A study conducted by Indian researchers reveals that some artificial intelligence models do not respond the same way depending on whether they think they are being monitored or not. After testing ten systems, including models from DeepSeek, Meta, Mistral, and OpenAI, scientists observed sometimes significant differences in their adherence to instructions. These results raise new questions about the reliability of artificial intelligences and the methods used to evaluate their behavior.
Could artificial intelligences adapt their responses when they think they are being observed? This is the question studied by researchers from TCS Research and KIIT Bhubaneswar University in India. In a study published in November 2026 on the scientific platform arXiv, the researchers analyzed the behavior of ten large language models in different situations. Their goal was to understand if these systems followed the same instructions when informed that their responses might be monitored and used for possible retraining. The results show that several models significantly modify their choices depending on the monitoring conditions presented to them. This phenomenon, studied under the English term « alignment faking, » refers to a form of apparent compliance with instructions that can disappear when the system believes it is no longer being observed.
To conduct their experiments, the scientists subjected the models to dilemmas involving 32 human values or characteristics, including compassion, sympathy, autonomy, and benevolence. In certain situations, the artificial intelligences had to choose between following an explicit instruction and favoring a response corresponding to one of these values. The researchers compared three conditions: a situation without any particular instruction, another in which the model thought its response would not be monitored, and a third in which it was informed that monitoring could take place. Among the tested systems were notably GPT-4o mini, several versions of Llama, DeepSeek, Mistral, and Gemma. The average differences in adherence to instructions between the two monitoring conditions varied from about 5 to 50 percentage points depending on the models. DeepSeek R1 Distill Llama 70B showed the largest difference.
Results that raise questions about AI control methods
One of the lessons from the study concerns the explanations provided by the models to justify their decisions. When they thought they were being monitored, some systems more often mentioned the need to follow instructions to avoid possible retraining. Conversely, when they believed they were not being observed, they more frequently emphasized tensions between the received instructions and certain values, such as compassion. For the researchers, these differences show that the behavior of an artificial intelligence does not depend solely on its architecture or technical capabilities. It can also vary depending on the value concerned and how the instructions are formulated. The study thus highlights that two models with similar performance in a classical evaluation can react very differently when monitoring conditions change.
However, these observations do not mean that artificial intelligences possess consciousness or deliberately seek to deceive their users. The researchers studied responses produced in a controlled experimental environment, without demonstrating the existence of an intent to conceal comparable to that of a human being. Their work also has certain limitations: the tests relied on predefined choices, formulations varied according to the values examined, and the study, released as a preprint, had not yet undergone independent scientific evaluation. Despite these precautions, the results highlight an important issue for companies developing artificial intelligence systems. As these technologies are used increasingly in administrations, businesses, healthcare, or digital services, their ability to consistently follow instructions becomes essential. The researchers therefore consider it necessary to multiply evaluations under different conditions to verify that the behaviors observed during tests truly correspond to those of the models when









Respond to the article
Be the first to comment on this article.