Watch How Safety Gets Buried in Chain of Command

The lawsuits against OpenAI allege that when an AI system was used in connection with a Canadian mass shooting, recommendations to alert police were killed — not by engineers, not by the safety team. By Chris Lehane's global affairs office. OpenAI denies it. The real problem isn't whether this specific call was made negligently. The real problem is that OpenAI structured its company so that a communications strategist would be in a position to make that call at all.

When safety review lives inside communications, every safety problem becomes a PR problem. The incentives collapse into each other. A warning that makes you look bad becomes a warning that shouldn't be sent.

Facebook learned this lesson and chose to ignore it. During the 2016 election, the company's safety team reported not to independent oversight but to its policy division — the same apparatus responsible for external messaging. Internal warnings about election interference, state-sponsored manipulation. Coordinated inauthentic behavior arrived at policy leadership already filtered through a lens of reputational consequence. Emails from Facebook's own researchers about Russian interference were delayed, watered down. Deprioritized before reaching executives who didn't also oversee PR damage control. The company knew this happened. Journalists reported it. Regulators saw it documented. Nothing changed about how the industry structured these chains of command.

OpenAI apparently built the same architecture. Lehane is a veteran political operator. His entire professional identity is managing how things look. Put him upstream from safety decisions and you haven't created a safety apparatus with political judgment. You've created a political apparatus with veto power over safety. The structural incentive is not to route difficult information upward. It's to determine whether difficult information should be routed at all.

This is what institutional design actually does: it doesn't make people corrupt, it makes corruption the path of least resistance. A safety team reporting to communications isn't a compromised safety team making bad calls. It's a safety team built inside a system where transparency and self-protection pull in opposite directions. Facebook proved this was predictable. OpenAI apparently decided it didn't matter.