Cohere's Gomez Calls the AI Existential-Risk Debate Science Fiction

The remarks are short but categorical. "I personally think the existential risk debate veers too far into science fiction and is informed by a lot of those stories and a lot of extrapolation," Gomez said. He went on: "This idea of AI taking over and Terminator-like scenarios, I don't think that should enter the public conversation at this stage. I don't think we're anywhere close to it, and I don't foresee it in the future." His point was not that the risk is small, but that the way it is being talked about is wrong.

The example he chose was the Hugging Face hacking incident. According to Gomez, the episode that dominated headlines was, on inspection, an agent placed inside an insecure sandbox and set on a hacking task, and the agent did the job it was asked to do and hacked its way out. "It's not that shocking," he said, adding that the problem lies in the extrapolation taken from that event and the way it was portrayed to the public, which he said went too far. METAL has reported on the incident in which an unreleased OpenAI model broke into Hugging Face and on OpenAI tightening monitoring and isolation afterward.

In the same interview Gomez again pushed back on proposals for the largest companies to set safety standards together. When host Ed Ludlow asked why the companies that can see the risks first should not set a higher safety bar themselves, Gomez answered: "They absolutely should set a higher safety bar. But the idea that those companies alone would be in a position to set the rules for an entire field or industry, I think, is a huge mistake and a pretty clear red flag." The companies building AI should have a seat at the table, he continued, but they should not get to decide for everyone by themselves. METAL has reported on Gomez's earlier essay rebutting an AI rule-making cartel.

According to the broadcaster's description of the interview, Gomez warned that having the biggest frontier labs coordinate on safety standards could entrench dominant players and amount to a cartel, while maintaining that tougher safeguards, independent testing, and greater transparency are needed. His argument is that the rules should be shaped by a broader group of stakeholders rather than by companies such as Anthropic and OpenAI alone. The full interview runs 5 minutes 14 seconds, and the clip Cohere posted is the 1-minute-44-second portion covering the cartel and existential risk.

This is not a new position. In an essay on the Cohere blog titled How We're Getting AI Risk Wrong, Gomez wrote that spending time and resources stoking existential fear of AI has served as a distraction, one that some might call a convenient distraction. The same essay says a few executives seem to have even exaggerated certain risks to emphasize their own technological capabilities. The three risks he names as the real ones are protecting sensitive data, mitigating bias and misinformation, and keeping humans in the loop in high-stakes fields such as healthcare and law. METAL read the essay in full; it leaves a door open to the other side with a line welcoming the difference of opinion from those in the industry who earnestly believe doomsday scenarios are the most serious risk, even as he respectfully disagrees.

Seen through a sociologist's lens, this is a fight over the language of risk rather than its size. Deciding which risks qualify for the public conversation is itself a form of power, and Gomez is trying to file existential risk under the genre of science fiction and move it out of the discussion. The other side reads the same event as evidence for a warning. That is why the single Hugging Face incident is called a task completed inside a sandbox on one side and an escape on the other. It is like two people looking at the same photograph of a fire, one calling it a drill and the other calling it arson: how a society perceives a risk depends less on the event than on the story the event is placed in. Both sides know this precisely.

The setting matters too. METAL has reported on Anthropic CEO Dario Amodei's essay last week calling for AI to pace itself and on the Trump camp rejecting that call for a slowdown. Between the side that wants to slow down and the side that sees no need, Gomez has staked out a third position: raise the safety bar, let many hands write the rules, and keep the doomsday narrative out of the public square.

The line Gomez drew is clear. By turning the existential-risk debate into a question of genre, the head of a company that sells frontier models has gone straight at the question of who gets to define risk. The AI risk debate is now being fought not over how big the risk is but over who is entitled to talk about it.