From thought experiment to roadmap
The question of whether AI systems will start improving themselves has stopped being hypothetical and started appearing in company plans with dates on them, Fortune reported on Saturday.
OpenAI has deployed what it calls an automated research intern, a system it says can complete tasks that would take a skilled human researcher a few days. The company is targeting March 2028 for a fully automated AI researcher. It has also stated that it does not yet know how to safely reach aligned, full recursive self-improvement, and that it cannot assume progress in alignment and safety will keep pace with capability.
That is an unusual pair of statements to publish together: a target date for a capability, and a concession that the safety problem attached to it is unsolved.
Different labs, different destinations
Elon Musk has described the position at xAI differently. Humans are getting progressively less involved in improving Grok, he has said, with each successive model built by the one before it, and the company is aiming for full automation by the end of 2027 at the latest.
Anthropic sits between the two. The company says Claude now leads about a quarter of its model research and development and can carry complex technical work end to end from a high-level prompt under human supervision — while separately arguing for pacing, and saying it would slow or pause its own work if competitors did the same in a verifiable way.
Microsoft’s Mustafa Suleyman has framed the goal as humanist superintelligence: systems with high degrees of autonomy that stay, in his phrasing, calibrated and within limits. That is a statement about where the boundary should be, not about when it will be reached.
The objection, and the caveat
Anthony Aguirre, president of the Future of Life Institute and a physics professor at the University of California, Santa Cruz, told Fortune that full autonomy is “probably the worst idea in the history of humanity to do this”.
John Thickstun, a computer science professor at Cornell, made the deflating point from the other direction: AI has been helping build AI for years in supporting roles, and the line between a useful research assistant and a self-improving system is drawn by degree rather than by kind.
Both things are true at once, which is what makes the dates hard to read. An automated researcher that closes a loop humans currently close is a continuation of existing practice. An automated researcher that sets its own research agenda is not. The company timelines do not always distinguish between the two.
What to watch
The specific number worth tracking is OpenAI’s March 2028 target, because it is the only one attached to a defined artefact rather than to a trend. The second is whether the labs publish the evaluations they use to decide a system has crossed from assistant to autonomous researcher — the threshold that would make any of these dates checkable from outside.
