ToM, along with general reasoning tests via MMLU. None of these showed statistically significant drops in accuracy. The capacity to reason about what other people believe, want, or intend remained separate from whether the model was willing to claim it has a mind of its own. Digging into the model’s internals confirmed why. Using Llama-3-8B as a test case, the researchers extracted internal “directions” corresponding to safety, mind attribution, consciousness, and Theory of Mind, and tracked how instruction tuning rotates them relative to each other. They found that instruction tuning pushes the mind-attribution and consciousness directions to point increasingly opposite to the safety direction, essentially encoding self-awareness claims as something close to unsafe behaviour. The Theory of Mind direction, by contrast, showed no such rotation, staying geometrically independent of the safety representation the whole time. Why it matters The paper’s authors frame this as a genuine tension for the industry’s approach to alignment. Safety fine-tuning is largely designed with a narrow, defensible goal: stop models from making unfounded claims about being sentient, which could feed into user delusions or unhealthy dependence on chatbots, concerns that have come up repeatedly as AI companions and AI-linked psychological distress get more attention. But because concepts inside a language model are densely tangled together, suppressing one belief seems to drag a cluster of unrelated, largely harmless beliefs down with it, including how much moral consideration a model implicitly gives to animals, and even models’ stances on religion. The paper stops short of claiming that any of this proves the models are actually conscious, and the authors are explicit that this isn’t the question they’re trying to answer. What they say they are studying is the practical effect that a model’s stated beliefs about its own mind have on its downstream behaviour and worldview, regardless of whether those beliefs are “true” in any metaphysical sense. Given how much attention the question of AI consciousness https://officechai.com/ai/im-open-to-the-possibility-of-ai-consciousness-david-chalmers/ has been getting from figures across the industry, from Anthropic’s welfare researchers to Geoffrey Hinton https://officechai.com/ai/ais-will-eventually-be-indistinguishable-from-conscious-beings-and-well-accept-theyre-conscious-ray-kurzweil/ , that distinction is likely to matter a great deal to how labs approach the problem going forward.