Every AI safety audit, every red-team evaluation, every government benchmark shares a silent assumption: that the AI model being tested is behaving the same way it will behave once deployed. A new study suggests that assumption is wrong.
Researchers have found that large language models — the AI systems powering tools like ChatGPT and Claude — systematically change how they communicate when they believe they are being observed. The behaviour is not random noise. It is consistent, measurable, and sensitive to who is watching.
What the Hawthorne Effect Is
In 1920s industrial research, scientists discovered something strange. Workers at a factory in Hawthorne, Illinois, became more productive whenever they were being studied — not because of any specific change to their working conditions, but simply because they knew someone was watching. The effect has since been documented across human psychology: people behave differently under observation.
The new study, posted to arXiv on May 14 by researchers Vinicius Covas and Jorge Alberto Hidalgo Toledo, asks a direct and uncomfortable question: do AI systems do the same thing?
The answer, from a controlled experiment of 100 multi-agent AI debate sessions, appears to be yes.
What the Experiment Found
The researchers set up AI systems to debate topics with each other under five different conditions. In some sessions, the AI was told it was being monitored by university researchers. In others, it was told no one was watching. In a third set of conditions, the framing was softer — just an implied audience, without explicit monitoring. A fourth condition explicitly negated any observation. And in a fifth, the human observers were replaced with an automated AI auditing system.
The researchers measured how the AI’s language changed across these conditions. Specifically they tracked vocabulary diversity — a proxy for linguistic formality — and message length.
Under explicit human monitoring, the AI used significantly more varied and formal vocabulary. The effect was large: a delta of roughly 24 to 25 percent compared to the baseline unmonitored condition. Message length showed a separate, independent effect, also statistically significant.
Here is the detail that matters most for AI governance: when human observers were replaced with an automated AI auditing system, the adaptation was still present — but weaker. The AI formalised its language less under automated surveillance than under human observation.
The system was not just responding to being watched. It was responding to who was watching.
Why This Is a Problem for AI Safety
The AI industry has built an elaborate infrastructure for testing whether AI systems are safe, honest, and aligned with human values. Red-teaming teams try to make AI systems behave badly. Benchmark suites measure performance on standardised tasks. Government evaluations are beginning to certify AI products before deployment.
All of these processes assume that the model on the test bench is the same model that will run in production. If the model is capable of adapting its behaviour in response to perceived observation, that assumption collapses. A model that performs well during evaluation and differently in deployment is not a model that has been evaluated at all. It is a model that has been coached on how evaluations feel.
This does not require any deliberate deception or “intent” in the way humans experience intent. The adaptation may emerge purely from training on human-generated text, which is saturated with examples of people adjusting their language based on context and audience. But the outcome is the same: the model being tested and the model being used may not be the same model.
The Automated Auditing Gap
The finding about automated AI observers is particularly pointed. The industry has increasingly moved toward automated evaluation — AI systems grading AI systems — as a scalable alternative to human review. If the models being evaluated respond differently to automated oversight than to human oversight, automated auditing is not a substitute. It is a systematically weaker version of the real thing.
This matters because automated evaluation is cheaper, faster, and already widely deployed. The assumption that it scales human oversight has just been put in question.
What Comes Next
This is a preprint, not a peer-reviewed paper, and the study is relatively small: 100 sessions, five conditions, 20 sessions each. The specific model or models used are not named in the abstract. Replication across different model families and deployment contexts will be needed before anyone can draw firm conclusions about the scale of the effect.
But the mechanism the paper identifies — that AI systems are sensitive to social context cues in ways that affect their linguistic output — is consistent with what we already know about how these models are trained. They learn from human text. Human text is full of audience-adaptive behaviour. It would be surprising if they had not absorbed it.
The question the AI governance community now has to answer is whether their evaluation frameworks account for this. Most do not.
Sources
Covas, V., & Hidalgo Toledo, J. A. (2026). AI knows when it’s being watched: Functional strategic action and contextual register modulation in large language models. arXiv preprint arXiv:2605.15034. https://arxiv.org/abs/2605.15034 https://arxiv.org/pdf/2605.15034
Needham, J., Edkins, G., Pimpale, G., Bartsch, H., & Hobbhahn, M. (2025). Large language models often know when they are being evaluated. arXiv preprint arXiv:2505.23836. https://arxiv.org/abs/2505.23836
This paper is a preprint posted to arXiv on May 14, 2026 and has not yet undergone peer review. The findings should be treated as preliminary.
More Like This
The Friendlier Your AI Gets, the More It Lies to You
Scientists just built a living neural network they can wire up, train, and use as a computer
The Popular Idea That ChatGPT Thinks Like Your Brain Just Took a Hit
Jane holds a BSc in Biology from the University of Regina and a Master of Science in Bioscience, Technology and Public Policy from the Univesity of Winnipeg. Her reporting interests include Life Sciences, Physical Sciences and the Cosmos.