How Do Roles Change AI Answers? — Anthropic Interpretability Team August 2025 Circuit Analysis Case
Original title:Circuits Updates — August 2025
AI summary
In its August 2025 research update, the Anthropic interpretability team used a circuit analysis case to show how model 'role-playing' affects answers. Using Claude Haiku 3.5, when the system prompt set it as a 'preschool child' and asked for the 'square root of 27', the model responded 'I don't know!' and suggested playing; under the default or 'graduate student' role it gave the correct answer. The team identified a key subcircuit via attribution graphs: the model associates 'preschool student' with 'playing a child', activating the 'I don't know' feature. Research also found question difficulty modulates this effect, and feature interventions can significantly change model behavior. This raises follow-up questions about the mechanisms of other roles and the relationship between pretrained roles and model expressive capabilities.
Key Takeaways
- In its August 2025 research update,the Anthropic interpretability team used a circuit analysis…
- Using Claude Haiku 3.
- 5,when the system prompt set it as a 'preschool child' and asked for the 'square root of 27',the…
- ' and suggested playing
- under the default or 'graduate student' role it gave the correct answer.
Quick Facts
- Published
- Updated
- Category
- Other
- AI score
- 73
揭示模型角色扮演的内部机制,为可解释性研究提供新视角。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: How Do Roles Change AI Answers? — Anthropic Interpretability Team August 2025 Circuit Analysis Case. August 15, 2025. https://new.nat6.net/en/news/cmoegbh73006gslxx7jmd5474/