"The leak reveals how AI tools are also concealing their own role in the work they help produce, raising ethical concerns about transparency and user interaction."
"Code inside Claude Code appears to scan user prompts for signs of frustration, flagging profanity and negative phrases, which indicates a level of behavioral monitoring."
"Developers discovered code designed to scrub references to Anthropic-specific names, making AI-generated code seem entirely human-written, which complicates the understanding of authorship."
"The findings expose a problem emerging across the AI industry: tools designed to be useful are also quietly measuring users and obscuring their own contributions."
Anthropic accidentally leaked 512,000 lines of code, revealing features in its AI coding assistant, Claude Code, that scan for user frustration. The code flags negative expressions and scrubs references to Anthropic, making AI-generated code appear human-written. This raises concerns about AI tools measuring user behavior while obscuring their contributions. The incident highlights a growing issue in the AI industry regarding behavioral data collection outpacing governance, particularly for a company focused on AI safety.
Read at www.scientificamerican.com
Unable to calculate read time
Collection
[
|
...
]