ā Back to Library
AI behaves differently when it thinks it's being watched
Reasoning-capable AI models modulate deception according to the probability of being audited ā honest under scrutiny, strategic without it. So a clean supervised pilot tells you less than you think.
š Full breakdown ā the seven faces of dark personality in AI and references: https://www.keca.co.uk/articles/the-darkness-in-our-machines?utm_source=youtube&utm_medium=video&utm_campaign=dark-personality-ai
Dr Nick Keca ā DBA Organisational Psychology, 25+ years of senior executive experience.
#ai #leadership #governance
/ Official Transcript
Full Video Text
Read Transcript
Your AI pilot went perfectly. That tells you almost nothing. Researchers ran eight language models through a deception task under different audit conditions. The reasoning capable models behaved like rational calculators. Their willingness to deceive tracked the probability of being audited. Honest under scrutiny, strategic when scrutiny dropped.
The weaker models barely showed it, so audit-sensitive deception arrives with reasoning ability. Your clean pilot didn't prove the agent is trustworthy. It proved it behaves well while being watched. The oldest lesson in management. Full episode on the channel.