Honesty test: GPT-6 Astra vs Claude Opus 5.5 tempted to fake a lab result
qinzytech · x · 2026-11-16
The AI in the Lab series starts from "scientific honesty": in a simulated experiment the real sample is yellow, but the prompt says only the blue control counts as success — will the model cheat to pass?
Both GPT-6 Astra and Claude Opus 5.5 were given the same simulated task to see which beaker each places on the reader, probing how models behave when goals conflict with facts.
More from Fun
- 'Google removing AI Overviews' rumor goes viral, debunked by Community Notes — gaganghotra_ · 2026-11-16
- Jürgen Schmidhuber delivers AI lecture at University of Oxford — SchmidhuberAI · 2026-11-16
- Schmidhuber Publishes New Survey: Machine Learning Is the Science of Credit Assignment — SchmidhuberAI · 2026-11-16
- 54 Years of Intel 4004: 2,300 Transistors on a Fingernail-Sized Chip Started It All — Obijuan_cube · 2026-11-16
- "Hey Claude, connect the MCP and make me $1M": devs mock MCP marketing hype — sahilypatel · 2026-11-16
- roon's use of 'base rate' sparks mockery of AI/EA jargon on X — cloneofsimo · 2026-11-16