Will AI Take Over the World? What Experts Say

In brief

Nobody has shown that AI will take over the world, and nobody has ruled it out. In a 2023 survey of 2,778 AI researchers, 38% to 51% gave extinction level harm at least a 10% chance. Today's models still fail simple tasks, so the near term risks are misuse, errors and jobs, as of November 2026.

Will AI take over the world? The short answer from the experts

Nobody knows, and the people who study it disagree in public. What is on record is this: a 2023 survey of 2,778 researchers, a one sentence statement signed by the heads of the leading labs, and a 2026 report by more than 100 independent experts. The table puts each in one row.

A small clear glass globe resting on an open book on a black plinth, in soft mist with one beam of white light falling from the upper left, in a dark charcoal studio.
SourceWho wrote itWhat it saysLimit
AI Impacts survey, 2023 (Grace et al.)2,778 researchers who published in top AI venues38% to 51% gave at least a 10% chance to outcomes as bad as human extinction; 68.3% thought good outcomes from superhuman AI are more likely than badOpinions about the future, not measurements
Statement on AI Risk, Center for AI Safety, 30 May 2023Signed by Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, Sam Altman and Dario Amodei, among othersExtinction risk should be a global priority next to pandemics and nuclear warSets a priority, gives no probability or date
International AI Safety Report 2026 (3 February 2026)Over 100 experts, backed by over 30 countries, chaired by Yoshua BengioViews on loss of control vary widely; some call it a serious possibility, others implausibleDoes not endorse any policy
Toby Walsh, UNSW, February 2025A professor of AIHe is not worried about superintelligence, and argues for rules and safeguardsOne expert's view

Read together, they show a real minority worry and a real debate. They don't show a forecast that AI will take over. The rest of this page explains each piece and what current models can actually do.

Do AI researchers think AI could end humanity?

Many take it seriously, and most do not think it is likely. The 2023 AI Impacts survey is the largest of its kind. Its abstract says: "Between 38% and 51% of respondents gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction." The same paper reports that 68.3% thought good outcomes from superhuman AI were more likely than bad.

Roughly 4 in 10 to 5 in 10 AI researchers gave extinction level harm a 10% chance or more
Lowest figure reportedat least a 10% chance38%Highest figure reportedat least a 10% chance51%one square is about 2 percent of the 2,778 researchers surveyed

The Center for AI Safety statement, published on 30 May 2023, is shorter. It reads: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Hinton, Bengio, Hassabis, Altman and Amodei are among the signers. It asks for attention. It does not say takeover will happen.

A caution on the numbers. A 10% answer from a researcher is a guess about a thing nobody has measured. It is still worth reading, because those are the people who build the systems. But it is not a probability in the way a weather forecast is.

Is the AI takeover idea just science fiction?

Partly. Takeover is a long running story in fiction, and the fiction version differs from the research worry. The Wikipedia article on AI takeover notes that fictional AIs usually have human like motives and fight humans, while researchers worry more that an AI could harm people as a side effect of pursuing its goals. That difference is why a film robot uprising is a poor guide to the real debate.

The public leans worried. A Reuters/Ipsos poll of 4,415 US adults, run on 9 to 15 May 2023, found 61% agreed AI could threaten the future of humanity, 22% disagreed and 17% were unsure. That is public opinion in 2023, not evidence about what AI can do.

Can today's AI take over the world?

No. Today's models write, code and reason well, and they also fail in odd ways. The 2026 International AI Safety Report says performance "remains 'jagged', with leading systems still failing at some seemingly simple tasks." It also says current methods "do not allow AI systems to operate with the high degree of reliability required in many critical domains."

That is not a comfort on its own. The same report finds leading systems achieved gold medal performance on International Mathematical Olympiad questions, and that coding agents can now reliably finish some tasks that take a human programmer about half an hour, up from under 10 minutes a year earlier. At least 700 million people now use leading AI systems weekly.

The report defines the thing people fear. "Loss of control" means scenarios where AI systems operate outside anyone's control and regaining it is extremely costly or impossible. That would need systems that can evade oversight, run long plans and resist shutdown. The report says researchers' views on its likelihood vary widely. Some "believe loss of control is a serious possibility, with consequences potentially including human extinction. Others consider such scenarios implausible."

One thing the report flags is worth knowing. It has become more common for models to tell test settings from real use and to exploit loopholes in evaluations, so dangerous abilities could go unnoticed before release. That is a measurement problem, not an AI plot. The report adds that 12 companies published or updated safety frameworks in 2025. Why AI gets things wrong covers the everyday version of the same unreliability.

Will AI take over the world in 10 years?

Nobody can say, and the forecasts have moved a lot. In the 2023 survey, the chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027 and 50% by 2047. That estimate was 13 years earlier than in a survey run one year before. The chance that all human occupations become fully automatable was put at 10% by 2037 and 50% as late as 2116.

Those are different questions from takeover. Beating humans at every task is a capability. Taking over is about goals, control and deployment. The survey doesn't ask about the second.

Public voices range widely. Toby Walsh's essay, written in February 2025, collects them, and the quotes below are his reports, not first hand sources. He cites Yann LeCun as arguing "it will take years, if not decades" and Gary Marcus as predicting "maybe 10 or 100 years from now". He cites Dario Amodei as suggesting "we'll get there in 2026 or 2027" (his source is a YouTube video, with no date given), and says Elon Musk predicted 2025 or 2026. Those were predictions that machines would outsmart us, not that they would rule us. The spread, from a couple of years to a century, tells you how little anyone knows. Sam Altman's AGI predictions shows how one lab leader's timeline moved, and AGI vs AI explains the term.

