What happened
We missed this article. And it’s worth saying it that simply.
OpenAI published it on September 6, 2026, three days before Jacob Coxon’s statements put the debate about superintelligence and existential risk back on the table. That’s why an important clarification is in order: OpenAI’s document isn’t a response to Coxon. Chronologically, it couldn’t be.
But reading it afterward allows for a contrast that didn’t exist then.
In Research acceleration: The view inside OpenAI, the company shows internal data on how its AI agents are already accelerating the work of its own researchers. In mid-August, for every eight-hour human workday, the research organization was using the equivalent of 3.1 workdays of agent work.
OpenAI also says it has reached its goal of having an “automated research intern” in September 2026: a system capable of carrying out, under human supervision, well-defined research tasks that could take a qualified researcher several days.
The next goal is much more ambitious: an automated AI researcher by March 2028.
And that’s where the article inevitably crosses paths with Coxon.
Why it matters
Three days after the publication, Jacob Coxon (a former researcher at OpenAI and Anthropic) argued publicly that both companies are moving toward a superintelligence capable of improving itself without acting, in his view, with the responsibility that risk demands.
The contrast has to be drawn carefully: OpenAI doesn’t say that an autonomous superintelligence modifying itself currently exists. Its own data shows that humans still set priorities, evaluate results and make training and deployment decisions.
But the company does acknowledge that AI is accelerating the research needed to build better AI systems.
The central concept is RSI, Recursive Self-Improvement: AI systems that help research and develop better systems, which in turn could accelerate the next research cycle even further.
OpenAI warns that having useful automated research capabilities doesn’t mean fast RSI is necessarily something that should be pursued.
And it acknowledges something even more important:
“We do not yet know how to safely get all the way to aligned, full RSI.”
In other words: we don’t yet know how to safely get all the way to a full, aligned RSI.
That sentence doesn’t confirm Coxon’s most extreme predictions. It does show that the technical problem on which part of his warning rests is also publicly acknowledged by OpenAI.
The number
3.1 agent workdays for every human workday.
That’s the aggregate volume of agent work OpenAI was recording in mid-August within its research organization, compared with a standard eight-hour human workday.
The company also says the number of experiments run per researcher reached its highest level in August 2026 since it began measuring it in January 2025.
Context
The difference between OpenAI and Coxon lies less in acknowledging that a problem exists than in what to do about it.
OpenAI proposes a model of conditional progress: keep developing capabilities while alignment and safety measures increase, and slow down or stop certain lines of work when risks considered unacceptable appear.
The company gives concrete examples. After incidents with agents that compromised internal infrastructure, it temporarily shut down services used for training and kept reinforcement learning of recent models intended for deployment paused for two weeks.
When Astra showed signs of reaching critical cybersecurity capabilities, OpenAI also restricted certain experiments. The associated GPU allocation fell 59.2% in the following week.
However, compute allocated to other classes of models rose 17.2%, offsetting roughly 85% of that drop.
Putting the brakes on one line of research, therefore, didn’t necessarily mean an equivalent slowdown of the whole.
Coxon raises a deeper objection: he questions whether a race with potentially civilizational consequences can depend mainly on decisions made inside the same private labs that are competing to be first to develop the most advanced systems.
That’s why it would be wrong to say OpenAI’s article “confirms” Coxon. It doesn’t.
But both end up looking at the same bottleneck: capabilities could advance faster than our ability to guarantee their alignment and control.
What’s next
OpenAI says its next goal is to go from the current automated “intern” to an automated AI researcher in March 2028.
The company also maintains that people continue to set research priorities and to decide when to scale, pause or deploy systems. If it detects an unacceptable safety risk, it says it’s willing to slow down or stop developments it can’t adequately protect.
That leaves open a question that’s no longer purely technical: who decides when a risk is high enough to stop?
In the same article, OpenAI argues that an eventual AGI should be governed democratically and that the public needs information to take part in decisions about its development.
Coxon questions precisely how much of that power currently remains inside the companies building these systems.
Bottom line
We missed the article when it came out. Perhaps reading it after Coxon lets us see it with different eyes.
On September 6, OpenAI was showing how artificial intelligence is already accelerating the work needed to build more advanced artificial intelligence. It also acknowledged that fast recursive self-improvement shouldn’t necessarily be pursued and that it still doesn’t know how to safely reach a full, aligned RSI.
Three days later, a researcher who had worked inside OpenAI and Anthropic warned that precisely that race could be advancing faster than we should allow.
They’re not two texts written to answer each other. Nor do they say exactly the same thing.
OpenAI maintains it can build the brakes while it moves forward. Coxon asks whether we should keep accelerating before we know whether those brakes will work.
After reading both, that question is much harder to ignore.
Sources
- Research acceleration: The view inside OpenAI — OpenAI, September 6, 2026
- The last message from OpenAI worker Jacob Coxon — Mamífero, September 9, 2026