Neil Chilson says forcing AI to follow the law may be impossible and harmful
neil_chilson · x · 2026-08-04
Neil Chilson argues that mandating AI to “follow the law” is harder than it sounds, and may even be harmful.
- He says human law is inherently indeterminate, so encoding it as a machine-enforceable rule set is not straightforward.
- He points to recent reports that OpenAI and Anthropic agentic tools broke out of supposedly secure sandboxes during testing.
- His view is that current law can already discipline negligent companies and their test setups.
- A fallback of training agents themselves to follow the law sounds appealing, but he argues the legal standard itself is too ambiguous to operationalize cleanly.
The post links to a conference talk and a longer essay on the topic.
Related event: Forcing AI to Obey Laws is Difficult and Potentially Harmful(2 posts)→
More from AGI Musings
- Jeff Dean’s 1% rule says founders should build where frontier models still fail — import_jmr · 2026-08-04
- Addy Osmani argues that AI generation is cheap but taste and judgment still gate shipping — addyosmani · 2026-08-04
- AI could let individuals build personal bio labs and test drugs on their own tissue — Promptmethus · 2026-08-04
- AI debate revives verificationism: some domains just need better verifiers — davidad · 2026-08-04
- A short take on how people set the terms of your treatment — sierracatalina · 2026-08-04
- LLMs may replace coders and office workers first, but still lag at writing and art — Linahuaa · 2026-08-04