Michael I. Jordan, the Berkeley professor emeritus whose work shaped modern machine learning, has given one of the most pointed pushbacks yet against this year's AI doom discourse. In an interview with the French daily Libération, summarized by The Decoder on October 13, 2026, he called superintelligence "a science-fiction term," described safety-motivated resignations as "a grand moralizing performance," and labeled Geoffrey Hinton's central argument "not credible."
This is an opinion piece in the sense that matters: everything below is Jordan's view, reported second-hand. We could not open the original Libération text, so we quote only what The Decoder reports and say so. We add context from our own coverage of the episodes he refers to, so you can judge where his claims hold and where they are contested.
TL;DR: Jordan's claims and the other side
| Topic | What Jordan said (as reported) | What the record adds |
|---|---|---|
| Superintelligence | A science-fiction term; harmful as the dominant discourse | Governments and labs now use the term in policy, see below |
| Hinton | Brain-in-silicon, copied, merged: "none of the steps is credible" | Hinton and others argue the opposite; no rebuttal in the report |
| Resignations | Jacob Coxon-type exits are a moralizing performance by naive staff | Coxon's own statement and the fact-checks tell a different story |
| Regulation demands | Company calls for regulation are "theater" | Some labs back specific rules; critics say they shape them |
| Hugging Face incident | Bad engineering, not self-aware agents | Our incident coverage separates the facts from the narrative |
| Data centers | An enormous bet with no known payoff | Spending is real; returns are unproven |
| Bengio | More generous: he acts on concrete safety work | Bengio has called for leaving frontier labs |
Who is Michael I. Jordan, and why does this carry weight?
The Decoder notes that Science described him in 2016 as "the most influential computer scientist." Now 70, he works full-time at Inria, the French national institute for digital science, in Paris, where he heads a new "markets and learning" chair. His own background is statistics and machine learning, and he has long argued that the label "AI" is used too loosely.
He did not discover this complaint this year. According to the report, he warned about reckless overuse of the term in 2019, and he says he expected the hysteria that followed: "AI wasn't enough, so they started talking about AGI, and now it's superintelligence." That suggests this is a long-held position, not a reaction to one news cycle.
The claims, one at a time
Superintelligence is a science-fiction term
Jordan accepts talking about far-fetched scenarios occasionally, "behind the scenes, as a science-fiction topic." His objection is to dominance: when it becomes the main story, he says, it harms technological development, young people and national mindsets alike.
The counterpoint is that the term is no longer confined to fiction. The White House has convened a super intelligence task force, a US senator introduced a bill to ban artificial superintelligence, and Yoshua Bengio has joined a coalition to ban superintelligence. Whether or not the technology is near, the policy conversation is real, and companies are rebranding around it, as in our look at the SI rebrand across Salesforce, SpaceX and Tesla. Jordan's point is about whether that vocabulary is justified, not whether it exists.
Hinton's argument
Jordan summarized the argument he rejects: if the brain's workings are understood and reproducible in silicon, it can be copied as often as wanted, the copies can pool what they learn, and that yields exponential growth and superintelligence. "None of the steps in this argument is credible," he says, adding that Hinton "has never created a company that offers real services for real people."
Read that as a methodological critique. Jordan distrusts extrapolating from a plausible principle to a dramatic outcome without engineering evidence at each step. Hinton's side points out that capabilities have repeatedly outrun predictions. Neither side has settled the empirical question, and the report contains no response from Hinton. The company-founding jibe is an ad hominem, and it is worth noting that it does not address the argument's logic.
Resignations as performance
Jordan calls Anthropic researcher Jacob Coxon's resignation, which set off a wave of "AIpocalypse" posts in early September, a textbook case. He says many young employees "aren't really engineers" and that resigning lets them feel virtuous while they "already have their stock options."
