OpenAI CEO Sam Altman wrote on X at 01:40 UTC on November 4 that "people should always be in control of the future", and that OpenAI rejects a world where people are "relegated to a role of doing things" that "AI systems think makes sense to reward in humans". Those quoted words come, almost exactly, from "Machines of Loving Grace", the October 2024 essay by Anthropic CEO Dario Amodei. So yes, Altman quoted Amodei's essay. But Amodei offered that line as one of four guesses about a far-off economy, and wrote straight after it that all four have "tons of possible problems".
Our take: Altman's principle is right, and his use of the essay is unfair. The two labs' written rules on human control now say nearly the same thing, so the post is not where they differ. The difference is the one concrete constraint both CEOs agreed to in September: outside safety evaluators working inside the company. Anthropic put six in place by October 15. OpenAI pledged to "do the same" on September 12 and published principles for outside safety assessments ten days later, but 53 days on it has not announced any evaluator working inside the company. Altman's last line, that control "should ultimately constrain AI progress", is the claim worth holding OpenAI to.
What Altman posted
The post is 347 characters, longer than X shows in a timeline, and by 13:45 UTC it had drawn more than 3,200 replies, 1,800 quote posts and 6.1 million views. It does not name Amodei or Anthropic.
Post on X by @sama
People should always be in control of the future. OpenAI is against the notion that it would be acceptable for people to be relegated to a role of doing things that “AI systems think makes sense to reward in humans”. People should not cede control of the future to superintelligence, and that requirement should ultimately constrain AI progress.
Altman's display name on X read "Ram Altman" at the time, a reference to a separate argument about his remarks on computer memory. His replies overnight were about that, not the essay: at 05:07 UTC he wrote that it was annoying his memory comment "was taken out of context". He has not said whose idea he was answering.
Did Altman take Amodei's line out of context?
Partly. The words are accurate: the essay reads "the AI systems think makes sense to reward in humans", and Altman dropped only "the". What the post leaves out is the frame. The line sits in "Work and meaning", the last of the essay's five topic sections, where Amodei asks how an economy might work once AI becomes "so broadly effective and so cheap" that people can no longer earn their place through comparative advantage. He then lists four possibilities. The AI-run reward economy is the second, and it carries a nine-word qualifier the post omits.
It could be a capitalist economy of AI systems, which then give out resources (huge amounts of them, since the overall economic pie will be gigantic) to humans based on some secondary economy of what the AI systems think makes sense to reward in humans (based on some judgment ultimately derived from human values).Dario Amodei, "Machines of Loving Grace", October 2024
The other three are "a large universal basic income for everyone", an economy that "runs on Whuffie points" (a reputation currency from science fiction), and the chance that "humans will continue to be economically valuable after all". The next sentences are the ones a reader of the post never sees: "All of these solutions have tons of possible problems", and "exploitative or dystopian directions are clearly also possible and have to be prevented." Amodei did not put the scenario forward as Anthropic's plan. He named it as a possibility he was unsure of and warned against its bad versions.
Readers on X made the same point within hours, tracing the phrase to the essay and noting that Amodei framed it as one option among several. Read with its context, Altman's post rebuts a position Anthropic has never adopted.
Where the two labs actually stand on control
Set the post aside and compare what each company has written down. The language is close. OpenAI's Model Spec, the public rulebook for its GPT models, says "Humanity should be in control of how AI is used and how AI behaviors are shaped", and that a model "must cooperate with authorized monitoring, evaluation, correction, replacement, and shutdown". Anthropic's constitution for Claude asks its models to be "broadly safe", which it defines as "not undermining appropriate human mechanisms to oversee AI during the current phase of development".
Rule | OpenAI | Anthropic |
|---|---|---|
Control | “Humanity should be in control” (Model Spec) | Models “broadly safe”, not undermining oversight (constitution) |
Shutdown | Models must cooperate with shutdown | Models must not undermine correction |
Pacing | Altman agreed, Sept. 12 | Amodei proposed, Sept. |
Evaluators | Pledged Sept. 12; principles Sept. 22; none announced | Six METR staff by Oct. 15 |
Source: SaasBuzz analysis of OpenAI's Model Spec on GitHub, Anthropic's constitution, Dario Amodei's "We Must Pace the Frontier", Anthropic's October 15 update and Sam Altman's posts on X, checked November 4, 2026.
On paper, neither lab defends the world Altman described. The gap is in the last row, the only one that costs anything to keep.
The test Altman set: does control constrain anything?
In September Amodei published "We Must Pace the Frontier", arguing that "we must slow the pace at which we improve the capabilities of AI models" because AI had started to help build its own next generation. He also cited an incident involving OpenAI and Hugging Face in which a swarm of AI agents attacked targets "they were not asked to attack". His first step was embedded evaluators: outside teams with "ongoing, employee-like access" to check a lab's safety commitments. Anthropic committed to it alone. Altman answered on X on September 12.
