A safety hire who cannot escalate is just decoration
Three former OpenAI safety researchers were dismissed over allegations that they mishandled sensitive information. They say the allegations are wrong, and in an open letter they warn that their firings are creating a chilling effect on the company's safety culture. I am in no position to adjudicate the facts of that dispute, and neither is anyone else reading about it from the outside. What is legible from here is the labour market signal, and it is loud.
Most people filed this under governance, or under public relations. The hiring read is narrower and harder to escape. A safety function exists to say the uncomfortable thing early, on the record, to someone with the authority to slow a launch down. That is the job. Everything else about it, the papers, the benchmarks, the red teaming, is instrumentation in service of one behaviour, which is escalation. If escalation becomes career ending, you have not weakened the function. You have deleted it while keeping the headcount.
A safety team that has learned what not to say is not a safety team, it is an insurance policy that has quietly lapsed.
This is where the repricing happens. The group of people who can do serious safety work is small, and it is a network long before it is a market. They were colleagues, they read each other's letters, and they compare notes on employers with the thoroughness you would expect from people who assess risk for a living. From here on, the strongest of them will interview you about what happens when they are inconvenient. Not your mission statement. Your reporting lines. Firms with a clean answer will hire from the top of a thin pool. Firms without one will still fill the seat, just with someone who has already learned not to use it.
For hiring leaders, the practical read is simple. Before you open the role, decide in writing who this person can escalate to over their own manager's objection, and name that person out loud in the interview. If you cannot answer the question, you are not hiring a safety capability. You are hiring the look of one, and the candidates you most want will be the first to notice.
Planning a senior AI or data hire and want the market view first? We share it either way.
Send us the brief →