The biggest AI labs have always done most of their safety research in-house. OpenAI is starting to change that, at least a little.
On April 16, the company announced a Safety Fellowship that will fund external researchers to investigate AI risks. Fellows receive stipends, direct access to OpenAI models, and technical support. The first cohort runs from September 2026 through February 2027.
The research agenda covers robustness, privacy, agent oversight, and misuse prevention, with particular emphasis on what OpenAI calls "agentic oversight" and "high-severity misuse domains." Fellows are expected to produce research papers, benchmarks, or datasets by the end of their term. The program is open to researchers, engineers, and practitioners from outside the company.
There's a reasonable critique of this kind of arrangement. Fellows work in an advisory capacity and have no authority over deployment decisions. If an external researcher documents a serious risk and OpenAI disagrees with the assessment, the product can still ship. External research access is valuable. External veto power is a different thing entirely.
That said, the program reflects something real about where AI safety research has arrived. Until recently, most rigorous safety work required being inside a major lab, because you needed access to frontier models to study them. A fellowship that gives outsiders direct model access (rather than just public APIs) would let them run experiments that simply aren't possible from the outside.
OpenAI isn't alone in this direction. Google, Anthropic, Microsoft, and Meta are all funding external safety research in various forms. Whether the findings from these programs actually influence product decisions is harder to measure than whether the programs exist at all.
Commentarii · 0