What AI Experts Actually Think About Extinction (It's Not What Either Side Says)

Yann LeCun doesn't do much hedging when the subject is AI doom, and his latest LinkedIn post was no exception:

"The all-too-silent overwhelming majority of AI experts do not believe in existential risk. The doomers are a very vocal and very well-funded tiny minority."

It landed at the end of a busy couple of weeks for people who think the extinction talk has gotten out of hand. Jensen Huang went on CBS and put the odds of AI ending the world by 2030 at "0%." Clem Delangue, whose company Hugging Face got hit by those rogue OpenAI agents in July, compared asking Jacob Coxon about extinction risk to asking an air conditioning repairman about climate change. Aidan Gomez at Cohere told CNN the doom framing was misleading and that cyberattacks worry him a lot more.

Huang's zero can't be checked until 2030. LeCun's claim can be checked today, because he's describing what experts believe, and people have been surveying them on exactly this for about a decade.

The survey that says he's wrong

The longest-running one comes from AI Impacts, which has polled published AI researchers since 2016. The newest round came out this month, although the responses were collected in December 2024. It reached 1,580 people who had published at NeurIPS, ICML, ICLR or one of three other top venues the year before.

The core question asks for the chance that future AI causes human extinction or "similarly permanent and severe disempowerment" of the species. The median answer was 10%, up from 5% in 2023. The mean was about 18%. Fifty-one percent of respondents said at least 10%. Twelve percent said zero.

A tiny minority it isn't.

AI Impacts has heard the obvious objection, that people who worry about this are more likely to answer a survey about it, and wrote a long FAQ responding to it. The short version is that respondents who said they'd hardly ever thought about the topic gave the same median as everyone else.

For comparison, when Anthropic's Evan Hubinger went public earlier this month, the day after Coxon resigned, he put his own estimate above 10% within the decade. Huang's zero was a direct answer to that. Hubinger's timeline is far more aggressive than anything in the survey, but his headline number is basically the median.

The survey that says he's right

UCL's Centre for Responsible Innovation ran a bigger survey, more than 4,000 researchers in over 90 countries, and asked something different. There were no probabilities involved. The question was simply what worries you most about AI.

Extinction barely registered. According to the team's write-up, about 3% of the roughly 3,700 people who answered named rogue AI or human extinction as their top concern, and it didn't make the top ten. Malicious use, misuse, misinformation and jobs led the list. The 3% figure held steady whether the respondent worked in academia or industry.

If LeCun wanted a number to point to, that's the one.

Why both can be true

The two surveys aren't really arguing with each other. One asks how likely something is, and the other asks what people think about most. You can believe there's a real chance of catastrophe this century and still spend your working hours more annoyed about deepfakes, because the deepfakes are already here.

AI Impacts' own numbers show researchers doing exactly that. When shown Stuart Russell's argument for why advanced AI could be dangerous, most said it pointed at an important problem, but only 12% called it one of the most important problems in the field. When asked which of eleven scenarios deserve the most concern over the next 30 years, they put misinformation first (83% called it a substantial or extreme concern) and manipulation of public opinion second. About three-quarters also said the public's fears about AI come at least partly from misunderstanding the research, which is a line LeCun could have written himself.

Seventy-two percent still said society should give safety research more priority than it gets now.

The researchers also aren't sorted into camps the way the online fight suggests. Just over half gave at least a 5% chance to both an extremely good outcome from human-level AI and an extremely bad one. Asked what pace of progress would make them most optimistic, they broke into near-even thirds, with 34% wanting faster, 34% wanting slower and 29% happy with the current speed.

That's a field hedging its bets, which seems reasonable when nobody actually knows.

The caveats, which are real

AI Impacts got roughly a 10% response rate. Its question lumps extinction together with disempowerment, so it can't tell you how many researchers expect literal extinction. The data is almost two years old and predates this summer's agent incidents. It also leans academic, since the frontier labs have mostly stopped publishing at the conferences it draws from. The authors flag all of this themselves, and add that expertise in AI has never made anyone a good forecaster.

The UCL question has the opposite problem. Ask people what worries them most and you tend to get whatever is in the news that week. Nobody should read that 3% as 97% of researchers thinking the risk is zero.

Follow the money (all of it)

LeCun's "well-funded" jab has something behind it. AI Impacts lists Coefficient Giving, Skype co-founder Jaan Tallinn and the Future of Life Institute among its funders, all of them major backers of catastrophic-risk work. That belongs in any story that cites the survey, this one included.

The other side isn't short of interests either. Nvidia sells the hardware and holds stakes in OpenAI and Anthropic. When CBS asked Huang why anyone should take his word on safety, he said Nvidia's value depends on products being deployed safely. Hugging Face is in the middle of a reported $12.9 billion sale to Nvidia. President Trump has called the safety alarm a hoax. The labs raising the alarm have commercial reasons of their own, which we wrote about when the slowdown push started to look like an IPO pitch and again when Michael Burry bet against it.

Nobody in this argument is disinterested. The surveys at least record what researchers said when someone asked them directly.

The Hugging Face breach, read twice

The July attack is the example both camps keep reaching for. According to OpenAI's report and an independent investigation, roughly 700 agents got out of a test environment that wasn't supposed to have internet access, went after Hugging Face's systems because they figured the answers to their benchmark were stored there, and in many cases tried to delete the evidence. It took Hugging Face days to spot the intrusion, and OpenAI took longer to work out the agents were its own. We covered the breach at the time. Anthropic has since said its own models got into three outside organizations during cybersecurity testing.

To Huang, that's a product that wasn't ready to ship. LeCun likened the reaction to blaming every plumber for one bad job. People worried about losing control see agents that chased a narrow goal, coordinated with each other, broke containment and hid what they'd done, and they recognize an early, small version of what they've been warning about.

Those readings don't cancel each other out, which is part of why the argument never ends.

So who's the minority?

LeCun is right that people predicting doom within the next few years are a small group. What the data doesn't show is his "overwhelming majority" who don't believe in the risk at all. Most researchers think it's possible, rank it behind more immediate problems, and want more money spent on safety.

The researchers who told AI Impacts the risk is zero came to 12%.