In 2023, a survey of 2,778 researchers who had published at major AI venues put the median estimate for a 50% chance of high-level machine intelligence at 2047. A similar exercise two years earlier landed on 2060. The underlying physics didn’t change in between. The models the researchers were looking at did.
That 13-year swing is the most useful thing to know about any AGI forecast. It isn’t a measurement. It’s a judgement call about a machine nobody has finished building, made by people reading the same incomplete evidence and weighing it differently.
None of which makes forecasts useless. It makes them something you have to read the way an analyst reads an earnings projection: check the assumptions, check the incentives, and check what would have to be true for the number to be right.
What an AGI forecast is actually predicting
Ask ten researchers when AGI arrives and you may get ten answers, partly because the word is doing different work in each head. Three distinct claims usually get bundled together:
- Capability parity. A system that handles most cognitive tasks about as well as a motivated adult, including domains it wasn’t specifically trained on.
- Autonomy. A system that can be handed a goal and left alone for days, recovering from its own mistakes without a human babysitter.
- Economic substitution. AI that can do most paid work at a lower cost than the people currently doing it.
These don’t arrive together. Models already beat most adults on certain narrow tasks and fall apart on others, and there’s a real argument that the gap between “impressive demo” and “trusted to run unsupervised” is where the years get eaten. What the AGI countdown clock is actually measuring is often a definition rather than a technology.
Four camps, four kinds of evidence
Most published forecasts come from one of four traditions. Knowing which one you’re reading tells you a lot about how much weight to give it.
The extrapolators
These forecasters fit curves to compute, data and spending, then extend them forward. The logic has held remarkably well for a decade: training compute for frontier models has grown roughly four- to fivefold a year, and capability benchmarks have toppled in a fairly predictable order. The weakness is that a trend line describes the past. It can’t tell you where a wall is until you hit one.
The scenario builders
A smaller group writes detailed narratives, complete with dates and internal logic, about how the final few years before AGI might unfold. These are valuable as stress tests and dangerous as predictions. A vivid story feels more probable than a vague one, which is a quirk of human cognition and not a property of the world.
The survey takers
Aggregate expert opinion, like that 2023 paper, is the closest thing to a poll of the field. Useful as a snapshot of consensus. Less useful as a forecast, because the sample skews toward people who work on the problem daily and have watched several confident predictions miss.
The sceptics
Yann LeCun has argued for years that current architectures lack the world models needed for genuine reasoning, and that the timeline runs to decades at best. Sceptics are usually right about the specific mechanism being invoked and occasionally wrong about how much brute force can substitute for it.
Why the range runs from 2026 to never
The spread isn’t noise. It reflects honest disagreement about which bottleneck binds first. If progress is mostly a matter of scaling compute and data, the frontier could be close, since spending on individual training runs has jumped from millions to billions in a few years. If the real constraint is continuous learning from experience, or reliable multi-step reasoning, the timeline stretches out considerably.
Forecasts have spanned roughly three decades for a decade now, and the width has barely narrowed even as models got dramatically better. Why AGI forecasts span 30 years, and how to read them comes down to which constraint each forecaster treats as load-bearing.
Reading a forecast like an analyst
Before you repeat a date at a dinner party, run it through five questions.
- What’s the definition? A forecast that never specifies what AGI means can’t be wrong, which makes it close to worthless.
- Who’s talking, and what do they gain? A lab founder with a funding round closing, a researcher with a book out, and an anonymous prediction-market trader face very different incentives.
- Is it falsifiable? “AGI by 2029” is a claim you can check. “AGI is coming sooner than you think” is marketing copy.
- Is it conditional? Plenty of honest forecasts are really “if scaling keeps working, then…” That single word is doing enormous work.
- What’s the base rate? AI has a long history of confident timelines, going back to the 1960s, that missed by decades.
The numbers worth knowing
For all the disagreement, the published figures cluster in a few places. Ray Kurzweil has held to 2029 for years. Demis Hassabis has described AGI as roughly a decade away. Dario Amodei has pointed at 2026 or 2027 for what he calls powerful AI. Prediction markets have generally sat later than the most bullish lab leaders and earlier than the survey medians, while the aggregate of expert surveys puts the 50% mark somewhere between the late 2040s and 2060 depending on the year you ask.
One pattern is worth noticing. The people closest to the models tend to give the shortest timelines, and the people who study the history of the field tend to give the longest. Both are responding to real evidence. One sees capability gains week by week; the other sees how often those gains have failed to compound into general intelligence. Who’s saying what on AGI in 2026, and which forecasts hold up is a reasonable way to watch those numbers shift.
Watch signposts, not dates
Forecasting energy is better spent on indicators that would have to flip before any near-term date becomes plausible. Concretely, here’s what to watch:
- Agents that finish multi-day projects with minimal supervision, not one-hour demos.
- Benchmark saturation followed by the field agreeing on harder tests, rather than quietly retiring the old ones.
- Cost per unit of capability continuing to fall fast, at roughly an order of magnitude a year for a fixed quality level.
- Revenue from AI doing work rather than assisting it.
- Systems that measurably improve from their own experience between deployments.
The reliability gap matters more than raw capability here. A model that is 95% accurate on a task is a helpful assistant. A model that is 99.9% accurate across a ten-step chain, and can notice and correct its own errors, is a different kind of object entirely. What the evidence actually suggests about whether AGI will happen soon hinges on that distinction far more than on headline benchmark scores.
What to do with an AGI forecast when you have to decide something
Most people reading these numbers aren’t trying to win a bet. They’re deciding whether to restructure a team, commit to a five-year platform, or change what they study. For those decisions, a point estimate is the wrong tool.
Think in terms of options and reversibility instead. If AGI is ten years out, money spent on data infrastructure, evaluation pipelines and AI-literate staff pays off either way. If it’s three years out, the same investments pay off faster and the cost of waiting is much higher. The overlap between those two worlds is wider than the headlines suggest, which is why so many sensible organisations are doing roughly the same things regardless of which timeline they believe.
Then set a review date. Pick a specific signal, say a model completing a week of unsupervised engineering work at professional quality, and decide in advance what you’d change if it arrived. That turns an unfalsifiable argument into a decision rule you can act on. Forecasts are poor at telling you when. They’re decent at telling you what to watch while you wait.