Nobody knows, and the experts disagree sharply. What has changed is that the worry is no longer fringe. In 2023 hundreds of AI scientists and industry leaders signed a statement saying that reducing the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. In the largest survey of AI researchers, the median estimate of AI causing human extinction or a similarly severe outcome was 5%. Some leading figures put it much higher, and others think the fear is overblown.
Some of the people making these predictions helped build modern AI, which is why their numbers are worth a close look.
Could AI really end humanity?
It is possible in principle, according to a large share of the people who build these systems, but far from certain. The scenario experts worry about is not a robot uprising from the films. It is a gradual or sudden loss of human control over systems that are smarter than us, or the deliberate misuse of powerful AI by people.
On 30 May 2023, the Center for AI Safety published a one-sentence statement: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Its signatories included Turing Award winners Geoffrey Hinton and Yoshua Bengio and the heads of several leading AI companies.
Why some experts are worried
The alignment problem
The central concern is called alignment: making sure a system that may become far more capable than us pursues goals that are good for us. A very capable system with the wrong goals does not need to hate us to cause harm. It only needs to want something that conflicts with what we need, and to be better than us at getting it.
Nick Bostrom’s famous thought experiment imagines an AI told to make as many paperclips as possible. Pursued single-mindedly by something powerful enough, even a harmless-sounding goal could consume resources humans depend on.
Instrumental convergence
Researchers also point to what they call instrumental convergence: almost any goal is easier to achieve with more resources, more influence and without being switched off. So a capable system might resist shutdown or seek power, not out of malice, but because those things help with whatever it was asked to do.
Misuse and loss of control
Not every risk requires an AI with goals of its own. Experts also worry about people using AI to design biological weapons, launch cyberattacks or entrench political power, and about a slower “gradual disempowerment”, in which more and more decisions are handed to systems that humans no longer fully understand.
What the experts predict
Geoffrey Hinton
Hinton, often called the “godfather of AI” and a 2024 Nobel Prize winner in physics for his work on neural networks, left Google in 2023 so he could speak freely about the risks. In a December 2024 interview on BBC Radio 4, he estimated a 10–20% chance that AI leads to human extinction within the next three decades, and said progress had been faster than he expected.
Eliezer Yudkowsky
Yudkowsky, co-founder of the Machine Intelligence Research Institute, is the most pessimistic prominent voice. In a 2023 essay for TIME, he argued that building superhuman AI under current conditions would most likely kill everyone, and called for an indefinite worldwide halt on large training runs. His 2025 book with Nate Soares is titled If Anyone Builds It, Everyone Dies.
Nick Bostrom
Bostrom’s 2014 book Superintelligence brought the topic to a mainstream audience. In his TED talk he describes machine intelligence as “the last invention that humanity will ever need to make”, which is exactly why getting it right matters so much.
Max Tegmark
Tegmark, an MIT physicist and president of the Future of Life Institute, organised the March 2023 open letter calling for a six-month pause on training systems more powerful than GPT-4. He argues that we can keep the benefits of AI if we insist on systems that are provably safe.
Yoshua Bengio
Bengio, who shared the 2018 Turing Award with Hinton and Yann LeCun, chaired the first International AI Safety Report, published in January 2025 with input from experts nominated by around 30 countries. The report did not give a single probability, but concluded that loss of control over advanced AI is a risk that experts take seriously and that the evidence is still uncertain.
What do AI researchers as a group think?
In October 2023, AI Impacts surveyed 2,778 researchers who had published at top AI conferences. The results, later published in the Journal of Artificial Intelligence Research, were striking:
| Question | Answer |
|---|---|
| Median chance of AI causing human extinction or severe permanent disempowerment | 5% |
| Mean (average) estimate | 16% |
| Share giving at least a 10% chance of such outcomes | Between 38% and 51%, depending on how the question was asked |
| 50% chance that AI can do every task better than humans | By 2047, 13 years earlier than the 2022 survey |
That is a minority worry, but not a small one. A 5% chance of catastrophe would be considered unacceptable for almost any other technology, from aircraft to nuclear power.
AI timeline: when could superintelligence arrive?
There is no agreement. The 2023 survey gave a 50% chance of AI outperforming humans at every task by 2047. Some leaders of AI companies have suggested human-level systems could arrive within a few years, while many academic researchers think it could take decades or might require breakthroughs nobody has made yet.
For risk, the exact date matters less than whether safety research and rules can keep pace with capabilities. Many of the worried experts say they cannot, at the current rate.
The other side: why many experts are not worried
Plenty of respected researchers think extinction fears are overblown:
- Today’s systems are limited. Yann LeCun, the third 2018 Turing Award winner and long-time head of AI research at Meta, has repeatedly called the idea that AI will wipe out humanity preposterous, arguing that current systems lack the understanding and planning needed to be a threat.
- Near-term harms are real now. Researchers focused on AI ethics argue that misinformation, bias, surveillance, job loss and the concentration of power in a few companies deserve attention today, and that talk of extinction distracts from them.
- Safety can be engineered. Some argue that, like other powerful technologies, AI will be made safe through testing, regulation and gradual deployment.
Both views deserve a hearing. What is not in doubt is that AI is already changing daily life, from jobs and the economy in 2026 to how wars are fought.
Can AI risk be reduced?
Most experts on both sides agree that it can, and that the main tools are:
- Safety research: methods to understand what AI systems are doing internally and to make sure they behave as intended.
- Testing before release: independent evaluations of dangerous capabilities, such as helping to make weapons.
- Regulation: the EU AI Act, which entered into force in August 2024, is the first broad AI law. Several countries have set up AI safety institutes.
- International cooperation: agreements such as the 2023 Bletchley Declaration, signed by 28 countries and the EU, recognise the risks of the most advanced systems.
The Doomsday Clock now counts AI among the threats it measures, and the 2026 statement specifically called for international guidelines on AI.
Frequently asked questions
Could AI cause human extinction?
Many AI experts think it is possible but not likely. The median researcher in the largest survey put the chance at about 5%, while Geoffrey Hinton has said 10–20% over 30 years. Others, such as Yann LeCun, think the risk is negligible.
Who are the main experts warning about AI risk?
Geoffrey Hinton, Yoshua Bengio, Stuart Russell, Max Tegmark, Nick Bostrom and Eliezer Yudkowsky are among the best-known voices, with different levels of concern.
What is the alignment problem?
It is the challenge of making sure an AI system’s goals and behaviour match what its designers and humanity want, especially as systems become more capable than the people overseeing them.
When could superintelligent AI arrive?
Forecasts range from a few years to many decades. Surveyed researchers gave a 50% chance of AI outperforming humans at all tasks by 2047.
How does AI compare with other risks to humanity?
Oxford philosopher Toby Ord ranks unaligned AI as the single largest risk this century, at about 1 in 10. See the odds of human extinction this century.
Sources and further reading
- Center for AI Safety: Statement on AI risk (2023)
- Grace et al., Thousands of AI authors on the future of AI, Journal of Artificial Intelligence Research
- IEEE Spectrum: AI researchers on existential risk
- Nobel Prize in Physics 2024: Hopfield and Hinton
- Future of Life Institute: Pause giant AI experiments (2023)
- Eliezer Yudkowsky in TIME (2023)
- International AI Safety Report (2025)
- European Commission: The EU AI Act
Related articles
Or start from the beginning: all predictions about the end of the world.
We explain what experts and sources have said in our own words and link to the originals so you can check them. If something here is wrong or out of date, let us know.