Ask a room of Google researchers when AGI arrives and you’ll get a dozen hedged answers. Ask the two people who founded DeepMind and you get something rarer: numbers, repeated often enough that you can hold them against a calendar.
That split matters. Google’s AGI predictions aren’t a corporate position so much as a set of overlapping personal forecasts from a handful of senior people, each attached to different definitions and different caveats. Pull them apart and you learn less about the future than about what the people building the systems think is still missing.
The Three Voices Behind Google’s AGI Timeline
Most of what gets quoted as “Google’s AGI prediction” traces back to three people: Demis Hassabis, Shane Legg, and to a lesser degree Sundar Pichai. They don’t agree, and the disagreement is informative.
Demis Hassabis: five years, framed as a coin flip
Hassabis has settled into a range. In 2023 he was saying AGI could be five to ten years out “if we get it right.” By early 2025, speaking at Davos and then at Mobile World Congress, he had tightened it to roughly five years, while adding the qualifier that it would take “one or two more big breakthroughs” to get there.
He is specific about what those breakthroughs are. Current models, in his telling, are missing continuous learning, long-term planning, persistent memory, and genuine invention. A system that forgets your entire working relationship every time a context window closes is not, by his definition, generally intelligent. That list is the most useful part of the prediction, because each item can be watched.
Shane Legg: 2028, held for fourteen years
Legg is the outlier who looks prescient or stubborn depending on the year. The DeepMind co-founder wrote on his personal blog back in 2011 that he expected human-level machine intelligence around 2028, and he has described it since as roughly a 50% proposition. He has not moved the date much through three model generations and two hype cycles.
That makes it the most falsifiable prediction on the board. If you want a single date to grade Google’s culture against, 2028 is the one, not the vaguer five-year framing that keeps resetting every January.
Sundar Pichai: deliberately dateless
Google’s CEO has said AGI is “not a matter of if, only when” and has gestured at “the next decade” without committing. That’s not evasion for its own sake. A CEO who names a year hands shareholders a metric and regulators a deadline. Hassabis can say five years because he runs a research lab; Pichai has to ship products against it.
Google Already Published a Definition That Makes the Date Look Early
In November 2023, a DeepMind team led by Meredith Ringel Morris published a paper called “Levels of AGI,” and it is the closest thing to an official Google yardstick. It grades systems on two axes: performance (Emerging, Competent, Expert, Virtuoso, Superhuman) and breadth (narrow versus general).
- Level 1, Emerging: equal to or better than an unskilled human at a wide range of non-physical tasks.
- Level 2, Competent: at or above the 50th percentile of skilled adults.
- Level 3, Expert: at or above the 90th percentile.
- Level 4, Virtuoso and Level 5, Superhuman: beyond the best humans, then beyond all humans combined.
By that rubric, a 2023-era chatbot already qualified as Emerging AGI. Which tells you how much work a definition does before a date means anything. If you want the full unpacking of why the same phrase produces wildly different timelines, this breakdown of what AGI actually means in plain English is a useful starting point. The companies themselves don’t even agree on the wording: OpenAI’s version is famously compressed into fourteen words that get argued over constantly.
Why the Gap Between Demo and Deadline Is the Whole Story
Hassabis’s five-year figure assumes progress continues at roughly the current slope. The counterargument isn’t that scaling fails; it’s that the gap between a striking demo and a system that does reliable work for eight hours straight has been stubbornly wide.
Gemini can pass a bar exam and still lose the thread three steps into a multi-file code change. It can summarise a hundred pages and still invent a citation. Those aren’t scaling artefacts so much as different problems, which is why researchers studying the distance between demos and real work keep landing on the same list of blockers: reliability, verification, and the cost of human oversight.
The most concrete measurement here is task length. Independent evaluators track how long a task an AI can complete unaided, and that number has been doubling roughly every seven months. If it keeps doubling, systems that handle a full workday arrive well before 2030. If it flattens, Hassabis’s five years slip by default. For a deeper look at where the shortfall actually sits, this analysis of where the gap to human intelligence really is is worth the read.
Google’s Dates Aren’t the Most Aggressive in the Industry
Compared with its closest rivals, Google is somewhere in the middle. Anthropic’s Dario Amodei has talked about “powerful AI” arriving in 2026 or 2027. Sam Altman’s “few thousand days” framing points closer to the early 2030s, though OpenAI staff have privately floated 2027 in internal documents. Yann LeCun, until recently at Meta, has argued for a decade or more.
Those spreads look like disagreement about physics. Mostly they’re disagreement about vocabulary. A company defining AGI as an autonomous system that can do economically useful research will hit its milestone years before one that requires human-level performance across every cognitive domain. If you want to see how much a single company’s dates have drifted, this running record of every AGI timeline OpenAI has published is a decent calibration exercise.
How You’d Actually Know a Prediction Was Right
Benchmarks are the obvious answer and the weakest one, because they saturate and leak. DeepMind’s own evaluation suites are more interesting than most, but as this look at what DeepMind’s AGI benchmark scores actually measure points out, a high score is often a statement about test design rather than capability.
The tests that would settle it look different. A model that learns a new skill mid-conversation and still has it a week later. An agent that takes a vague goal, works for twenty hours, and returns something a competent human would sign off on. A system that proposes a research question nobody asked and then answers it. None of those are benchmarks yet, which is why the prediction remains open.
Signals Worth Watching Between Now and 2030
Forget the interviews. Watch what Google ships and what it stops saying.
Model releases are the obvious signal, but the more telling one is whether any Gemini generation is shipped with genuine on-the-job learning, where weights or an adapter update from use rather than from a fresh training run. That single feature would move Hassabis’s list further than another ten points on an exam.
Then there’s Google’s own safety paperwork. The Frontier Safety Framework defines Critical Capability Levels, including one for machine learning R&D autonomy, the point at which a model can meaningfully accelerate AI research itself. The first time Google announces that a model has tripped that threshold, the five-year estimate has effectively been confirmed in public, whatever the blog posts say.
And listen to the phrasing. Hassabis has been unusually consistent, which makes drift meaningful. If “five to ten years” becomes “a couple of years” in a keynote rather than an interview, that’s not PR, that’s a researcher updating on evidence. Legg’s 2028 stays the number to hold everyone to. It’s close enough now that nobody gets to revise it quietly.

