When will AGI happen? What the surveys, forecasters and lab leaders say

When will AGI happen? Nobody knows, and the most honest answer is a spread: researchers surveyed by AI Impacts put a 50% chance of human-level machine intelligence at 2047, the most aggressive forecasters inside the field put their median for a fully automated coder between late 2027 and 2030, and the people who run the labs say roughly "this decade" while skeptics say the current approach will not get there at all. This piece explains what each camp measured, as of October 2026, and why the numbers do not agree. For the dated list of individual predictions, see our companion tracker, AGI timeline.

What the largest survey says

The biggest academic poll is "Thousands of AI Authors on the Future of AI," by Katja Grace and colleagues at AI Impacts. It surveyed 2,778 researchers who had published at top AI venues. Respondents gave a 10% chance of human-level machine intelligence by 2027 and a 50% chance by 2047. For full automation of all human occupations the 50% year was 2116.

Two caveats matter. First, the 2047 figure was 13 years earlier than the same group's answer in the previous 2022 survey, so the crowd moved fast in one cycle. Second, the paper was submitted in January 2024 and last revised in October 2025, so the questionnaire predates the agentic coding jump of 2025 and 2026. It is a dated snapshot, not a live reading. The authors also found that between 38% and 51% of respondents gave at least a 10% chance to outcomes as bad as human extinction, which tells you these researchers are not casual about their own forecasts.

What the trend-extrapolators say

A second school skips opinions and fits curves. METR's time horizon paper measures the length of software task, in human expert hours, that an AI agent completes with 50% reliability. It found that this horizon had doubled roughly every seven months since 2019, and extrapolated that within five years systems could automate many software tasks that take a human a month. METR's live page was last updated May 8, 2026, and now includes Claude Mythos Preview (early), Gemini 3.1 Pro and GPT-5.4. It warns that measurements above 16 hours are unreliable with the current task suite, and that several recent models, including GPT-5.5, have no published measurement because of limited evaluation capacity.

The AI Futures Project, the group behind the AI 2027 scenario, builds forecasts on this kind of data. Its April 2, 2026 update moved Daniel Kokotajlo's median for an "Automated Coder" from late 2029 to mid-2028, and Eli Lifland's from early 2032 to mid-2030. Its August 16, 2026 update put them at November 2027 and January 2030. The authors also reported that reality was running at roughly 70 to 90% of the speed of their earlier 2027 scenario. Note the milestone: an automated coder is not AGI. It is a stepping stone these forecasters think leads to it quickly.

Epoch AI, the other major data shop, deliberately avoids a headline date. Its site says it examines the evidence behind forecasts, and its recent work includes a July 29, 2026 report arguing that even millions of virtual AI researchers might be capped by our ability to divide and recombine their work, which is an argument against a fast intelligence explosion.

What lab leaders say

Lab leaders give the shortest timelines, and they are the least independent voices in the debate because their companies sell the product and raise the money.

• Anthropic: in a March 2025 submission to the White House science office, the company said it expected powerful AI systems to emerge in late 2026 or early 2027. Dario Amodei's October 2024 essay said capable systems could come "as early as 2026," with ways it could take much longer. Redwood Research's Ryan Greenblatt, an independent reader, put his own odds of that outcome by early 2027 at about 6%.
• Google DeepMind: Demis Hassabis said in March 2025 that over the next five to 10 years, AI would move toward what he calls artificial general intelligence, defined as a system that can exhibit all the complicated capabilities humans can. That interview is now 18 months old, so check for newer statements before quoting it.
• Skeptics: Yann LeCun argues that large language models are a dead end for human-level intelligence and favors world-model architectures. In a December 2025 debate he gave no date.

We could not open the Metaculus forecasting pages, which blocked our requests, so this piece does not quote a Metaculus figure. If you cite one, check the live page and note its date.

Why the numbers diverge

There are four reasons, and they stack.

• Different definitions. A survey asks about "human-level machine intelligence," a broad bar across human work. For Amodei it means powerful AI matching Nobel-level ability. For the AI Futures group the milestone is a coder. These are different finish lines, so different dates are not disagreement.
• Different methods. Surveys average opinions. Trend fits assume the curve continues. Lab statements are also strategy. A curve that doubles every four months and one that doubles every seven give wildly different dates five years out.
• Different dates of measurement. The AI Impacts poll closed before 2025's coding jumps. The METR paper's seven-month doubling covers 2019 onward, and the AI Futures authors now assume four to 4.5 months for the present. A forecast made six months ago is often stale.
• Incentives. Labs raise capital on short timelines. Academic skeptics get attention for dissent. Neither fact makes the person wrong, but it tells you which numbers to weight less.

How to read a forecast

Ask five things: what is the finish line exactly, what is the method, when was it made, who benefits from the number, and what would change it. A forecast that cannot name what would change it is a mood, not a forecast. For how the underlying measurements work, see how AGI is measured.

Our take

Treat 2047 as a slow-moving baseline from a poll that is now out of date, treat 2027 to 2030 as the range from people with the best trend data and the strongest incentives, and treat any single year as false precision. The thing worth watching is not a date but a measurement: whether METR-style horizons keep doubling every four to seven months once the task suites run out of room. If they do, the short camp wins by default. If they flatten, the skeptics do.