Machine learning pioneer Michael I. Jordan argues that public debate about existential AI threats has become detached from the engineering problems that need attention. In an interview with Libération, he called the summer's AI discourse “totally frustrating” and said superintelligence scenarios belong largely to science fiction.
Jordan, professor emeritus at UC Berkeley, now works full-time at Inria in Paris, where he heads the “markets and learning” chair. His criticism spans the language used to describe AI, public warnings by industry figures and the scale of investment in data centers.
Jordan says the debate has escalated beyond AI
Jordan says he warned in 2019 about reckless overuse of the term AI. In his view, the discussion has since moved from AI to AGI and then to “superintelligence,” a progression he believes has helped create public hysteria.
He does not argue that people should never discuss far-fetched risks. Rather, he objects to those scenarios becoming the dominant public conversation. He says that emphasis can harm technological development, affect young people and shape countries’ outlooks.
That distinction matters to Jordan: speculation can be explored as science fiction, but it should not stand in for evidence about how current systems work or fail. He says debate about present-day shortcomings is more useful than treating speculative futures as established outcomes.
He questions companies’ calls for regulation
Jordan describes AI companies’ demands for regulation as “theater.” He argues that such appeals can present companies as responsible while leaving them seeking substantial control over the rules that govern them.
He also criticized the public attention around the resignation of young Anthropic researcher Jacob Coxon. The resignation, which the article says prompted a wave of collective concern about an “AIpocalypse” scenario in early September, is an example of the kind of spectacle Jordan distrusts.
Jordan argues that some young employees are drawn to broad, science-fiction discussions without experience developing real products for real people. He says public resignations can make their authors appear virtuous and important, while claims of sacrificing money may overlook stock options they already hold. He characterizes the pattern as a “grand moralizing performance,” combining what he sees as naivety with cynicism.
Warnings differ, Jordan says, and engineering still matters
Jordan is particularly dismissive of Geoffrey Hinton’s account of how superintelligence could emerge. As Jordan describes it, the argument assumes that researchers understand the brain well enough to reproduce it in silicon, duplicate it, and combine what copies learn to create exponential growth. He says none of those steps is credible.
He is more receptive to Yoshua Bengio’s warnings when they address weaknesses in current AI, such as agent swarms spiraling out of control. The contrast in Jordan’s comments is between immediate technical concerns and claims about a distant, vastly more capable intelligence.
He applies the same practical lens to the hack of Hugging Face by OpenAI agents. Jordan calls it bad engineering, saying that in chemical engineering, people would not allow explosions simply to observe how badly things could go. He criticizes letting agents act without adequate constraints, and says turning the incident into a story about self-aware systems trying to destroy humans is a science-fiction interpretation.
Data center spending is a bet, not a guarantee
Jordan also questions the scale of data center investment. He calls it an enormous bet whose payoff is unknown, and says greater spending will not necessarily produce a comparable leap in what language models can do.
For many uses, he suggests, organizations may need specialized knowledge tied to a particular field or local context. He argues that this can be added to open-weight models integrated into a company, without requiring a gigantic data center.
His broader point is that AI progress should be judged through concrete capabilities, careful engineering and useful applications. In Jordan’s view, focusing public attention on existential scenarios risks crowding out those questions, along with practical discussion of current system limits and the investments being made to expand them.