Will AI Take Over the World?

What Does "AI Taking Over the World" Actually Mean?

"Taking over the world" usually bundles three different claims together: AI doing more economic work than humans, AI systems making more decisions that affect people's lives, and a science-fiction scenario where AI pursues its own goals against human interests. Only the third one is what "takeover" literally means.

The first claim, AI doing more of the work, is already true and mostly boring: spreadsheets, drafts, code, customer replies. The second, AI making more decisions, is also happening in loan approvals, hiring screens, and content moderation, and it deserves real scrutiny because a biased or broken model making decisions at scale already does harm. The third claim, the one people usually mean when they ask this question, is about autonomy: a system that sets its own objectives and acts on them regardless of what a human wants. There is almost no evidence for it today, and it is the claim this piece focuses on.

What Do AI Researchers Actually Think?

The largest survey of AI researchers ever conducted found a median 5% probability that AI causes human extinction or a similarly permanent catastrophe. More than half of respondents gave at least a 5% chance to that outcome, and the average estimate, at 9%, was nearly double the median.

AI Impacts surveyed 2,778 researchers who had published at top AI venues for its 2023 Expert Survey on Progress in AI, led by Katja Grace. Asked to weigh the long-run impact of high-level machine intelligence, the median respondent put 5% odds on an "extremely bad, e.g. human extinction" outcome, the same figure as the prior year's survey. The distribution has a long tail, with 57.8% of respondents giving at least a 5% chance to that outcome. The same survey found the median aggregate forecast for a 50% chance of human-level AI moved up to 2047, thirteen years earlier than the 2022 estimate. So the researchers closest to this work take the risk seriously without treating it as likely.

Why Do Experts and Superforecasters Disagree on AI Risk?

AI domain experts put a median 3% probability on AI causing human extinction by 2100. Superforecasters, people with a documented track record of accurate predictions, put that number at 0.38%, nearly eight times lower. Months of structured debate between the two groups did not close the gap.

The Forecasting Research Institute ran the most rigorous version of this question anyone has attempted, the Existential Risk Persuasion Tournament, which put 80 domain experts and 89 superforecasters through months of research, argument, and rebuttal on AI and other long-run risks. On AI, the two groups ended further apart than on any other risk category they studied. The tournament's own researchers said the most puzzling result was that both groups barely moved, even though they were incentivized to persuade each other and had months and thousands of forecasts to do it. We read that as evidence that the question is hard and that neither group is clearly right.

Why Is Artificial Intelligence Bad, and Which Risks Are Real Right Now?

The documented risks right now are concentration, misuse, and reliability. A small number of companies control the compute and infrastructure much of the world now runs on, bad actors already use AI for fraud and cyberattacks, and models still fail in ordinary, expensive ways that have nothing to do with scheming for power.

Concentration comes first. A handful of labs and cloud providers control the compute, data, and model access that everyone else builds on, so an outage, price change, or policy shift at one company can ripple through an entire economy. Misuse is next: AI already lowers the cost of phishing, deepfakes, and automated scanning for software vulnerabilities. Reliability is the least dramatic risk and the most common. The AI Incident Database tracked 362 documented AI-related incidents in 2025, up from 233 in 2024, according to Stanford's AI Index Report, and nearly all of that growth is mundane: wrong outputs shipped as fact, biased screening tools, chatbots giving bad advice. None of it required a system with its own agenda, only a system that was wrong and that nobody caught in time.

Why Is Artificial Intelligence Good, and What Does It Actually Deliver?

AI already drafts, codes, catches errors, and handles repetitive work at a scale no equivalent team of humans could match at the same cost. AI experts rate the technology's likely impact far more positively than the public does, and the gap is largest on personal benefit. The realistic upside is productivity that compounds over time.

In Pew's 2025 survey of 1,013 U.S.-based AI experts and 5,410 members of the public, 56% of experts said AI will have a positive effect on the United States over the next 20 years, versus 17% of the public. Experts are more positive because they spend their days watching the technology do useful things the public mostly does not see, which is different from being naive about risk. We see the same pattern in our own work. The upside of AI right now is narrow and real: faster first drafts, fewer dropped tasks, and more work done per person. None of that resembles a general intelligence taking on the world's problems on its own initiative.

What Does the Public Think About AI?

Half of Americans say they are more concerned than excited about AI in daily life, up from 37% in 2021, and only 10% say they lean toward excited. AI experts feel almost the opposite: 47% are more excited than concerned, and just 15% lean toward worried.

Pew has tracked this since 2021, and over that time worry has gone up while excitement has stayed flat. More than half of Americans, 57%, rate AI's societal risks as high, against 25% who rate its benefits as high. On jobs, only 23% of Americans think AI will have a positive impact on how people do their work, per Pew's 2026 tracking. The public has clearly decided AI is a big deal. What worries it is control, jobs, and whether the people running these systems are accountable to anyone, not a robot uprising.

Can AI Become Self-Aware?

No credible research has demonstrated self-awareness or consciousness in a deployed AI system, and there is no established test that could even detect it if it existed. Current models predict text and select actions, and they have no persistent goals, no body, and no stake in their own continuation between conversations.

People often conflate two things here. A model can appear to reason, hold a consistent persona across a conversation, and describe its own "thinking" in plain language, and none of that requires consciousness, any more than a thermostat needs consciousness to "decide" to turn on the heat. Whether any information-processing system can be conscious at all is a contested question in philosophy of mind that predates AI by decades, and nothing about current transformer-based models has settled it. From running these systems daily, we can say they have no memory between sessions unless we explicitly give them one, no body, and no persistent objective that survives past the task in front of them. "Self-aware" in that question carries a science-fiction meaning, and nothing the systems in production today do fits it.

Will AI Replace Humans?

So far, AI replaces specific tasks rather than whole people. It is already displacing routine work in translation, first-draft writing, and boilerplate code, while creating new work like agent oversight and prompt design. Some jobs shrink and some grow, and in almost every job the shape of a normal workday changes.

We wrote a full breakdown of this by profession in our companion report on AI and jobs, so we will not duplicate it here. In short, the public is more pessimistic than optimistic on this question, and Pew's data backs that up, with only about a quarter of Americans expecting AI to help how they do their jobs. What we see building these systems for clients fits that caution better than either extreme: AI takes over specific, bounded tasks well, but it does not take over judgment, accountability, or the messy parts of a job that involve reading a room.

What's Our Take, Building AI Agents Every Day?

Our take is that today's AI fails like ordinary software. The failures we see in production are mundane: a scraper that misreads a date, an agent that retries the wrong step, a model that states a wrong phone number as fact. None of that resembles a system scheming for power.

That does not mean the frontier labs get a pass on scale risk. The concentration and misuse problems above are real and getting more attention: the International AI Safety Report, led by Yoshua Bengio and backed by more than 30 countries and over 100 independent experts, tracks this ground every year, and its most recent update found that the number of companies publishing formal frontier AI safety frameworks more than doubled in 2025. We think that is the right kind of response to a technology this consequential, because it is independent, international, and focused on what labs actually do rather than what they promise. Our own working rule, after shipping agents into real businesses week after week, is simpler and less dramatic than either the doom framing or the hype framing: today's systems fail like software rather than like adversaries. So the risks worth your attention right now are reliability, oversight, and who controls the infrastructure, and a system plotting against you is not one of them.