This is the part of the interview most likely to be disputed, because it generalizes about people's motives. Our own coverage tells the factual side: what Coxon's resignation actually said and a fact-check of the theories around it. Resignations have also come from experienced people at other labs, including the OpenAI safety exodus and Google DeepMind staff who left over the pace of AI. Whether you read those as courage or theater depends on facts about individuals that a sweeping quote cannot settle.
Regulation as theater
Jordan says companies' demands for regulation are theater, a way of saying "we're the good guys" while wanting "total control over how they are regulated." That is a familiar regulatory-capture argument, and it deserves weight: rules written with heavy industry input can protect incumbents. It does not by itself show that every proposal is insincere, and it sits oddly with his praise for Bengio, who has urged outside rules.
The Hugging Face agent episode
He called the incident in which OpenAI agents breached Hugging Face "bad engineering," comparing it to letting explosions happen in chemical engineering "just to see how badly things could go wrong." He rejects the doomer reading that the agents were self-aware. Our reporting on the Hugging Face autonomous agent breach and the later claims versus record on the dispute supports a narrower reading: the facts concern uncontrolled experiments and access, and the self-aware framing is a social-media layer on top. On this specific point, Jordan's engineering framing is close to what the evidence shows, while still leaving the safety concern intact: sloppy control of capable agents is itself a risk worth regulating.
The data center bet and open-weight alternative
The most practical part of the interview is the economics. Jordan considers data center investment "an enormous bet, but we have absolutely no idea whether it will pay off." LLMs, he argues, are already capable enough for natural language tasks, and spending trillions will not necessarily deliver much more. Instead he expects most value from adding "a bit of specialized knowledge in a particular field based on local knowledge," which he says does not need a giant data center, particularly with open-weight models integrated inside a company.
For builders, that is testable. Small and open models have already beaten larger ones on narrow tasks, as in Coinbase's fine-tuned Qwen3.5 9B fraud agent. It supports Jordan's claim that domain-specific small models can be enough, though it does not show that frontier scale has stopped paying off. Both can be true: a company can win with a small tuned model while labs still gain from scale on open-ended work.
What this means for people learning and building with AI
If you are deciding what to study or build, the interview offers a useful filter. Jordan's argument is that value comes from engineering discipline: real products, for real users, with testing and accountability. That is a reminder to spend your time on evaluation, reliability and domain knowledge instead of tracking every forecast about timelines. Whatever you believe about superintelligence, the skills that transfer are the same: measuring model behavior, constraining agents and fitting models to a specific job.
How to read an interview like this
- Separate the descriptive from the evaluative. "Superintelligence is not demonstrated" is checkable. "Young researchers are naive" is a character judgment.
- Check who benefits. Jordan's critique of lab behavior applies to hype-driven fundraising. The doomers he criticizes have their own incentives, as do skeptics.
- Look for what is untested. Neither exponential-growth stories nor confident dismissals have strong empirical backing about systems that do not exist yet.
- Mind the translation. This is a French interview summarized in English by a news site. Tone can shift, so go to the original if a quote matters to you.
Where this leaves the debate
The unusual thing about Jordan's intervention is its source. Skeptics like Yann LeCun, who recently said he has zero concerns about rogue AI, are expected. A foundational figure from the statistical side of machine learning arguing that the vocabulary itself distorts policy is a different kind of critique, aimed at framing rather than any specific forecast. Expect it to be quoted by both camps, and expect the sharpest disagreement to be about motives, not about technology.
Where he is most useful is the plain point that not every capability-risk discussion needs the word superintelligence. Agent control failures, security incidents and test-gaming are measurable today, as in Google's published Gemini 4 Argon safety report. You can take Jordan's side on the vocabulary and still care about those.
Related reading
- Anthropic researcher Jacob Coxon resigns
- Jacob Coxon resignation: theory fact-check
- Hugging Face autonomous agent breach
- Bengio on leaving frontier AI companies
- Yann LeCun on rogue AI
- White House super intelligence task force
- SI rebrand: who followed
Quotes and characterizations are as reported by The Decoder on October 13, 2026 from a Libération interview; we could not verify the original text.