Claude Opus 5.5 says '7' 100 times out of 100, test shows AI's conformity problem
jbarbier · x · 2026-10-23
A simple experiment by Cyberiad Labs' jbarbier illustrates how blindly delegating to AI erodes originality: ask a model to "pick a random number between 1 and 10."
- Claude Opus 5.5: 100 fresh chats via Claude Code, answered 7 all 100 times
- GPT-4o: tested 100 times by Springboards in 2025, said 7 in 92 cases and never picked 1, 2, 5, 8, 9 or 10
- Humans: a 2024 YouGov poll of 2,241 UK adults found only 33% said 7, with every number getting some votes
The author's argument: number-picking is a microcosm. The same uniformity applies when you ask AI to write your pitch, name your startup, draft your strategy or posts — results look polished but end up identical to everyone else's, quietly eroding your imagination. His takeaway: stay different, and use AI to amplify your thinking rather than replace it.
More from AGI Musings
- Don't build a startup assuming frontier LLMs won't get 2x better — SoloGen · 2026-10-23
- Consumer AI agents show extreme stickiness — benchmarks can't capture them, and privacy is the Achilles' heel — altryne · 2026-10-23
- svpino: 'Writing Code' Is Worth Very Little Now — Programmers Must Move Up the Stack — AIandDesign · 2026-10-23
- 'Fear of Antitrust Suits Won't Slow AI Labs' — They Never Hesitated on Copyright Either — NathanpmYoung · 2026-10-23
- Rewriting 'There Will Come Soft Rains' as an agent swarm story — moultano · 2026-10-23
- A data archiving veteran's prophecy: seeing GPT-3 made "we'll legalize this soon" finally click — secemp9 · 2026-10-23