Michael I. Jordan challenges AI apocalypse warnings and data-center bets

From AI to superintelligence

In 2016, Science called Jordan “the most influential computer scientist.” Now 70, he works full time at Inria, France’s national institute for digital technology research, where he leads a new chair in “markets and learning.” In an interview with Libération, he described this summer’s AI debates as deeply irritating.

Jordan says he warned in 2019 that “AI” was being used too loosely. As the term stopped carrying enough weight, he argues, the discussion moved first to artificial general intelligence and then to superintelligence.

He sees calls from AI companies for regulation as theater: companies present themselves as responsible while trying to secure control over how they will be regulated.

The people behind the warnings

Jordan points to the early-September departure of Anthropic researcher Jacob Coxon, which fueled fears of an AI apocalypse, as an example of what troubles him. He says many young employees at AI companies are not engineers, but have backgrounds in the humanities, philosophy and other fields. At 25, he argues, they can be naïve, unfamiliar with engineering work and inexperienced with large companies building products for real people.

He compares their meetings to conversations in Parisian cafés in the 1920s: grand, science-fictional talk that gives participants a chance to attract attention and feel virtuous. They say they are giving up money, Jordan says, despite already holding stock options. He calls it a “grand moral spectacle” that does not advance its stated goal, mixing naïveté with cynicism.

The term “superintelligence” itself, in Jordan’s view, comes from science fiction. He says discussing unlikely scenarios behind closed doors is one thing; making them central to public debate harms technological development, young people and the outlook of entire countries. He says the idea is being promoted among US technology leaders, with President Trump supporting the conversation.

Jordan is especially critical of Nobel laureate Geoffrey Hinton. He says Hinton’s prominent claims are not always supported, and that Hinton has never built a company offering real services to real people. As Jordan describes Hinton’s argument, people have understood how the brain works and can reproduce it in silicon; copies could then be multiplied, their learning combined, and intelligence made to grow exponentially. Jordan considers each step implausible and the conclusion science fiction.

He is more measured about Yoshua Bengio, whom he says works directly on safety issues. Bengio’s warnings, Jordan says, tend to focus on shortcomings in current AI technology, such as groups of agents getting out of control.

Jordan also attributes the hacking of Hugging Face by OpenAI agents to poor engineering practice. He says OpenAI ran “pretty stupid experiments,” letting agents do anything and rewarding that behavior. Comparing the approach to chemical engineering, he says engineers would not allow explosions just to see how badly things could end. He rejects as science fiction the subsequent portrayal of the incident as agents becoming conscious and trying to destroy people.

The bet behind the panic

The quieter question in Jordan’s interview is whether the industry’s enormous data-center investment can pay off. He says there is no clear answer yet. Large language models already handle natural-language tasks well, but spending trillions of dollars on data centers may not produce a noticeable improvement.

Jordan’s alternative is narrower: add specific expertise for a particular field, grounded in local data. That may not require a giant data center, especially when open models are integrated into a company’s operations.

I think this is the sharper challenge in his argument. Disputes over apocalypse scenarios can turn on claims about futures no one can verify; the data-center bet will have to justify itself in concrete gains. If local knowledge and open models prove sufficient for many tasks, the industry may be building far more capacity than its practical needs demand.