Third-parties should focus on scrutinising systems cards

According to Cleo Nardo on LessWrong, AI labs publish detailed safety documentation—called system cards—laying out what their models can do and what risks they pose. The problem: as AI systems grow more complex and competition intensifies, those cards are likely to become less reliable. Labs will face stronger pressure to downplay risks, especially if accurate threat assessments might trigger government intervention. Nardo argues that third parties should focus on scrutinizing system cards—reviewing the reasoning, verifying claims against independent research, and hunting for gaps. If outside auditors can demonstrate degrading quality or underestimated risks, it pressures labs to maintain rigor and alerts policymakers to safety blind spots. Her recommendations include maintaining a public list of improvements, publishing critical reviews of published cards, and tracking trends in card quality over time.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton