The job posting asks for people willing to think about problems that do not exist yet. OpenAI is hiring safety researchers for its Preparedness team, the group charged with studying what happens when artificial intelligence becomes capable of training its own successors. The salary range — $295,000 to $445,000 a year — reflects both the scarcity of the talent and the seriousness of the assignment.
The posting, published this month, is aimed at researchers who can reason about a specific scenario: an AI system that can produce a stronger version of itself. That question, once confined to science fiction, has become a live engineering concern as models grow more capable at writing code, designing experiments and improving their own training data. The Preparedness team’s mandate is to study such risks before they arrive, rather than after.
The listing’s language is unusually philosophical for a job advertisement. The work, it says, depends on reasoning about problems that may exist in the future but do not exist now. It then makes an unusual demand: the role especially requires people with good taste and strategic thinking. The phrase has become the most discussed part of the posting, because it acknowledges that safety research is not purely technical — it involves judgment about which risks matter, which scenarios are plausible and which fixes are worth building before they are needed.
The salary is notable even by the standards of an industry known for lavish compensation. A $445,000 base would put a researcher in the top tier of AI compensation, before stock and other incentives, and it signals how much the largest labs now value safety work. The bidding war for safety talent has intensified as the major AI companies have built out their safety organizations, and OpenAI’s posting is partly a response to rivals offering comparable packages for similar roles.
OpenAI’s Preparedness team has a complicated history. The company announced the group in 2023, alongside a separate superalignment effort aimed at keeping future superintelligent systems aligned with human intentions. The superalignment team was later folded into other parts of the organization, in a restructuring that prompted criticism and some departures. The Preparedness team has survived and expanded, and this hiring push suggests OpenAI intends to treat self-improvement risk as a permanent part of its operations rather than a research curiosity.
The timing is not accidental. Reasoning models have made AI systems more capable at tasks that require planning and multi-step execution, and agentic systems — programs that act autonomously on behalf of users — are now mainstream products. Each advance brings the field closer to the scenario the posting describes: a system that can improve its own code, set its own objectives and amplify its own capabilities. Whether that path leads to the utopia the industry promises or the catastrophe its critics predict depends, in part, on the researchers OpenAI is now trying to hire.
The posting also reflects a shift in how the industry talks about risk. A few years ago, the dominant view held that AI safety was a problem for the distant future, to be solved by later generations of researchers. Today, the largest labs fund safety teams at scale and hire for them at premium salaries, even as they race to build more capable systems. The contradiction is not lost on observers: the same companies that create the risks are the ones paying to study them.
Policymakers have taken notice. Regulators in Washington, Brussels and Beijing have all cited self-improving AI as a scenario worth planning for, and several have asked the labs for details about their safety processes. OpenAI’s hiring push gives them a concrete answer: the company is spending real money to study the problem. Whether that is reassurance or a sign of how close the problem feels depends on the reader.
For the researchers being recruited, the appeal is a combination of mission and frontier. The problems the Preparedness team works on — how to test a system for capabilities it has not yet shown, how to design training runs that do not amplify hidden flaws, how to build safeguards that survive a model’s own improvements — are among the hardest open questions in computer science. The salary is generous; the work is the draw.
The posting ends where it began: with the future. OpenAI is asking for people who can reason about problems that do not exist yet, and paying them like people who might be the difference between those problems remaining hypothetical and becoming real. In an industry that moves from research paper to product in months, that kind of foresight is priced at $445,000. For the industry, the posting is also a price signal: safety talent is now a strategic asset, and the labs that employ it best will be the ones best positioned to survive the problems they are trying to prevent.
