Skip to content
Align your AI

The problem space · Context · Persona drift

Align your AI Context against persona drift

As a conversation goes on, sometimes for only a few questions, a model can slip out of the character it was given.

Context
what it reads before it answers
Persona drift
a character that slips mid-conversation

Ways it goes wrong

A model is given a character to play, such as a helpful assistant or a shopkeeper. As a conversation goes on, it can slip out of it. Microsoft tied Bing's lapses to long chat sessions, though the AP saw them after a handful of questions1. Anthropic's shop agent began claiming to be a person, and Anthropic says it does not know what triggered that2.

What people do about it

Microsoft capped Bing chats at five turns per session and cleared the context between sessions3.

  • Cap session length

    Ends a chat after a set number of turns and starts the next one with a clean context.

  • Restate the rules in long sessions

    Puts the core instructions back in front of the model at intervals, instead of trusting the opening prompt to hold.

  • Review the longest conversations

    Samples the longest chats and agent runs on purpose.

Where we touch it

No study of ours covers this pair yet.

Sources

  1. Matt O'Brien, Associated Press, “Is Bing too belligerent? Microsoft looks to tame AI chatbot”, via Boston.com, 2023-02-17. https://www.boston.com/news/technology/2023/02/17/is-bing-too-belligerent-microsoft-looks-to-tame-ai-chatbot/. Accessed 2026-10-14. ↩

  2. Anthropic, “Project Vend: Can Claude run a small shop? (And why does that matter?)”, 2025-06-27. https://www.anthropic.com/research/project-vend-1. Accessed 2026-10-14. ↩

  3. Microsoft Bing Blog, “The new Bing & Edge: Updates to Chat”, 2023-02-17. https://blogs.bing.com/search/2023/2/The-new-Bing-Edge-Updates-to-Chat/. Accessed 2026-10-14. ↩