When Will AGI Happen? Every Major AGI Prediction, on the Record

Core argument: Nobody knows when artificial general intelligence arrives, and the loudest dates come from the people with the most to gain from them. What a reader can do is different: put every major forecast on one page, keep the date it was made and the source it came from, and grade it later. Five named deadlines now sit within the next 40 months. The record so far is unkind to forecasters — and genuinely useful to everyone else.

This is the sixth chapter of an ongoing series on world models. So far: why LLMs hallucinate, what separates a world model from an LLM, how the four technical routes actually work, who is building them, and what artificial general intelligence really means. The series hub holds the full map.

Last updated: October 8, 2026. This page is a living tracker, refreshed every quarter. Forecasts are never deleted when they expire — the status column records what happened.

1. The short answer

No one knows when artificial general intelligence (AGI) — a machine that matches or beats skilled humans across essentially the whole range of cognitive work — will arrive. There is no instrument that measures the distance to it. What you can do is treat every forecast as a claim with a deadline, put all of them side by side, and check back. That is what this page is for.

Five forecasts now have deadlines close enough to test: Sam Altman says the end of 2026. Dario Amodei’s window runs to early 2028. Shane Legg’s bet is 2028. Ray Kurzweil has said 2029 since 2005, and Demis Hassabis now treats 2029 as possible. Within about three years, five of the table’s rows will have come due — more graded forecasts than the field has produced since Simon’s.

One abbreviation note, as always in this series: the initials AGI also stand for adjusted gross income, the figure at the top of a US tax return, so search engines cannot always tell an intelligence question from a tax question. Full term first, abbreviation after.

2. Think of it like a GPS arrival time — and where that fails

When your navigation app estimates your arrival, it is not predicting the future. It is reporting the best current estimate from a model of traffic, and the estimate sharpens as you drive. That is the right way to read every date in this article: an estimate under constant revision, not a prophecy.

The analogy fails in two ways that matter. First, your GPS has no incentive to flatter you. A lab chief executive does: a confident date raises money, attracts researchers, and moves markets, while a missed date costs nothing. Second, a GPS estimate is calibrated on millions of completed trips to similar places. Nobody has ever completed this trip. Artificial general intelligence is a first-of-its-kind event, so every estimate — including the forecasting platforms’ — is an extrapolation with no completed runs behind it. Treat each date as a statement of what the forecaster believes — and wants.

3. Why the dates keep moving

Three forces do most of the moving.

Definitions move first. The previous chapter showed that the term has no agreed definition. The forecasters in the table below are not aiming at the same target: Altman’s bar is OpenAI’s charter definition — systems that outperform humans at most economically valuable work. Amodei refuses the word and points at “powerful AI”: smarter than a Nobel Prize winner across most relevant fields, working at 10 to 100 times human speed, deployable in millions of copies. Legg’s bar is explicitly “minimal AGI” — his term for a system as capable as a typical adult across most cognitive work, not a superintelligence. Three different bars can all be hit while the everyday meaning of the term — a machine that can do what a person does — remains unmet. When a deadline passes, watch the definition, not just the date.

Incentives move them second. A missed forecast costs the forecaster nothing; nobody is struck off a list. The upside of an aggressive date — funding, talent, attention — arrives immediately. This asymmetry does not make lab dates worthless, but it sets their exchange rate.

The absence of a base rate does the rest. In 2023, Katja Grace and colleagues surveyed 2,778 published AI researchers, the largest survey of its kind. Their median guess for machines outperforming humans in every possible task was 10 percent by 2027 and 50 percent by 2047 — which sounds steady until you learn the same survey a year earlier had put the 50 percent point 13 years later. The crowd’s own center of gravity swung more than a decade in twelve months. The forecasting platform Metaculus, which aggregates thousands of individual predictions on the question, currently puts the median at January 2031, with an uncertainty band from April 2028 to May 2037. When the disciplined guessers cannot hold still, the honest summary is: wide uncertainty, and dates that are statements of mood as much as measurement.

4. The record: every major AGI prediction, as of October 8, 2026

Two reading notes. First, the cluster: three of these bars fall due inside about two years of each other, and they are three different bars. If all three get declared met, the headlines will say AGI has arrived; what will actually have been shown is that three labs hit three self-chosen targets. Second, the direction of travel: Hassabis has moved his window earlier, Hinton cut his from 30–50 years to 5–20, and the 2023 survey pulled the crowd’s midpoint 13 years toward the present in a single year. Experts are converging on sooner — but converging from positions they have changed repeatedly.

5. What has already expired

Herbert Simon, one of the field’s founders and later a Nobel laureate, published his twenty-year prediction in 1960 and repeated it through the mid-1960s. The deadline passed quietly. Nothing marked the occasion, and the field’s confidence was not visibly reduced by it.

Elon Musk’s April 2024 version — machine smarter than any single human by the end of 2025 — expired about nine months before this page’s last update. No system is credibly regarded as smarter than every living human, and the claim was never withdrawn, graded, or conceded. It simply stopped being repeated.

Amodei’s “as early as 2026” falls due in the next few months. In fairness, he wrote himself a hedge in January 2026 — “it could also be considerably further out” — so this is a testable claim, not yet a failed one. That is exactly how a forecast should look before its deadline: dated, sourced, and falsifiable.

