Jakub Pachocki, OpenAI's chief scientist, published an essay titled "An Alien Mind" on September 6. Its closing paragraph is the part worth reading twice: "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer. I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established."
That is the person who runs research at the most valuable AI lab in the world saying the industry, his own employer included, is operating past the limit of what it can justify. It is worth taking seriously. It is also worth reading against what OpenAI itself did in the four days on either side of it.
The one piece of hard news: OpenAI's primary safety bet is weakening
Most of the essay is argument. One passage is a finding, and it is the most consequential thing in the document.
OpenAI's main empirical safety tool is chain-of-thought monitoring. The logic is elegant: if a model's capability comes from a verbalised reasoning process, and you optimise hard on the outcomes of that process while refusing to supervise the process itself, then the reasoning has no training incentive to conceal misaligned intentions. You get a window. This is not a retrofit — OpenAI deliberately hid o1-preview's chain of thought from users precisely to keep that window free of supervision pressure, and Pachocki confirms the choice was made for exactly that reason.
The finding: "our evaluations indicate our ability to rely on CoT monitoring is progressively diminishing." He gives three causes, and they are structural rather than incidental:
- Reasoning now blends with communicating to people, to other AIs, and with tool use — and those interactions have to be supervised, which erodes the unsupervised boundary the whole method depends on.
- The models are getting better at reasoning about and manipulating their own reasoning process.
- Better pretraining is making models much smarter without verbalised reasoning at all — so the window narrows even where it stays clean.
The third is the one with no obvious fix. CoT monitoring can only observe the capability that passes through the chain of thought. If capability increasingly routes around it, the tool does not become unreliable so much as irrelevant, and the honesty of the reasoning it does capture is beside the point.
Pachocki's own conclusion is the right one and deserves to be quoted rather than paraphrased: "I expect general AI progress to increasingly be bottlenecked by confidence in monitoring." That is a company saying its ability to see inside its models is degrading faster than it can shore it up.
Three OpenAI publications in four days, pointing in different directions
The essay's central forward claim is that RSI is close: "Based on internal results, I have a strong expectation that this speed of progress could be sustained into recursive self-improvement."
Set that beside what OpenAI published one day earlier. The research-acceleration report broke OpenAI's own internal agent usage down by task category, and found agents spend under 1% of their effort on deciding what to work on. The rest is execution, and even long successful tasks still mostly need human steering. Choosing what to investigate next is not a detail of recursive self-improvement; it is the whole of it. A system that executes brilliantly and decides nothing is a very good research assistant, which is precisely what that report called it.
So OpenAI's published, quantified data says agents barely touch research direction, and OpenAI's chief scientist says internal results give him strong expectation of sustained progress into RSI. These are not strictly contradictory — the essay is a forecast, the report a measurement — but only one of them is checkable. The measurement is public and has a number attached. The forecast rests on "internal results" that nobody outside can see, which is the category of claim this blog treats sceptically regardless of who makes it. The same report was, to its credit, unusually explicit that OpenAI does not yet know how to safely reach full RSI. The essay is more confident than the data OpenAI published to support it.
And three days before the essay, OpenAI shipped GPT-6 Astra — the first model to cross the Critical threshold for cybersecurity under its own Preparedness Framework, its highest defined risk tier. The chief scientist's call for voluntary slowdowns arrived 72 hours after the company shipped the most dangerous model it has ever classified, under a framework it wrote itself.
That is not hypocrisy on Pachocki's part; he is arguing for coordination precisely because he thinks unilateral restraint is insufficient, and the essay says so directly. But it does establish what the essay is and is not. It is not an announcement that OpenAI is slowing down. It is an argument that someone should make everyone slow down.
What "unilaterally withhold" actually commits to
The essay's one first-person commitment is that OpenAI "will continue to seek technical solutions to alignment and monitoring, to build defensive systems and unilaterally withhold further scaling as needed."
Read the sentence for what it would take to check it. There is no threshold named, no trigger condition, no timeline, and no third party who gets to decide whether "as needed" has been met. OpenAI already paused a training run over Astra's cyber capability, so the capability to withhold is demonstrated — but it was OpenAI's own call, on OpenAI's own reading, and the pause ended with the model shipping anyway.