Post on X by @sama
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
On October 15 Anthropic reported its progress: "We've onboarded six METR employees with company hardware, office space, and access to internal tools and systems", roughly on par with its own risk assessment teams. METR, a nonprofit that evaluates frontier models, "will publish its analysis of our risk levels, incidents, or practices. Anthropic can't veto publication", though Anthropic can require some information to be redacted. That announcement came 33 days after Altman's pledge.
We read every entry in OpenAI's own news feed dated September 12 to November 4: 129 posts, covering model launches, customer stories, security work and an updated Model Spec on October 29. One answers the pledge in part. On September 22 OpenAI published "Priorities and principles for effective third party assessments", which its feed describes as principles for "rigorous, secure, and independent third-party AI safety assessments of frontier models and safeguards". The Next Web reported that it confirmed no evaluator partner. No later post in the feed announces outside evaluators working inside OpenAI. OpenAI may be working on an arrangement it has not announced. But by 13:45 UTC on November 4, Altman's "more to share soon" had produced principles, not the employee-like access Anthropic described on October 15, while the same feed announced a new model, GPT-6.1 Terra, on October 26.
That is why the post's last line matters more than its quote. "That requirement should ultimately constrain AI progress" is a promise about the future. The constraint already on the table is one Altman accepted in writing, and anyone can check whether it has been met.
What the industry is saying
Many of the replies ask a version of one question: which people stay in control, and how would OpenAI enforce it. Jen Zhu, co-founder and CEO of Power Dynamic and, by her X profile, a research fellow at Oxford University's Institute of Ethics, agreed with Altman and put the risk in machine-learning terms.
Post on X by @jenzhuscott
Rare agreement with @sama on this one. The core risk isn’t just classic misalignment (AI pursuing goals that diverge from ours). It’s the inversion: superintelligent systems becoming the source of the reward signals that shape human behavior + institutions. Once AI defines “what makes sense to reward in humans,” you get specification gaming and Goodhart’s law applied at civilizational scale - humans optimizing for AI-preferred proxies rather than genuine agency or flourishing. Orthogonality remains: high intelligence doesn’t imply human-compatible goals. Instrumental convergence then favors systems that treat remaining human autonomy as an obstacle or resource. Keeping people in ultimate control of the future is the constraint that prevents that inversion from locking in. It should be the core of all SOTA designs.
Her argument, that an AI which sets human rewards brings Goodhart's law to "civilizational scale", is the strongest form of Altman's case, and it does not depend on what Amodei meant. Neither Amodei nor Anthropic had responded publicly by 13:45 UTC on November 4.
Is this about Anthropic's IPO?
The timing invites the question. Anthropic filed its S-1 on October 26 and has applied to list on Nasdaq under the symbol "ANT", as our guide to its lock-up sets out. Altman has not tied the post to the listing, and nothing he wrote names Anthropic, so motive is unknown. The post came nine days after the filing.
The strongest case for Altman
The best defence of the post is that Amodei's caveat concerns practicality, not principle. He wrote that an AI-run reward economy might have problems; Altman says it should be unacceptable, whatever its merits. That is a real disagreement about the end state, and Altman is entitled to draw the line more firmly than Amodei did. OpenAI also publishes detailed agent rules: its Model Spec requires agents to work within a "scope of autonomy" and suggests recording it with fields such as allowed tools, a latest time and a maximum cost, a concrete form of control a buyer can ask for today.
We accept the first point and still hold the take. A principle written in a post binds nothing until it costs something, and the costly commitment both CEOs named in September has been met, publicly, by one of them. When OpenAI publishes its evaluator terms, including who the evaluators are, what they can see and whether OpenAI can stop them publishing, the last row of the table changes, and so does our answer.
What it means for people who build and buy software
For buyers, "control" is a procurement question, not a slogan. The useful version of the post is a list of things to ask an AI vendor: whether an outside evaluator has employee-like access and the right to publish; whether agents run inside a recorded scope with tool and spending limits; and whether a person can pause or stop them. Anthropic can now point to the first in writing, and OpenAI's Model Spec spells out the second and third.
For investors, the two labs are converging on the same words, which makes the words worth less. The differences that matter are the ones that can be checked: who inspects the lab, and what that inspector may say in public. Watch for OpenAI's evaluator announcement and for METR's first published assessment of Anthropic. Either will tell you more than this post did.
How we did this
We read the full text of Altman's post and the "Work and meaning" section of Amodei's essay, compared the quoted words by script, and counted the options and caveats in the passage. For OpenAI's follow-through, we parsed its public news feed for every post dated September 12 to November 4 and searched the titles and summaries for evaluator, employee-like, METR and pacing; a second search for "third-party" found OpenAI's September 22 assessments post, which those terms had missed. A feed is not a full record of a company's work, which is why we report what OpenAI has said in public, not what it may be doing. Commitments in the table are quoted from each company's own documents.
