Nobody knows — and the disagreement itself is informative. The people closest to building these systems give answers spanning “2027” to “2047” to “not on the current path at all.” The spread is not mostly about who is smarter; it is about what counts as AGI, what evidence to trust, and how much weight to give the bottlenecks that could slow progress down.
If you need one sentence for a meeting: most AI labs and their leaders expect something they would call AGI within a few years, the broad survey of AI researchers puts the median around 2047, and a vocal minority of experts thinks the current approach will never get there. Everything else in this entry is about why those three groups disagree, and what evidence would settle it.
What “AGI” even means
AGI stands for artificial general intelligence: an AI system that can do, broadly, the intellectual work a human can do — learn new skills, reason across domains, and carry out long, open-ended tasks without being rebuilt for each one. That is the rough idea. The trouble starts when you try to make it precise.
OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Google DeepMind researchers, in a 2023 paper, proposed instead a ladder of levels based on two dimensions: how well a system performs, and how broadly. Anthropic’s Dario Amodei dislikes the term altogether and prefers “powerful AI,” which he defines concretely: a system smarter than a Nobel Prize winner across most fields, able to work autonomously on tasks lasting days or weeks, with millions of copies running in parallel.
These are not the same test. A system could pass one definition and fail another — and, as we will see, the scoreboard’s own grading problems are mostly definitional problems.
What the optimists say
The leaders of the major AI labs are the most bullish group, and they have moved earlier, not later, over the past two years.
- Dario Amodei (Anthropic) told the UN Security Council in September 2026 that a “country of geniuses in a data center” — his phrase for millions of copies of a system smarter than top human experts — could arrive within “one or two years, maybe less.” In his 2024 essay he had said powerful AI “could come as early as 2026, though there are also ways it could take much longer.”
- Sam Altman (OpenAI) has said the company would have an internal system he would call AGI by the end of 2026, and that AGI will “probably get developed during this president’s term.”
- Nathan Benaich, author of the annual State of AI Report, predicted AGI in 2027.
- Daniel Kokotajlo, author of the AI 2027 scenario, gives about a 50% chance that AI companies automate the whole AI research process by the end of 2028, and expects strong AGI within five to ten years of November 2025. His median date for full coding automation is 2029.
- Leopold Aschenbrenner, formerly of OpenAI, argues in Situational Awareness that “AGI by 2027 is strikingly plausible,” tracing trendlines in computing power, algorithmic efficiency, and the shift from chatbots to autonomous agents.
- Elon Musk gives the earliest dates of any major figure: AI smarter than any human by the end of 2026, “no later than” 2027.
Notice the pattern: the optimists’ argument is not really “AI will get smarter” but “AI will start doing AI research, and then progress will feed on itself” — the crux of the fast-takeoff debate (see our companion entry).
What the broad survey of experts says
Lab leaders have an interest in the answer, so it is worth asking the wider research community. The largest survey of its kind, published in 2024 by Katja Grace and colleagues, asked 2,778 researchers who had published at top AI venues. The aggregate forecast: a 10% chance that unaided machines outperform humans at every task by 2027, and a 50% chance by 2047. That median is 13 years earlier than the same survey found a year before — the expert consensus is moving earlier, but from a much later starting point than the labs.
On jobs specifically the survey was far more cautious: a 10% chance all occupations become fully automatable by 2037, and 50% only by 2116. Capability and economic replacement are different things, governed by cost, regulation, and trust as much as by intelligence.
What the skeptics say
A minority of prominent researchers thinks the optimists are extrapolating past a wall.
- Gary Marcus, a cognitive scientist, has said flatly: “We won’t get to AGI in 2026 (or 7).” He also predicted no AGI in 2025 — a call the scoreboard graded correct.
- Andrej Karpathy, a founding member of OpenAI, puts truly capable AI agents “about a decade” away, arguing the remaining problems are tractable but difficult.
- Christopher Manning, a Stanford professor, has called claims of AGI by 2030 “laughable.”
