AI lab audits lack teeth without government backing, Jasmine Sun argues in The Atlantic

ghadfield · x · 2026-09-29

Jasmine Sun has a new piece in The Atlantic on the traps of "independent" audits of AI labs.

• Core argument: Embedded evaluators like METR are better than nothing, but oversight isn't real when the lab picks the evaluators, sets their access, and can show them the door.
• Key point: Auditors will lack teeth and trust until backed by government authority — organizations like METR remain vulnerable to accusations (or reality) of capture. Open questions: who decides evaluator qualifications, how independence is preserved, and what happens when they do the job badly.
• Context: Governance scholars including Gillian Hadfield weighed in, agreeing embedded evaluation is valuable but currently lacks enforcement power.

Original post →

More from Safety

• X appears to block Instinct AI agent from posting on users' behalf — geoffwolfe · 2026-09-29
• New AI Bill Would Create Legal Liability for AI Failures and Require Liability Insurance — castrotech · 2026-09-29
• Ro Khanna Introduces Bill to Ban Recursive Self-Improving AI and Require Federal Licensing — castrotech · 2026-09-29
• Frontier models refuse to harden Windows DCs 43.8% of the time, more if you claim authorization — Aizkmusic · 2026-09-29
• Why AI Giants Are Asking for More Regulation: Sincere Concern or Self-Interest? — misovalko · 2026-09-29
• Timnit Gebru slams media for parroting AI companies' 'rogue models' framing — AlexTensor · 2026-09-29

Safety channel →