When Will AGI Actually Arrive? Every Forecast, and Who Is Making It

I wrote a whole article on what AGI actually means and deliberately refused to say when it arrives. This is the follow-up, because "when" is the question everybody asks next and almost nobody answers honestly.

The honest answer starts with an admission: nobody knows. But that is not the same as "all guesses are equal". Some forecasts come from methods with track records; some come from people selling something. Learning to tell them apart is the actual skill here, and it takes about ten minutes.

The short answer

The forecasts, side by side

Here is the landscape. Note that each row is answering a slightly different question — which is the whole point of the next section.

Look at the spread: 2026 to 2047. Two decades between credible sources.

Most articles pick one row, put it in a headline, and move on. That is why the topic feels like noise.

Why the forecasts differ so wildly

Three reasons, and only the third is about disagreement.

1. They are answering different questions

This explains most of the spread. "Weak AGI" and "high-level machine intelligence" are not the same bar — Metaculus itself has run both, and its own forecasters put roughly five years between them.

Stretch that further, to "outperforms humans at essentially all tasks including physical ones", and you have added the entire robotics problem to a software forecast. That alone can move a date by decades.

2. The methods have different biases

Each forecasting method fails in a characteristic direction, which is useful once you know it:

• Prediction markets reward being right and punish being vague, which is good. But they are populated by people who follow AI closely — a self-selected group that skews enthusiastic.
• Expert surveys capture people who build the systems, and researchers are famously conservative about their own field's timelines. They also historically under-predicted specific capabilities that then arrived early.
• Compute extrapolation is the most mechanical and therefore the most legible — but it assumes the thing being extrapolated is the thing that matters, which is exactly what is in question.
• Lab leaders have information nobody else has, and a direct financial interest in the answer. Both facts are true at once and neither cancels the other.

3. Genuine uncertainty about whether scaling is enough

Underneath the definitional noise there is a real scientific disagreement: does continuing to scale current approaches get you to general intelligence, or is something conceptually missing?

If scaling suffices, the near-term forecasts are reasonable. If something is missing, no amount of compute fixes it and the far forecasts are right. Nobody knows which, and that is not a failure of the forecasters.

How to read any AGI forecast in ten seconds

Four questions. Ask them in order and almost every AGI headline becomes legible.

Run that on the next AGI headline you see. It will usually collapse into "someone asserted a near date using an economic definition, and stands to benefit."

The signal that matters more than any date

Here is the thing I would actually take away from all of this.

Almost every forecaster has moved their estimate earlier, and the moves are large. The largest recent expert survey shifted its median forward by roughly 13 years compared with the survey conducted a year before. That is not a small correction. That is a field re-evaluating.

A 13-year move in one year tells you more than the resulting number does, for two reasons:

• It shows the direction of surprise. When forecasts move, they have been moving one way. Capabilities have repeatedly arrived earlier than surveyed researchers expected.
• It tells you the current number is unstable. A median that moved 13 years in twelve months will very likely move again. Treating 2047 as a settled fact would be a misreading of the same survey that produced it.

So if you want one thing to watch, do not watch the date. Watch the revisions.

Which forecast should you actually use?

It depends entirely on what the answer changes for you.

Notice how often the answer is "this forecast is not relevant to your decision". Most practical questions are answered by what models can do today — which is why I spend more time on what the benchmarks actually measure than on predicting dates.

The AGI timeline matters enormously for policy and safety. For most professional decisions it is, honestly, a distraction dressed as strategy.

FAQ

When will AGI arrive?

There is no consensus date. Prediction-market forecasters have centred on the late 2020s to early 2030s, compute-extrapolation models point to the 2030s, and the largest recent expert survey puts the median at 2047 for machines outperforming humans at essentially all tasks. The spread is roughly two decades, and most of it comes from the forecasts answering differently-worded questions rather than from disagreement about AI itself.

Why do AGI predictions vary so much?

Three reasons. Most of the spread comes from definitions — Metaculus forecasters put about five years between "weak AGI" and "strong AGI" on their own questions, and adding physical tasks to the definition can move a date by decades. The rest comes from method bias, since markets, surveys and extrapolations each fail in a characteristic direction, and from a genuine open question about whether scaling current approaches is sufficient.

Do experts think AGI is coming sooner than they used to?

Yes, and this is the clearest signal in the data. The largest recent expert survey moved its median forward by roughly 13 years compared with the survey a year earlier. The direction and size of that revision is more informative than the resulting number, because it shows capabilities have repeatedly arrived earlier than researchers expected and suggests the current estimate will move again.

Is AGI arriving in 2026?

Not on any strict definition. Aggregated community forecasting put roughly a 10% probability on pure AGI in 2026. Some Metaculus questions on looser "weak AGI" criteria could resolve near-term, and some lab leaders have described the current period as the beginning of an AGI era, but those claims rest on economic or benchmark definitions rather than the capability definition most people have in mind.

Which AGI forecast is the most reliable?

None individually. Prediction markets reward accuracy but draw on a self-selected, enthusiastic population; expert surveys capture the people building the systems but are historically conservative; compute extrapolations are transparent but assume the thing being scaled is the thing that matters. The most defensible approach is to read the resolution criteria first, then treat the range as a probability distribution rather than picking a favourite number.

Should AGI timelines change what I do?

For most professional decisions, no. Whether to adopt a tool, hire, or learn a skill is answered by what models can do today rather than by a forecast — and current capability is measurable in a way a date is not. AGI timelines matter substantially for policy and safety work, where the cost of preparing too early is far lower than the cost of preparing too late.

• Metaculus — the prediction-market forecasts and the weak/strong AGI split
• AI Multiple, AGI and singularity timing: ~10,000 predictions analysed
• 80,000 Hours, What happened with AGI timelines — on the scale of recent revisions
• FutureSearch, AGI timeline tracker: how top forecasters updated, 2023–2026
• Jayant Solanki, what AGI actually means, and who decides we have reached it

Jayant Solanki

AI-Ready SEO, GEO & AIO strategist based in Indore, India. I write explainers that define their terms and label their evidence, and that say plainly where the argument stops being factual.

Google Analytics Certified (2026–2027) · verify

More about Jayant →