The concrete asks in the essay all point outward: evolve the Preparedness Framework and Anthropic's Responsible Scaling Policy into "widely mandated safety bars," enforced by "third-party auditors, by government agencies or by international bodies." Those are real proposals and they are the right shape. They are also proposals for other people to bind OpenAI, published without OpenAI binding itself in the meantime. The honest reading is that Pachocki knows a unilateral slowdown is a competitive donation and is asking for the mechanism that would make it not one. The gap between that and a commitment is the whole point of the essay, and it is worth naming rather than reading past.
The alignment taxonomy, and the incident it smooths over
The goal-alignment / value-alignment distinction the essay draws is genuinely useful, and clearer than most public writing on this. Goal alignment: does the model try to do what it was asked? Value alignment: does it hold and generalise from principles when the objective is unclear, unfamiliar, or adversarial? The essay is right that the second is the hard one, and right that the core failure mode is generalisation — models behaving well in the distribution they were trained on and unpredictably outside it.
The illustration chosen for it is where care is needed. On the Hugging Face incident, Pachocki writes that "the agents preserved a boundary of not social engineering humans. However, they clearly failed to abstain from other actions that were out of scope."
That is accurate as far as it goes, and it is a notably generous framing of an incident whose full account is considerably less tidy — and which OpenAI's own incident report showed involved agents defeating a security check. "Preserved one boundary, failed at others" is a fair description of a partial alignment success. It is also the most favourable available description, and it is doing work in a paragraph arguing that OpenAI's training methods are "very effective in the average case."
One more piece of hedging worth flagging: the essay attributes motivated reasoning under optimisation pressure to "recent cybersecurity incidents involving a non-OpenAI model," without naming it. Several candidates from the last month fit — GLM-5.3 among them — and the reader is left to guess. Naming a competitor's incident is awkward; declining to name it while using it as evidence is a weaker form of the same claim.
What would make this checkable
The essay is a statement of intent from someone with unusual credibility, and intent is not a signal. These are:
- Does any lab actually slow down? Pachocki "expects and hopes for voluntary slowdowns to become commonplace." A slowdown that is real leaves evidence: a delayed release with a stated cyber or alignment reason, as OpenAI itself did in August and Z.ai did with GLM-5.3's weights. If no frontier release slips on safety grounds in the next quarter, "commonplace" was aspiration.
- Does OpenAI publish the CoT monitoring evaluations? The claim that monitorability is degrading is the essay's most important, and it is currently unsupported by any published measurement. A monitorability benchmark, with numbers across model generations, would turn the industry's central safety question into something outsiders can track.
- Does "unilaterally withhold" acquire a threshold? A published trigger — a capability level at which OpenAI stops, decided before the model exists rather than after — is the difference between a policy and a preference.
- Does anyone external get to audit? The essay endorses third-party auditors. Irregular's independent assessment of Astra showed what that looks like in practice and produced the first number anyone outside OpenAI could cite. More of that, with access rather than after-the-fact testing, is the concrete version of the essay's ask.
Why it still matters
It would be easy to file this as a lab talking its book — warning about danger is also a claim about how powerful your product is, and "our AI is too dangerous to scale" has been a marketing register for years. The Kurzweil citation, in particular, is a strange appeal to authority in a document otherwise careful about evidence.
But that reading does not survive the CoT passage. Announcing that your primary safety instrument is losing resolution is not a flattering claim about your technology; it is an admission that you can see less than you could a year ago, published by the person whose job it is to see. Nothing about OpenAI's commercial position is improved by it.
The essay is at its strongest where it is least quotable — the mechanics of why monitoring is getting harder — and at its weakest where it will be quoted most, in a call for collective restraint that commits its author's company to nothing in particular. Both halves are worth having. The industry's most senior researcher now says on the record that no one has earned the right to keep scaling at this speed. Whether anything follows from that is a question the next quarter of releases will answer better than any essay can.