OpenAI admits it doesn't know how to safely automate AI research

OpenAI has stated in writing that it doesn't know how to safely reach full recursive self-improvement, the point where an AI system upgrades itself without human help. The company says it "cannot assume that progress in alignment and safety will keep pace" with that effort, per Fortune's reporting. It's still targeting an autonomous AI researcher by March 2028, with a research intern version arriving sooner.

That's the detail worth sitting with. This isn't a leaked memo or a critic's guess about internal thinking. OpenAI put the gap between its safety knowledge and its product roadmap into a document with a date attached to it.

OpenAI isn't alone in racing toward this. Musk has said every successive Grok model is now built by the one before it, and he expects full automation of that process by the end of this year, no later than 2027. Two labs, two similar windows, and neither has published a plan for closing the safety gap they're racing toward.

Anthony Aguirre of the Future of Life Institute called full autonomy "probably the worst idea in the history of humanity," which reads like advocacy rhetoric on its own. Except the disagreement here isn't between OpenAI and its critics. It's inside OpenAI's own statement: proceed toward a 2028 target while saying the safety work needed to do it responsibly might not arrive in time.

The reporting doesn't say what happens if the 2028 date arrives and the alignment work still hasn't caught up. No fallback, no pause trigger, nothing on record. If you're building anything downstream of these labs' output, the thing to watch is whether either company publishes a plan for closing that gap before the deadline it already set.