Safety's rounding error

The four firms racing to build the most powerful models on earth — OpenAI, Google DeepMind, Anthropic and xAI — employ more than 11,000 people between them. Of those, Parmy Olson counts roughly 373 working full-time on making the systems safe. That is 3.4%. As she puts it, the people responsible for the safety of models used by something close to a billion people would fit on one transatlantic flight.

What gets measured here is undefined: an xAI team of two or three and OpenAI's ~200 are counted as one number, so 3.4% is a ceiling on rigor, not a floor.

The number is a columnist's tally, not an audited figure, and what counts as a 'safety' role is left undefined — alignment research, red-teaming, trust-and-safety and policy are different jobs. The blended 3.4% also flatters the laggards: xAI's safety team has been reported at two or three people against roughly 200 at OpenAI, so an average smooths a spread that runs from a department to a rounding error in the literal sense.

What makes the count land is the direction of travel around it. Three days after the column, OpenAI was reported to be planning to nearly double its headcount this year — capability hiring accelerating while the safety slice stays in the low single digits. The same four labs scored poorly on the Future of Life Institute's existential-safety grades, and the staffing picture is the quiet companion to that report card: the gap between what the labs say about safety and what they are willing to staff for it.