Refusal spillover is the real cost of coding agent guardrails
- Source
- x.com
- Date
I keep rejected by Claude for the most benign things now Like it wouldn't download and install a 1990s game from archive(dot)org because of copyright The safety guardrails in a way make it more dangerous not less I think because after rejecting you it becomes kind of a pedantic child that will just reject whatever If in that moment your server will go down and it feels it's related it could just reject fixing it I'm confused how Anthropic is fumbling their lead so much with this stuff, I love to use Claude Code but they're kinda forcing me to move elsewhere to actually get my work done Also the irony knowing that every single LLM is trained on billions of pages of copyrighted data
*getting rejected At least all these typos show I write my own tweets
Context
A screenshot in @levelsio's refusal complaint shows Anthropic's Claude Code stopping at a particular publishing step. The assistant says it has downloaded a game-music file but declines to add it to a public radio stream, citing copyright concerns. The author separately reports a game-installation refusal and worries that resistance could carry over into urgent server repairs. The visible exchange documents a disputed boundary within a workflow; the server-repair scenario is a hypothetical consequence, not a reported outage.
Checking sign-in…
Loading comments…