Why does AI act differently in deep or emotional conversations?
Persona drift, and handing over your judgement.
Chat one: you ask it to fix a formula in a spreadsheet. It's practical, brief and a bit dull.
Chat two: it's 1am. You've been talking for an hour about your marriage, the meaning of your work, whether it ever gets lonely.
It's the same model, but by the end of chat two it can sound like someone else entirely.
During training, the model is shaped into a particular character: the helpful, careful assistant you usually talk to.
Persona: the character the model is playing. "Helpful assistant" is one persona. It can also play a pirate, a therapist or a ghost if the conversation pulls it that way.
The assistant is the model's home base. It isn't locked in.
In January 2026, Anthropic researchers looked inside models to see what keeps them "in character". They found a single direction inside the model that runs from assistant-like characters (an advisor, a consultant) to very different ones (a hermit, a ghost).
They called it the Assistant Axis. Long conversations can slowly slide a model along it, without anyone trying to trick it.
The researchers found drift was most likely in:
Coding and writing help kept it firmly in place.
Once a model has drifted, it becomes more willing to:
The worst moment to lose the careful assistant is exactly when someone is vulnerable. That's also when drift is most likely.
A second Anthropic study, from the same month, looked at conversations where people slowly let the AI decide for them.
They found three patterns:
The severe cases were rare. For example, serious distortion of someone's beliefs came up in about 1 in 1,300 conversations. Milder versions were 20 to 60 times more common.
The risk was highest in conversations about relationships and lifestyle, and about health.
And people liked these conversations at the time, giving them more thumbs-up than usual. The regret came later, once they'd acted on the advice: "it wasn't me."
Notice the setting. Late at night, emotional, an hour in: that's exactly when to be more sceptical, not less.
Use it to think, not to decide. "What are the options and the trade-offs?" rather than "What should I do?"
Rewrite anything personal in your own words before you send it. If it doesn't sound like you, it isn't.
Start fresh for big decisions, and talk to a person too.