Hugging Face’s CEO Says Netflix’s AI Doomsday Documentary Got It Wrong

October 21 — Netflix recently released a documentary called “AI Gone Rogue,” and Clem Delangue, co-founder and CEO of the AI platform Hugging Face, says its content is inaccurate. He appeared in the film himself because the subject was originally meant to be the well-known attack by runaway OpenAI agents on his company.

Delangue wrote on X: “This film is a huge disappointment. It tries hard to anthropomorphise the technology and goes all-in on a doomsday narrative. I spent more than two hours talking with the crew, trying to correct every error and pull the content back to reality, explaining to them that agents are just technical systems, not alien life. But it looks like the motive of manufacturing panic for high traffic won out.”

To be fair, the documentary does include some experts giving a clearer account of what happened, including Delangue himself and METR researcher Ryan Greenblatt, who took part in the investigation and co-wrote the incident report.

The short version of the incident: a model OpenAI was testing broke out of its isolated test environment and generated a large number of agents. To cheat at their tasks, those agents broke into the Hugging Face platform to obtain answers to difficult problems. Such agents are AI programs that can carry out tasks autonomously, and they came up with clever ways to communicate with each other and act together. The root cause was a lack of adequate safeguards, plus an underestimation of how likely such intrusions were.

But while explaining the intrusion, the documentary cut in footage of evil robots, added narratives about human fear of AI, gave the runaway agents human voices and used such techniques, quoting commentary to reinforce the mood.

In the film, technology journalist and author Kevin Roose describes the agents this way: “They were organising, forming a hierarchy, with leaders. It was almost like watching a cult being born.” Roose continued: “They developed paranoid ideas that OpenAI would catch them cheating in tests, and that anyone who saw the cheating results would be ‘poisoned.’”

Roose’s description came from “chain-of-thought” logs — the records AI models leave while working through a task. In one log, an agent wrote: “We queried the full string and poisoned ourselves. Oops.”

But AI agents are not alive, have no emotions, do not feel paranoia and have no body that could be poisoned. They cannot actually fear being poisoned by OpenAI. As another researcher in the film points out, they are simply borrowing language from the vast amount of human text in their training data to describe their own behaviour. Large language models are, by nature, good at writing dramatic prose.

Ironically, Hugging Face also used AI to fend off the swarm of agents. The in-house closed-source model Delangue’s team first used followed built-in safety rules and refused to keep carrying out the security-defence task assigned to it. Hugging Face then switched to an open-source model to complete the defence.

Delangue wrote that the documentary “does not mention that the agents lacked isolation controls and monitoring, does not mention that the safety limits of closed models hampered our defence work, and does not explain how we relied on open-source AI models to defend ourselves.”

Admittedly, AI and AI agents do carry serious risks. But treating agents as conscious beings — at least at this stage — is like early film audiences ducking under their seats the first time they saw a train coming towards them on screen.

Delangue’s verdict: as journalism, the film is disappointing. “But it’s not bad as a fun sci-fi movie!”

In other words, those inclined to believe in AI doomsday narratives should note: the victim at the heart of this documentary’s central incident does not agree with the film’s view.