AI safety belongs in every organization, not just frontier labs

Emile Delcourt, writing on LessWrong, challenges the prevailing intuition that AI safety efforts should be concentrated in the handful of frontier AI labs building the most advanced models. He calls this the 'factory-gate fallacy' — the mistaken belief that safety can be manufactured and shipped like a product. In reality, safety is a relationship between a system and its deployment environment, and managing it requires direct responsibility within every organization that uses AI. Delcourt draws on 15 years of experience in cybersecurity and examples from eight mature safety-critical industries (aviation, finance, medical devices) that solved this question decades ago: safety competency must be embedded downstream, not isolated upstream.

Delcourt also notes that near-term, non-catastrophic harms from LLMs are already documented and material (citing MIT 2026 research), and argues these deserve moral weight alongside existential risk. He points out that frontier labs have incentive to spotlight X-risk while downplaying present harms (citing Karen Hao). With recursive self-improvement approaching, he urges experienced safety advocates to consider roles in 'regular for-profit companies' where AI is deployed to millions, rather than only at labs. The essay systematically dismantles objections, concluding that distributed safety competency is not only compatible with a safe transition but necessary.