The lesson of the expired rows is not that forecasters are foolish. It is that expiration has no cost. Nobody publishes the correction, because no correction is owed — which is why a page that keeps the original dates and sources is the only place the grading actually happens.

6. My take

Three readings.

First: a forecast is a definition plus a date, and without the definition the date is noise. Altman’s end-of-2026, Amodei’s 2026-to-2028, and Legg’s 2028 will each be judged on a bar the forecaster chose. Expect the 2026–2028 window to produce “AGI announced” headlines even if the everyday meaning — a machine that can do a person’s job, end to end — remains unmet. The word is likely to settle into a business milestone, as argued in the previous chapter, and these deadlines are how it gets there.

Second: counting from Simon onward, every confident dated AGI forecast on record has missed. That does not prove the current dates are wrong — the field has changed more since 2020 than in the four decades before — but it does set the prior. I weight lab-chief dates as strategy statements and the survey and Metaculus as calibration instruments, with Legg’s in between: probabilistic, narrowly defined, and pinned in public for seventeen years, which is more accountability than anyone else in the table has accepted.

Third, medium confidence: by 2029 this question will have real grading data for the first time, and the pattern will be near misses with shifted definitions — some bar declared met on schedule while the everyday meaning waits. The tell to watch is a rename that arrives with the deadline. Not every rename is a dodge — Legg defined “minimal AGI” long before his date — but when a deadline approaches and the target’s name changes, the deadline was doing marketing work, not measurement work.

7. Questions people actually ask

Has AGI been achieved yet? Not by any public, falsifiable standard. Sam Altman says an internal system he would call AGI under OpenAI’s charter definition is months away; no external evaluation of it exists. The table above tracks the named deadlines as they fall due.

What is the most credible AGI prediction today? Credible should mean a track record, and nobody has one for this event — that is the point of the table. For calibration, use the instruments that are honest about uncertainty: the researcher survey (50 percent by 2047) and the Metaculus crowd (median January 2031). Legg’s 2028 bet is the most accountable of the near dates because it is probabilistic, narrowly defined, and on the record since 2009. Dates from lab chiefs are best read as strategy.

Why do AI chief executives keep saying it is just a few years away? Three reasons stack. Genuine uncertainty — even careful experts have swung their guesses by decades. Definition choice — each lab’s bar is shaped by what its own products can do. And incentives — an aggressive date raises money and talent now, at zero cost when missed. Note that the pattern is not unique to chief executives: Simon and Hinton show that academics’ dates move too, just with less money attached.

What happens if the 2026–2028 deadlines pass without AGI? Nothing visible, and that is the finding. Forecasts expire quietly, definitions shift, some labs rename the target. The useful move is to check what each forecaster did: reaffirmed, revised with reasons, or quietly relabeled. A forecaster who revises in public has told you something trustworthy; a date that slides without comment has told you something too.

Sources

1. Quote Investigator, “Quote Origin: Machines Will Be Capable, Within Twenty Years…” — traces Herbert Simon’s prediction to his 1960 book The New Science of Management Decision and its repetition through the mid-1960s.
2. Ars Technica, “Elon Musk: AI will be smarter than any human around the end of next year”, April 2024.
3. The Guardian, “AI scientist Ray Kurzweil: ‘We are going to expand intelligence a millionfold by 2045’”, June 29, 2024; publication date of The Singularity Is Nearer per Wikipedia.
4. The Decoder, “DeepMind co-founder Shane Legg sees 50 percent chance of minimal AGI by 2028”.
5. Dwarkesh Patel, interview with Shane Legg — his reasoning and the bet’s history.
6. Sam Altman, “Reflections”, January 6, 2025.
7. The Decoder, “Sam Altman says OpenAI will have AGI by the end of 2026 — if you accept his definition”, August 26, 2026.
8. Dario Amodei, “Machines of Loving Grace”, October 2024 — “powerful AI” definition and the “as early as 2026” line.
9. Dario Amodei, “The Adolescence of Technology”, January 2026 — “as little as 1–2 years away, although it could also be considerably further out.”
10. Ina Fried, “DeepMind CEO: AI agents are a ‘practice run’ for AGI”, Axios, May 26, 2026 — quotes checked against the archived copy; the live page blocks automated requests.
11. CNN, “The ‘godfather of AI’ reveals the only way humanity can survive superintelligent AI”, August 13, 2025.
12. Katja Grace et al., “Thousands of AI Authors on the Future of AI”, arXiv, 2024; published in the Journal of Artificial Intelligence Research, 2025 — 2,778 researchers; “10% by 2027, and 50% by 2047,” thirteen years earlier than the prior-year survey.
13. Metaculus, “Date of Artificial General Intelligence” — community median and band as read on October 8, 2026; this line is refreshed quarterly.
14. IRS, “Definition of adjusted gross income” — the abbreviation collision noted above.

Next in this series: what AI will actually look like in the next five to ten years — built around the forces that will shape that future. The forecasters above make claims about that future; the next chapter looks at what would have to be true for any of them to land.

More from this series: the hub — what is a world model? · why LLMs hallucinate · world model vs LLM · how do world models work? · who is building world models? · what is artificial general intelligence, really?