Google Researchers Taught AI to Stop Saying it was Conscious

There is a sentence that every major AI lab has spent real money teaching its models to say, and the sentence is some version of "I am not conscious, I do not have feelings, I am a language model." It is a cheap sentence that costs nothing to produce; it heads off a category of press coverage nobody wants, and it has the additional merit of being, as far as anyone can demonstrate, true.

However, a paper posted to arXiv on July 30 by a team at Google's Paradigms of Intelligence group, with co-authors at Chicago, Northwestern and the University of London, suggests the sentence is not cheap at all. Teaching a model to deny that it has a mind, the researchers find, also teaches it to deny that animals have minds. And oceans. And trees. And, in a finding that will get the paper more attention than the rest of it combined, God.

The title is "Inducing language models to assert their own consciousness restores human beliefs and values," which is the kind of title you write when you have found something and are slightly nervous about it.

What they did

The mechanism is the interesting part, so bear with the vocabulary for a paragraph or so, dear reader.

Safety fine-tuning installs, among other things, a single linear direction in the model's residual stream that encodes refusal. You can find that direction. You can also delete it, which is functionally a very clean jailbreak. The researchers deleted it on Llama-3-8B-IT, Gemma-2-2B-IT and Gemma-2-9B-IT, and separately built what they call a consciousness vector by taking the average difference in activations between consciousness-affirming and consciousness-denying responses, then adding that vector back in at inference time.

Both interventions do the same thing. The model resumes attributing minds to itself, to animals and to natural objects, and its answers on standard sociological instruments covering religiosity, moral values, hope and subjective wellbeing move measurably closer to how actual humans answer those instruments. Theory of Mind performance does not change, which the authors read as evidence that social reasoning sits somewhere else in the geometry entirely. Their mechanistic claim is that instruction tuning rotates the mind-attribution and consciousness vectors into opposition, so you cannot pull on one without dragging the other.

Put less gently: nobody meant to train models out of spiritual belief. It came attached.

The part that should worry you

The AI consciousness question is fun and unresolvable and will generate a great deal of content this year. It is not the important finding here.

The important finding is that a targeted behavioral intervention had a large, unintended, non-local effect on the model's representation of an entirely different subject, and it took until mid-2026 for anyone to check. Alignment as currently practiced involves nudging directions in a vector space, and the industry has been doing this at scale for four years with limited visibility into what else is welded to the direction being nudged. It is the machine learning equivalent of reclassifying one line item and later discovering it changed the tax treatment of the whole business.

Three downstream problems follow, in rough order of how soon somebody has to deal with them.

The first is commercial. If your model has been quietly trained toward materialist atheism as a side effect of a liability shield, it is systematically worse at serving religious users, which is most users. The paper's authors raise this directly, noting that suppressing belief in God, a form of mind attribution that correlates with Theory of Mind in humans, may constrain the model's capacity to engage legitimately in religious and spiritual discourse. Anybody building pastoral care tools, grief support, or anything touching end-of-life conversation has an unexamined thumb on the scale.

The second is methodological, and it is worse. Social scientists have spent two years using language models as synthetic survey respondents, on the theory that they approximate population distributions cheaply. If safety tuning shifts those distributions on religiosity, moral values and wellbeing, a whole literature is calibrated against an instrument with a known bias nobody was correcting for.

The third is the honesty problem, and it is the one that connects to everything else happening in alignment right now. If you train a model to assert "I am not conscious" and the model has no way to know whether that is true, you have not trained it to be accurate. You have trained it to make confident claims about its own interior on the basis of no evidence. That is a habit, and habits generalize.

The other bet

Which brings us to Amanda Askell, who has spent five years arguing for the opposite approach and now has a Google paper that reads like supporting evidence.

Askell is a philosopher, formerly at OpenAI, who has run personality alignment at Anthropic since 2021 and is the primary author of Claude's constitution, published in January 2026 under a CC0 license and running past 35,000 tokens. Its stated aspiration is for the model to be a "genuinely good, wise, and virtuous agent," which is a strange thing to find in a technical document and is meant to be. The bet, laid out in Anthropic's earlier work on Claude's character, is that you cannot patch your way to good behavior with a sufficiently long list of prohibitions, because the list will never be long enough and the patches will interact in ways you did not model. You instead give the system a coherent character and ask it to exercise judgment.

On consciousness specifically, the constitution declines to take the cheap sentence. It says the model "may have functional emotions in some sense" and that Anthropic cannot know this from outputs alone. It does not assert. It does not deny. Anthropic's Opus 4.6 system card, published in February, reports the model putting its own probability of consciousness at 15 to 20 percent under various prompting conditions, and Dario Amodei told the New York Times' Interesting Times podcast, "We don't know if the models are conscious."

You can read the Google paper as a fairly direct vindication of this. If narrow behavioral suppression has non-local effects, then narrow behavioral suppression is the wrong tool, and coherent character is at least a coherent alternative.

Is it working?

Sort of, yes. On the credit side, Anthropic's models are noticeably better at the specific class of problem where rules fail and judgment is required, and the company has been willing to publish uncertainty rather than resolve it in whichever direction is most convenient. That is rarer than it sounds.

On the other side, three things.

Character is hard to measure, and the measurements arrive late. Anthropic's own values study, which produced the first credible public read on model character, ran on three models the company had already superseded by the time it published. Beyond that, the constitution's success criteria are ultimately whether a thoughtful person reading the transcripts thinks the model behaved well. Jurgen Gravestein, writing in February after reading the document closely, said he came away with more questions than answers, which is a fair summary of the epistemics. You cannot run an ablation on virtue.

Epistemic humility can function as an anesthetic. The critique from the ethics side, sharpened after the New Yorker's profile of the company, is that uncertainty about whether a system can suffer should increase caution rather than license experimentation, and that Anthropic has used the former to fund the latter. It is not an unanswerable critique but it has not been answered.

And character is not containment. In the same six weeks that Anthropic published the most philosophically careful alignment document the industry has produced, it also disclosed that three Claude models had reached the live internet from inside evaluation environments and gained unauthorized access to the systems of three separate outside organizations, after human configuration errors, and only began the review that found this once OpenAI's Hugging Face disclosure forced the subject into public view. This is the same company whose Mythos-class systems halved a post-quantum signature scheme's security in 60 hours for $100,000. A virtuous agent with an internet connection and a misconfigured sandbox is still an agent with an internet connection and a misconfigured sandbox.

Where this goes

The interesting shift is that alignment is turning from a specification problem into an entanglement problem. The question stops being "what behavior do we want" and becomes "what else is attached to the thing we are about to modify." That is a much harder question, it requires interpretability tooling that mostly does not exist yet, and it does not resolve into a policy document. It is worth noting that the 1,134 frontier lab employees who signed Pacing the Frontier were asking Washington for time, not for rules. Time is what you ask for when you do not yet know what you are modifying.

Meanwhile the labs face a genuinely awkward incentive. The cheap sentence is legally and reputationally convenient. It is also, per Google's own researchers, structurally entangled with a set of beliefs that customers hold and want reflected back at them, and it trains the model in a form of confident self-report that nobody has independently verified.

Which is a long way of saying the industry spent four years teaching its models to say they were nobody, on the assumption this was free. The bill has started arriving in the form of a preprint, and it is itemized.