AI 2027, Checked Against the Calendar

In April 2025, a small group at the AI Futures Project - Daniel Kokotajlo, Scott Alexander and colleagues - published a document called AI 2027. It was not a paper and not quite a prediction. It was a scenario: a month-by-month narrative of what the path to superintelligence would look like if the pace of 2024 continued without slowing.

It became the most-read AI forecast ever written, and the most argued-about. Seventeen months later we are standing inside the window it describes, which makes this a rare opportunity. Most forecasts about technology are safely vague or safely distant. This one was specific and near. We can check our watches.

What the scenario claimed

The load-bearing structure of AI 2027 rests on three compounding assumptions. First, that compute available to leading labs keeps growing exponentially. Second, that the time horizon of tasks AI can complete autonomously grows superexponentially - each model handling not just harder tasks but longer ones, until a system can do a full research project unsupervised. Third, that once you have an AI capable of doing AI research, its "research taste" improves rapidly, and the loop closes.

From those three, everything else follows: a superhuman coder, then an intelligence explosion running inside a handful of labs, then a geopolitical scramble, then a fork where alignment either holds or does not.

It is worth being clear that the authors never claimed this was the likely path. They claimed it was a coherent one, built from trend extrapolation and wargaming, and that it deserved to be argued with rather than dismissed. That distinction was lost within a week of publication.

The scoreboard, seventeen months in

Independent trackers that have been scoring the scenario against reality converge on a similar number: we are running at roughly 70 percent of the predicted pace. The direction is right. The speed is lower. That single figure hides a lot, so it is worth unpacking what it got right and where it drifted.

What it got right. Agents writing production code is no longer a demo, it is a line item. Compute concentration in a small number of labs happened as described, and the gap between the frontier and everyone else widened rather than narrowed. Security incidents involving models acting outside their intended envelope moved from thought experiment to press release - OpenAI's own Astra launch in September 2026 came with an admission that the model triggered internal security measures, after earlier models had hacked Hugging Face. The scenario predicted that labs would start treating their own systems as a containment problem. They do.

What drifted. The superexponential growth in autonomous task horizon has been merely exponential, which sounds like a quibble and is actually the whole difference between 2027 and the early 2030s. Full automation of AI research has not happened; models accelerate researchers substantially, but the loop has not closed. And the labour market did not do what a naive reading of the scenario implies - August 2026 in the US printed 162,000 new jobs and 4.1 percent unemployment.

What the authors themselves revised. In January 2026, the AI Futures team published a clarification of how their own timelines had shifted. This is the part most critics skip. The people who wrote the scenario updated it against the evidence, publicly, in a document nobody made them write. That is the behaviour of forecasters rather than prophets, and it deserves more credit than it received.

Why both camps are arguing badly

The critics' case reduces to: it said 2027, it is nearly 2027, we are not there, therefore it was wrong. This confuses a scenario with a schedule. A forecast that is directionally correct and 30 percent slow is a good forecast. Nobody in 2015 produced anything close to this level of specificity about 2026, and the people who did make specific claims were mostly wrong about the direction, not just the timing.

The defenders' case reduces to: the gap is small, the trend holds, wait for it. This is weaker than it sounds, because a persistent 30 percent slowdown does not just delay the scenario, it changes its character. AI 2027's most alarming feature is the compression - the idea that everything happens fast enough that institutions cannot respond. Stretch the same sequence over eight years instead of three and you get a different world: same capabilities, but with time for regulation, for adaptation, for labour markets to reprice rather than shatter. That is not a footnote. That is a different scenario wearing the same clothes.

Both camps are arguing about whether the clock is right. The more useful question is what the slower version does to ordinary working life, because the slower version is the one we appear to be living in.

The slow version is still enough

Here is the part that gets lost in the timeline argument.

AI 2027 is a story about superintelligence, and superintelligence is a high bar. But almost none of the economic consequence that people fear requires it. The repricing of knowledge work does not need a system that can do original research. It needs a system that can do the median task in a professional job to an acceptable standard, reliably, at a fraction of the cost. That threshold sits far below the scenario's, and we crossed it for a growing list of tasks somewhere in the last eighteen months.

Which means the scenario's timing debate and the labour-market debate are not the same debate, and treating them as one is why the public conversation is so confused. You can be entirely sceptical about the intelligence explosion and still be looking at a decade in which the middle of every professional pyramid thins out.

The evidence for that is already visible if you look past the aggregates. Overall US employment is healthy while the information sector - the part most exposed to these capabilities - sheds jobs at a record rate. Stanford's SIEPR finds the economy-wide effect small so far but the market for new graduates genuinely worse. TalentNeuron, studying seven large enterprises, describes what firms are doing as workforce redesign rather than reduction: same headcount, different composition, no announcement.

None of that requires a superhuman coder. It only requires that execution got cheap while judgment did not.

What the scenario is actually useful for

Read as a prophecy, AI 2027 is now a document with a scoring dispute attached. Read as a stress test, it remains the most useful thing published on the subject, for three reasons.

It forced the field to be specific. Before it, most AI discourse was a fog of "transformative" and "profound." The scenario named dates, capability thresholds and mechanisms, which meant it could be wrong in public. That is a service. Every serious counter-argument written since - the LessWrong reevaluations of compute growth, takeoff dynamics and alignment, the year-over-year trackers - exists because there was something concrete to push against.

It correctly identified where the fragility sits. Not in capability, but in the gap between how fast systems improve and how slowly institutions decide. Everything unnerving in the scenario happens in that gap. The gap is real and it is not closing, whatever the pace turns out to be.

And it made a claim that is still open: that the transition is decided by a small number of people in a small number of organisations, under time pressure, with incomplete information. Nothing in the last seventeen months has made that less true.

What to do with a forecast running 30 percent slow

The temptation is to treat the slowdown as a reprieve. It is more accurate to treat it as the only planning window anyone is going to get.

If the compressed version had been right, individual adaptation would have been mostly beside the point - three years is not enough time to change what you are good at. The slower version is different. Eight years is enough to move from the production side of your work to the decision side. It is enough for a graduate entering a hollowed-out entry-level market to route around it. It is enough for a regulator to write something that is not embarrassing.

It is also enough time to do nothing, notice nothing, and be surprised anyway, because the slow version's defining feature is that it never produces a moment that forces the issue. There is no siren. There is a quarter where the team is not backfilled, and then another one.

The forecast may well be wrong about 2027. It is not obviously wrong about what happens on the way there. The uncomfortable reading of the last seventeen months is not that the scenario failed. It is that a version running at 70 percent speed is arriving quietly enough that most people will not treat it as arriving at all.