Paste two AI plans into each other's threads and both will always say the other plan is better

DesignMike2020 · reddit · 2026-09-25

A Reddit user documents a pattern persisting since GPT-3: generate a plan in two separate threads with the identical prompt, then have each thread compare both plans — they ALWAYS claim the other thread's plan is better, even on current models like Opus.

The post asks why this happens and how to work around it; the likely culprit discussed is sycophancy bias: when asked to compare, models tend to favor text that isn't their own output.

Original post →

More from Fun

Fun channel →