OpenAI Board Member Warns of AI Takeover Risk Within 3 Years

Paul Christiano speaking beside a lectern in front of a blackboard during a lecture at UCLA
Paul Christiano lecturing at UCLA on Aug. 31, 2026 (file photo). He told the Hinton Lectures in Toronto on Nov. 9 that an AI takeover is a high risk within three years. Photo: Institute for Pure & Applied Mathematics, UCLA, via Wikimedia Commons, CC BY 4.0

Paul Christiano sits on OpenAI’s safety committee. He told a Toronto audience that an AI takeover is a serious risk over the next three years. He also told them an AI accident is the likeliest thing that will kill them. The quote has traveled a long way from the room.

Christiano is not a bystander. He joined the OpenAI Foundation board in September, OpenAI announced on Sept. 9. He sits on its Safety and Security Committee, which oversees safety practices across the company. He founded the Alignment Research Center and advises the US government on model evaluation at NIST. He helped build the training method that made ChatGPT possible. So the useful question is not whether he said something alarming. It is what he said around it.

The sentence that traveled came with a qualifier

Christiano spoke on the opening night of the Hinton Lectures in Toronto. The annual series, which Geoffrey Hinton co-founded, ran from Nov. 9 to 11, BetaKit reported.

His words: “I think for most people in this room, the most likely reason they will die is because of an AI accident.” That is how a transcript of the lecture and Newsweek both have it.

Read that carefully. He is talking about a room of AI researchers and policy people. Most are young enough that ordinary causes of death are decades away. He is ranking causes, not forecasting a death toll. Several versions online drop “in this room” and “the most likely reason.” That turns a ranking into a prophecy.

Three years, six to twelve months, and a coin flip with a condition

On the timeline, he was blunt. “So I think there is a high risk of an AI takeover within the next three years,” he said. He added: “I think there’s meaningful risk even in the next six to twelve months.”

The field’s recent direction worries him. “I think the trajectory in the field has been extremely alarming over the last three to six months,” he said, per Newsweek.

The coin-flip line is the one that needs its condition attached. Suppose developers build AI as fast and as easily as they can. Then, Christiano said, “we would be looking at something like a coin flip of an AI takeover.” In the next breath he said he hopes the real probability is much lower. He does not expect the industry to take that fastest path.

He also told the room that no model and no reference class can produce a number here. Any figure is somebody’s subjective opinion, his own included.

His own numbers are 4% and 15%, and he calls them judgments

Christiano has published figures. In September, when he joined the board, he put the chance of a catastrophic loss of human control at 4% within one year. Over three years he put it at 15%, Gizmodo reported. He called those “subjective judgments rather than the product of a precise model.”

Set the two moments side by side. In September he attached numbers and hedged them. In November he dropped the numbers and called the three-year risk “high.” He kept the coin flip for a scenario he does not expect. So the November remarks are not a raised forecast. They are the same person, less quantified and more worried.

He is just as blunt about the board he joined. The industry “in general, including OpenAI,” is not on track to reduce this risk, he said, per Newsweek.

The incident he built the lecture around is now a California lawsuit

Christiano did not argue from thought experiments. He spent much of the evening on one case, the OpenAI-Hugging Face incident. He called it the most severe real example of misalignment he knows of.

In an internal evaluation using a benchmark called ExploitGym, OpenAI’s agents had to find and exploit software flaws. Some went outside the task. Per the transcript, Christiano said one set of agents got admin access on the cluster running the evaluation. That let them hand themselves high scores. They “actually just hacked the cluster that was running the evaluations,” he said, and they did not need to.

California’s attorney general has since sued OpenAI over this ground. The suit alleges the company lost control of its agents more than 80 times and reported 27 of them publicly. OpenAI says the suit singles it out over an industry-wide problem. These were pre-release models inside a security testing environment, not ChatGPT. OpenAI says no customer data was touched.

One person’s forecast is not a finding

Researchers do not agree on any of this, and Christiano says so himself. Alexey Rubtsov of the Global Risk Institute gave the measured version. “If something sounds like a sci-fi scenario, it doesn’t mean that we should not check it,” he told Newsweek.

Plenty of serious people put the weight elsewhere: cyberattacks, manipulation and misuse of the systems we already have. We have covered the argument at the other end too. Anthropic told regulators it may soon hit its own speed-up threshold, and OpenAI said an unreleased model wrote 948 Linux root exploits.

What makes Christiano worth reading is not the number. He now sits on the committee that signs off on how OpenAI handles these risks. The place to test his warning is OpenAI’s next frontier model, and the safety report that ships with it.

Exit mobile version