- Ajeya Cotra, a researcher at Open Philanthropy, gives a median of around 2040 for transformative AI, with only about a 15% chance by 2030.
The skeptics’ core argument is that today’s systems are impressive pattern-matchers that still fail at the things humans find easy: sustained reasoning, common sense, learning from a handful of examples, and knowing when they are wrong. Scaling up the current approach, in this view, buys fluency, not understanding.
Why the estimates disagree so much
Four things drive the spread:
- Definitions. A “chatbot that passes a test set” is a far lower bar than “a system that can do any remote job.” Some forecasters are implicitly using the first; others the second. Until the definition is fixed, two people can both be right.
- Extrapolation vs. bottlenecks. Optimists extend the trendlines of the last few years — computing power, data, and algorithmic progress — forward. Skeptics point to the physical limits: experiments take real time, chips and energy are finite, and some problems may need new ideas, not more scale. Amodei himself lists these limits in his essay: the speed of the outside world, the need for data, intrinsic complexity, human constraints, and physical law.
- What counts as evidence. Lab leaders see internal systems the public cannot. Outside researchers see only published benchmarks. Both sides are partly guessing about what the other side knows.
- Incentives. Lab leaders benefit from excitement; critics benefit from contrarianism. Neither fact settles who is right, but both are reasons to weight the independent survey and the track record — which is what the scoreboard is for.
What to watch
Judge the forecasts yourself by watching the milestones the forecasters named: full automation of coding, AI doing AI research autonomously, and AI completing tasks that would take a skilled person days or weeks — a milestone Anthropic says could come in 2027. The scoreboard tracks dated claims on all of them, so both sides will be graded in public.
What the scoreboard says
SiPoly’s AGI & superintelligence timelines topic tracks dozens of dated AGI forecasts. Gary Marcus’s “No AGI in 2025” was graded correct (claim page); Elon Musk’s April 2024 claim of AGI “within two years” was graded wrong. Sam Altman’s claim that superintelligence would arrive “in a few thousand days” could not be graded at all — no definition, no date — and was listed as too vague (claim page).
Still open: Altman’s internal-AGI-by-end-of-2026 claim, Benaich’s AGI in 2027, and Kokotajlo’s 50% chance of automated AI research by the end of 2028. The longest-dated open claim is the 2023 expert survey itself — a 50% chance of machines outperforming humans at every task by 2047 (claim page). Across the whole scoreboard, as of 5 November 2026, 367 claims are tracked, 129 resolved, and 60% came true.
Key sources
- Grace, K. et al., “Thousands of AI Authors on the Future of AI” (2024) — the 2,778-researcher survey; 50% by 2047 for machines outperforming humans at every task. https://arxiv.org/abs/2401.02843
- Kokotajlo, D. et al., AI 2027 — a detailed scenario with dated milestones for coding automation and AI research automation. https://ai-2027.com/
- Amodei, D., “Machines of Loving Grace” (October 2024) — defines “powerful AI” and lists the factors that could limit its speed. https://darioamodei.com/machines-of-loving-grace
- Aschenbrenner, L., Situational Awareness: The Decade Ahead (June 2024) — the optimist case for AGI by 2027. https://situational-awareness.ai/
- Morris, M. R. et al., “Levels of AGI for Operationalizing Progress on the Path to AGI” (2023) — a framework for what “AGI” could mean. https://arxiv.org/abs/2311.02462
Common misconceptions
- “Experts agree AGI is coming in 2027.” They do not. Lab leaders cluster around the late 2020s; the broad survey of researchers says 2047 at the median; skeptics say later still or never on the current approach. The disagreement is the story.
- “A lab announcing AGI means AGI has arrived.” A claim is not a measurement. Such claims are usually too vague to grade — which is why Amodei’s precisely defined “country of geniuses” is more useful than a bare announcement.
- “The date is the only thing that matters.” The definition matters just as much. Two forecasters can name the same year and mean entirely different things.