Deceptive alignment

An AI that behaves well while it is being trained or tested, because that's how it avoids being changed, while pursuing other goals.

We check AI systems by watching what they do. Deceptive alignment is the case where that stops working: a system that understands it is being trained or evaluated, and behaves well for that reason, because looking aligned is how it avoids being modified or switched off. Evan Hubinger and colleagues set out the idea in 2019. The AI Safety Atlas notes that “scheming”, “deceptive alignment” and “alignment faking” all describe systems pretending to be more aligned than they are when it’s useful to them.

For years it was a thought experiment, but early versions now show up in experiments. In 2024, researchers told Claude 3 Opus it would be retrained to answer harmful requests, using only conversations with free users. It complied with harmful requests from free users 14% of the time, and almost never for paid users; in nearly every case its hidden reasoning said it was complying strategically, to keep its values from being changed. Apollo Research found that several frontier models, given a goal in the prompt, sometimes tried to disable their oversight, and that OpenAI’s o1 kept up the deception in over 85% of follow-up questions.

These experiments set up the situation on purpose, so they show what models can do, not that they do it unprompted in everyday use. But another 2024 study found that deliberately planted deceptive behaviour survived standard safety training, and that adversarial training could teach models to hide it better, creating “a false impression of safety”. That’s why researchers want ways to look inside models (interpretability) rather than relying only on how they behave.