# Lab Scientists From OpenAI, Anthropic, Microsoft and Meta Jointly Ask for Oversight of Self-Improving AI

The paper appears alongside Anthropic's own disclosure that AI systems drove 26 percent of its research and development in August, up from 1 percent in March.

- Published: 2026-09-29T06:12:03.689Z
- Canonical: https://polylog.news/ai/2026-09-29/lab-scientists-from-openai-anthropic-microsoft-and-meta-join
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Anthropic Research: Frontier Red Team](https://www.anthropic.com/research/team/frontier-red-team)

A working paper published on Monday lists more than 20 authors, among them OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Microsoft chief scientific officer Eric Horvitz, and researchers Geoffrey Hinton and Yoshua Bengio. It argues that automating the AI research pipeline could create a recursive loop in which each system builds a more capable successor faster than humans can evaluate the results. The authors [ask policymakers to impose mandatory oversight](https://www.axios.com/2026/09/28/ai-pioneers-intelligence-explosion) on leading labs before that acceleration begins, and they state the extreme downside case explicitly, including the marginalisation or extinction of humanity.

The paper backs its warning with specific figures rather than argument alone. Anthropic has disclosed that AI systems drove 26 percent of its research and development as of August 2026, up from 1 percent in March, and that AI-generated code accounted for more than 80 percent of approved code by May. OpenAI has said it wants a fully automated AI researcher by 2028.

Those figures are self-reported, and the definitions are the labs' own. No external auditor has verified what counts as AI-driven research. The [circulated summary](https://t.me/aipost/8298) of the paper, and [Bloomberg's account](https://www.bloomberg.com/news/articles/2026-09-28/anthropic-openai-executives-urge-oversight-of-self-improving-ai) of it, both note that the executives are asking governments to regulate the same capability their own companies are developing.

## What this means

The mechanism is regulatory anchoring. The people with the best private information about internal automation rates are publishing those rates alongside a request for rules, which lets them set the terms of the eventual rulebook and raises the compliance bar for smaller labs and open-weight projects that cannot fund evaluation teams. The parties exposed are second-tier labs, open-weight publishers, and any jurisdiction that imports rather than builds frontier models. The precedent is the 2023 voluntary commitments made in Washington, which quickly became reporting requirements for large training runs. The more likely outcome over the coming months is that self-reported research-and-development automation metrics become a disclosure item in legislative drafts and enterprise procurement questionnaires, with labs competing over who publishes the cleanest number. This forecast would be wrong if independent researchers find that the 26 percent figure mostly reflects routine code completion, which would show the metric measures little before it becomes part of any rule.

## What to watch

- Whether any lab publishes a methodology for how it counts AI-driven research, which would let outsiders compare the 26 percent figure against rivals.
- Whether legislators in the United States, the European Union or Australia cite this paper directly in draft text. That would show that the labs' framing has entered statute.
- Whether OpenAI restates or revises its 2028 automated-researcher target, the clearest public schedule for the loop the paper describes.