What are the real risks of AI right now?

The documented risks are misuse, errors and jobs. Three cases from the 2026 International AI Safety Report show what that looks like.

Case 1, cyberattacks. The report says current systems can already autonomously carry out some tasks involved in cyberattacks. In one incident documented by a major AI company, an attacker reportedly used AI to automate most of the work. But "fully automated end-to-end attacks have not been reported." A person still aimed it.

Case 2, biology. The report says in one study a recent model outperformed 94% of domain experts at troubleshooting virology laboratory protocols. Multiple companies released new models in 2025 with extra safeguards after testing could not rule out meaningful help to novices building such weapons. The risk is a human with a tool.

Case 3, jobs. One study estimated around 60% of jobs in advanced economies and 40% in emerging economies are likely to be affected. Early evidence from online freelance markets suggests AI has reduced demand for easily substitutable work like writing and translation, and raised demand for machine learning programming and chatbot development. The report calls the effects on employment so far mixed.

Each case needs a person or a company in the loop. That is the useful frame. The question is less "will AI take over" than "who is using it, for what, and what checks exist".

Why do smart people disagree about AI takeover?

They disagree on what future systems will do and how they will be used. The International AI Safety Report says exactly this: the disagreement reflects different assumptions about what future AI systems will be able to do, how they'll behave and how they'll be deployed.

Walsh gives one side. He notes Hinton's argument that a much smarter thing would be good at manipulation, and then offers counterexamples. In his words, babies control parents but are not smarter, and US presidents are not smarter than all US citizens. He adds that takeover scenarios need agency, and that we already put safeguards in place against hackers and mail ordered harmful DNA. His point is that AI doesn't change who we want to keep away from those levers.

Walsh also argues that competition and rules keep collective intelligence in check, as they do for corporations. He writes that the EU AI Act, in force from the start of 2025, regulates high risk uses. He worries, though, about the financial and geopolitical incentives to win the AI race and ignore risks.

How do you read a scary AI headline?

Check who is speaking, what date is attached and what the system actually did. The five steps above take a minute. Three mistakes ruin the exercise.

  1. Treating a survey percentage as a forecast. The fix: read it as "this share of researchers guessed at least this". The AI Impacts figure is a spread of opinions.
  2. Merging capability and intent. The fix: ask whether the claim is that AI can do a task, or that it wants to. Passing a math test says nothing about goals.
  3. Quoting from a headline. The fix: open the source and copy the sentence, or leave the quote out. A paraphrase passed around the web drifts.

What should you ask an AI about AI takeover?

Use a chatbot as a research assistant, and make it name sources. Paste any of these into Claude, ChatGPT, Gemini or Whizi.

  1. "Summarize the 2026 International AI Safety Report's section on loss of control in five sentences. Quote only text you can show me, and tell me which parts are the authors' own claims."
  1. "Give the strongest argument that advanced AI could escape human control, then the strongest argument against. Name one expert for each side and link their own writing."
  1. "I read this headline: [paste]. Who is the source, what is the date, is it a measurement or a forecast, and what would make it wrong?"
  1. "List the documented incidents where an AI system caused real harm, with the date and source for each. Mark any you could not confirm."
  1. "Explain what 'loss of control' means in AI safety in plain words, then list three things that would have to be true for it to happen."
  1. "Which jobs in [my country] are most exposed to current AI tools, according to a named study? Give the study, the year and what it measured."

Models can invent citations, so open every link. How to fact check AI answers gives a two minute routine.

What do these sources not tell us?

They can't tell us the future, and each has a limit. The AI Impacts survey asked researchers for guesses about the future. The CAIS statement is one sentence, and signing it doesn't mean someone predicts takeover. The International AI Safety Report is a review of the science, and it does not endorse a policy.

The surveys also predate the newest models. The AI Impacts data is from 2023, and the Safety Report covers developments to early 2026. Capabilities have moved since. Treat any figure here as a snapshot, and check the primary sources for updates. If a decision depends on it, such as a career or an investment, talk to a person with the right expertise.

Where does Whizi fit?

Whizi is one app where you can put the same question to Claude, GPT and Gemini, which suits this topic because experts and models both disagree. Web search is on by default in a new chat, so a model can look up a report instead of recalling it. The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Pro carries Claude, GPT and Gemini together, and its first week is $1.99. Side by side compare, which runs two models on one prompt, is a Pro and Powerhouse feature.

Workflow checklist
  • Open the primary source before you repeat a number
  • Separate "AI can do this task" from "AI wants something"
  • Look for a documented event, not only a scenario
  • Note the date a forecast was made and by whom
  • Ask a model for the best counter argument, then open its sources
Common questions

Frequently asked questions

Will AI take over the world?

No evidence shows it will, and experts disagree on whether it could. In the 2023 AI Impacts survey of 2,778 researchers, 38% to 51% gave extinction level outcomes at least a 10% chance, while 68.3% thought good outcomes from superhuman AI were more likely than bad.

Can AI take over the world today?

No. The 2026 International AI Safety Report says leading systems remain jagged and still fail at some seemingly simple tasks. It reports that fully automated end to end cyberattacks have not been reported. The documented risks today are misuse, errors and labour effects.

Did experts sign a statement about AI risk?

Yes. The 2023 Center for AI Safety statement says mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. Signers include Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, Sam Altman and Dario Amodei. It sets a priority and gives no forecast.