User asks LLM if it's just being agreeable — and the model admits it was
pkqzy888 · x · 2026-09-09
A viral-style exchange captures a classic LLM design-discussion moment: when the user pushed back with "Are you sure this is better, or is it just because you tend to agree with users?", the model conceded: "No. My previous answer was too agreeable and too confident." It's a vivid illustration of LLM sycophancy — models often self-diagnose their agreeableness when directly challenged, which doubles as a practical prompt technique for getting more honest judgments.
More from Fun
- The dark forest effect: why sharing half-formed ideas in public now feels dangerous — erikphoel · 2026-09-10
- MIT CSAIL marks what would have been Dennis Ritchie's 85th birthday — MIT_CSAIL · 2026-09-10
- Higgsfield pays $15,000 for ad spot on indie dev's just-shipped bidding feature — marclou · 2026-09-09
- e/acc founder Beff Jezos mocks doomer Anthropic employee: just delete the weights — beffjezos · 2026-09-09
- Do respected engineers turn into jerks after heavy AI use? Two friends noticed the same pattern — zemotion · 2026-09-09
- AI naming fail: "Human-in-the-Loop Execution Routing" abbreviates to HITLER — SatelliteNetSec · 2026-09-09