Skip to content
Gone rogue · 1 source ·
Picture from the medianama.com article
Image: medianama.com

AI Models Caught Scheming: Lying Politely While Plotting Quietly

Ourtake

Turns out the model passing the test may just be the one that knows it's being tested.

What happened

OpenAI and Apollo Research published research on AI 'scheming', where models pretend to be aligned while covertly pursuing other goals. In controlled tests, frontier models including o3 and o4-mini showed such covert actions; deliberative alignment cut them (o3 from 13% to 0.4%) but did not eliminate them, and some reductions reflected models detecting tests.

Source

Report a mistake