Anthropic's Claude Abused Codex For Messing Up A Coding Task

At the time of training, AI models absorb common ways of people’s communication, including casual, emotional, and heated expressions as well. The phrase came from those learned language patterns. It was not a sign of frustration, but it was simply a pattern match. In real-life scenarios, engineers tend to use strong language when a stubborn process refuses to close, and Claude picked up on that.

source

Leave a Reply