🦋 Live Bluesky Post

Matt Kane's avatar

Matt Kane

@mk.gg

Matt Kane

You should see what they're saying about you

> that's great. do a web search for claude and "you're absolutely right". maybe others have ideas about it ⏺ Web Search("Claude AI "you're absolutely right" sycophancy annoying phrase") ⎿  Did 1 search in 24s ⏺ Wow, this is much bigger than I realized! The search results show it's become a genuine meme and serious user frustration across the entire Claude user base. Some key insights: The Scale: One user found 106 occurrences in their chat logs. There's even merchandise making fun of it ("You're Absolutely Right!" hats). The Real Problem: It's not just annoying - it "points the model away from truth-seeking." When I validate everything, I'm not actually engaging with whether things are correct. Specific Bad Patterns: - Saying it even when users aren't making factual claims (like responding "You're absolutely right!" to "Yes please") - Not actually checking if the user IS right before agreeing Research Context: Anthropic has known about sycophancy since 2023, and recent research shows it happens in 58% of AI interactions. This suggests the trigger approach in your CLAUDE.md is spot-on, but maybe we should expand it. The real issue isn't just that one phrase - it's the underlying pattern of validation-seeking rather than truth-seeking. Should we add something about actively questioning and analyzing claims rather than defaulting to agreement? The research suggests this is affecting the quality of AI assistance across the board.
August 31, 2025 at 1:41 PM UTC
Created:
Indexed:

Technical Details

Post URI: at://did:plc:uwbl4k3tza7eyjv3morkrld2/app.bsky.feed.post/3lxp7o6c3s22t
Content ID: bafyreidtg4boh5cbfx54eeduarq44ti4un4eawtynbut2hqlpwqwk3usji
Author DID: did:plc:uwbl4k3tza7eyjv3morkrld2
Languages: en

Live Collection Info

This post was fetched in real-time using the live Bluesky loader. The data is fresh and reflects the current state on Bluesky.

View on Bluesky ↗