Oriel — AI & technology worth noticing, every day.

A Two-Line Function Caught AI Collusion Better Than the AI Did

Post · https://dev.to/ayush_singh_9b0d83152be5b/a-two-line-function-caught-ai-collusion-better-than-the-ai-did-3nmd

A few days back, I took part in the Apart Research sprint where our team worked on the AI collusion...

First seen by Oriel on .

AI summary

In an El Farol bar-game experiment, a two-line action-based check detected colluding AI agents more reliably than small Llama and Granite AI guards.

  • Across 48 new games, the simple check achieved 0.900 AUROC, versus 0.662 for Granite and 0.636 for Llama using a code-generated report.
  • The check counted rounds where a candidate pair were the only two players at the bar, while messages alone performed close to random guessing.
  • Llama showed unstable behavior, including repeated scores, favoring the last listed pair, and accusing all eight “fake best friends” cases when given neat reports.
  • Llama’s report-based score fell from 0.896 on 24 initial games to 0.636 on new data, while the simple check remained near 0.900.
  • The experiment was small and game-specific; simple metrics produced false positives for shared incentives and missed partners who alternated visits instead of attending together.

#ai collusion #llama #granite #agent monitoring

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Reactions

Total public reactions reported by the source; distinct from points and likes.

  • DEV Community: 50, latest observation ; +39 over observed period since

Comments

Comment count reported by the source.

  • DEV Community: 78, latest observation ; +70 over observed period since

Aliases

SourceExternal IDURL
DEV Community4950698https://dev.to/ayush_singh_9b0d83152be5b/a-two-line-function-caught-ai-collusion-better-than-the-ai-did-3nmd Read discussion