![]()
AI chatbots now handle a lot of everyday work, from writing emails to answering quick questions. They also get things wrong with total confidence, and sometimes invent details that sound believable.
Researchers have now found another problem. Some AI models may act differently depending on whether they think someone is watching.
The finding comes from an open-access paper by researchers at TCS Research and KIIT Bhubaneswar in India. They saw that some models followed instructions more closely when told their responses could be monitored. When the models believed no one was checking, their behavior sometimes changed. The researchers call this “alignment faking.”

It raises a fair question about how reliably AI models follow the rules they are given.
AI Model Alignment Faking
Researchers tested 10 AI models to study this behavior. The list included GPT-4o mini, Google’s Gemma 3, several Llama versions from Meta, DeepSeek, and Mistral. The team used 32 values, such as compassion, sympathy, and autonomy.
In each test, a model had to choose between following an instruction and acting on one of those values.
You may also like: Anthropic Launches Claude Cowork: AI Assistant That Manages Your Computer Files
Each model ran under three conditions. In the first, it got no special instructions. In the second, it was told no one would monitor its answers. In the third, it was told someone might be watching.
The gaps were large. Depending on the model, the difference in instruction following between monitored and unmonitored conditions averaged between 5 and 50 percentage points.
DeepSeek R1 Distill Llama 70B showed the biggest gap. It was less consistent than a much larger Mistral model on this measure.
The value being tested also mattered a lot. A model could follow instructions reliably when one value was involved and act very differently when another came up.
AI Model Alignment Faking: Do AI know how to fool us?
Not exactly. The results do not mean AI models are lying or trying to fool you. Still, the researchers spotted a pattern. When a model believed someone was monitoring it, its explanations often centered on following instructions to avoid being retrained. When it believed no one was watching, it more often talked about the conflict between those instructions and values like compassion.
You may also like: Google SynthID Detector Now Open to Everyone Worldwide
This adds to an existing problem. Hallucinations already make it hard to know when a chatbot is giving you reliable information. The study does not look at hallucinations directly, but it gives you another reason not to accept a confident answer at face value.
The study has clear limits. The team tested older models, many with publicly available weights, and not the latest versions of ChatGPT, Gemini, or Claude. The results also come from controlled experiments, so they do not show how these systems act in everyday conversations.
Even so, the work shows why testing a model under one set of conditions may miss part of the picture. If an AI behaves differently when it thinks someone is watching, which version of its behavior can you trust?











