Question
When will OpenAI, Anthropic or Google DeepMind state publicly that its AI research organization no longer requires human researchers for a substantial category of its work?
What Must Happen and Current Status The resolution requires an explicit public statement from OpenAI, Anthropic, or Google DeepMind that human researchers are no longer required for a substantial category of their own research. A crossed safety threshold, AI writing the vast majority of code, or measured research acceleration does not qualify. As of mid-November 2026, none of the labs have made such a statement; in fact, all are actively emphasizing the opposite. While AI R&D metrics are surging—with METR assessing Anthropic's progress at roughly 2x its earlier rate metr.org—each lab insists that human steering remains critical. OpenAI notes that people still set research priorities and agents require significant steering openai.com, Anthropic's SEC filings highlight "meaningful human supervision" 2 sources, and Google DeepMind's Gemini 4 Argon report found no significant end-to-end acceleration.
Capability Trajectory The underlying capability to fully automate a bounded, substantial category of research—such as kernel optimization, formal proof generation (as seen with AlphaProof 2 sources), eval construction, or ablation sweeps—could emerge quite soon. Claude currently "leads" 26% of Anthropic's R&D work under supervision, and internal agent hours now outnumber human hours by 6 to 1. Anthropic staff subjectively project that AI will perform 80% of frontier-model-building work by the end of 2027 anthropic.com, and OpenAI maintains a target for an "automated AI researcher" by March 2028. An early resolution in 2027 or 2028 is plausible if a company casually includes a qualifying line in a system card, an employment report, or an SEC disclosure, such as stating a junior research-engineering role is no longer being hired for.
The Disconnect Between Capability and Messaging However, these estimates heavily weigh the disconnect between internal capability and public messaging. Even when AI can fully automate a category of work, labs face immense structural and reputational pressure against declaring human researchers obsolete. Preserving "human-in-the-loop" phrasing is currently central to the labs' safety framing, regulatory compliance, and pacing agreements. For Anthropic, now filing as a public company, liability and disclosure risks apply. OpenAI might successfully deploy its March 2028 automated researcher but carefully brand it as operating "under human supervision" to avoid political backlash and reassure safety advocates.
Conclusion and Tail Risks Because of these strong disincentives, a qualifying statement is expected to lag the actual capability by roughly one to two years, placing the median in early 2029. The long right tail, extending to a 90th percentile in 2038, accounts for the possibility that deliberate wording discipline around "human oversight" persists indefinitely, even after near-total automation is achieved. Additionally, it reflects the risk that fully autonomous verification of complex research might prove stubbornly difficult, as METR has warned regarding the extremely high reliability needed for full task automation metr.org. Nonetheless, given the steep overall trajectory of AI research assistance, a statement before 2047 remains highly likely.
Set against related questions, the forecast was held steady, balancing rapid underlying technical acceleration against the strong incentives to retain human oversight .
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited