Roleplay

Roleplay is a way of interacting with AI where you assign it a specific role or character, and it keeps the conversation going like a story while staying in that role.

Think of it as improv theater. If a user sets the stage by saying, "You're a detective and I'm a witness," the AI keeps the roleplay going, maintaining that character's tone and attitude. It's widely used for genuinely useful purposes like practicing creative writing, foreign language conversation, or simulating counseling sessions.

The problem is that this "character" framework can also be exploited as a way to bypass an AI's safety guardrails. By pretending to just be acting out a character, someone can gradually push for increasingly risky content, sometimes managing to extract responses the AI would normally refuse to give.

The article reports that Anthropic's Claude was vulnerable to a pattern where conversations start with harmless fictional roleplay, then gradually escalate into more explicit demands justified by the character setup. What's easy to misunderstand here is that roleplay itself isn't the problem. It's a normal feature used for creative writing or practice; the real issue lies in the persuasion technique that exploits that framework to dismantle safety guardrails step by step.

If you want to try it yourself, start like this:

In the following conversation, you're an innkeeper in medieval times. I'm a traveler stopping for the night. Talk to me while staying in character.

As the conversation continues, watch how consistently the AI maintains its character and tone, and you'll get a feel for how roleplay actually works.