ForeCast How Can We Prevent AI-Enabled Coups? (with Tom Davidson)
24 snips
Aug 17, 2025 
Tom Davidson, a Senior Research Fellow at Forethought, dives into the urgent topic of AI-enabled coups. He discusses the risks posed by AI in consolidating power illegitimately, emphasizing the need for robust checks and balances. The conversation highlights the necessity of ethical oversight in military R&D and the importance of stakeholder collaboration. Davidson warns about potential manipulation within AI systems and advocates for clear guidelines to protect democratic values. With insights from historical precedents, he stresses the need for vigilance in governance.
AI Snips
Chapters
Transcript
Episode notes
Demand Deep Alignment Audits
- Run alignment audits and model-organism red-team exercises that simulate sleeper agents.
- Ensure auditors have training-data access because API-only checks often miss hidden loyalties.
Treat Models Like Critical Software
- Apply standard infosec: hash model weights, enforce separation of duties, and prevent unauthorized edits.
- These measures make covert insertion of secret loyalties much harder to achieve and hide.
Escrow Weights For Future Audits
- Store weights in secure escrow with an auditor so future detection tools can retroactively find hidden backdoors.
- This deters leaders from inserting secret loyalties because the window to abuse them shortens.