AI 2027 Scorecard: 4 Hits, 3 Misses, 15 Months On

In short: AI 2027's forecasts for money, hiring and geopolitics are landing. Its forecasts for the actual machine are slipping by 10-100x. That split matters more than the score, because it means the economic consequences of superintelligence have arrived before the superintelligence has — and the response an SME should have to that is close to the opposite of what the headlines suggest.

The AI 2027 scenario was published in April 2025 by Daniel Kokotajlo and colleagues at the AI Futures Project. Unlike most AI forecasting, it is specific enough to be graded: it names dates, compute figures in FLOP, dollar amounts, power draw in gigawatts, and labour-market outcomes. Fifteen months have passed. Roughly seven twelfths of its 2026 predictions can now be checked against published numbers.

So we checked them. Every figure below is the scenario's own stated number against the best published data as of late July 2026, with sources. Where the evidence is contested, we say so rather than picking the flattering number.

The scorecard

Scoring key: HIT — the stated figure or outcome materialised within a reasonable margin. PARTIAL — directionally right, materially off on magnitude or definition. MISS — the stated figure or outcome did not occur.

Four hits, two partials, three misses. On its own that is a respectable record for a document written before any of it happened, and far better than the average pundit. But the score is the least interesting thing in the table.

The pattern: it predicted the consequences better than the cause

Sort those nine rows not by verdict but by what each one depended on, and the noise resolves into a signal.

Everything AI 2027 got right is something that resolves on belief. Capital expenditure is a bet placed on future capability. A hiring freeze is a bet. Export controls and compute shares are bets made by states. None of them require the AI to actually work — they require enough people to believe it will.

Everything it got wrong is something that resolves on capability. The 10^28 FLOP training run is a physical fact about a machine, and it is 10 to 100 times short. The 1.5x research multiplier is a measured productivity fact, and the best controlled measurement puts it at 1.18x. The self-improving research loop that the whole back half of the scenario depends on has not started.

Kokotajlo has said as much himself: he moved his own median estimate for recursive self-improvement from roughly 2028 out to 2030. The author of the aggressive timeline revised it downward while the financial and labour-market consequences he described arrived on schedule.

That is the finding worth carrying: we are living through the economic consequences of AGI without AGI. Roughly $700B a year in capital expenditure is chasing perhaps $30B of leading-lab revenue. A generation of junior engineers is being turned away by employers citing automation that, measured properly, is delivering 18%.

Where the story is actually wrong, not just early

Row 3 is the one everybody cites as proof the scenario is coming true, and it is the row that deserves the most scrutiny — because AI 2027 got the outcome right through the wrong mechanism.

The scenario's causal claim is that AI automates coding, so junior coders are displaced. The data does not support that mechanism carrying the effect. Analyses of the collapse attribute roughly 10% of it to AI, with the balance falling to budget cuts and interest rates. The tell is in the hiring pattern itself: postings labelled "entry-level software engineer" grew 47% between October 2023 and November 2024 while actual hiring into those levels fell 73%. Firms are posting junior roles and filling them with experienced engineers. That is not a story about automation replacing juniors. It is a story about a buyer's market letting employers buy seniority at junior prices.

A right prediction for the wrong reason is more dangerous than a wrong one, because it produces confident decisions built on a mechanism that is not operating. If you believe AI ate the junior developers, you stop hiring juniors and you buy tools. If you understand it is roughly 90% macro, you reach the opposite conclusion, and you reach it while your competitors are looking the other way.

What this means if you run a small or mid-sized business

Three decisions follow from the split between what landed and what slipped.

1. Plan for 1.2x, not for AGI

The capability line is the one that slipped, so stop planning around a 2027 step-change and plan around the multiplier you can actually measure. METR's number is the most instructive figure in this entire scorecard, and not because of its size: the same study found experienced developers were 19% slower with AI tools a year earlier, then 18% faster a year later. A 37-point swing on the same population and the same class of tool.

That swing is not a tool upgrade. It is a learning curve. Which makes AI adoption a training and process problem, not a procurement problem — the oldest lesson in Kaizen, arriving in new clothing. The organisations getting the 1.8x are not the ones that bought the best tool; they are the ones that had a defined process to slot it into. In a weak organisation AI reliably surfaces the existing mess rather than fixing it.

Practical version: pick one process you can already describe end to end, measure its current cycle time, then introduce the tool. If you cannot describe the process, the tool has nothing to attach to and you will be in the -19% cohort for months.

2. Rent, do not commit

$700B of annual capex chasing $30B of frontier revenue is not a stable price structure. Inference prices have one direction to travel while that gap persists, and the scenario's own model-cost trajectory assumed exactly this: each generation's small model arriving roughly 10x cheaper.

Practical version: avoid multi-year commitments at today's prices, keep switching costs low, and prefer per-token or per-seat arrangements you can exit. Anything you would build to own today gets cheaper to rent in twelve months. The build-versus-buy line has moved, and it will move again.

3. Hire the talent everyone else is refusing

This is the contrarian one and it follows directly from the mechanism error. If ~90% of the junior collapse is macro rather than automation, then the pool is mispriced, not obsolete. Entry-level candidates are taking 5-6 months and 200+ applications to place. Meanwhile the specialisation everyone wants shows a 63% talent shortage with 500,000+ open roles — so the shortage and the surplus are sitting in the same market at the same time, which is the definition of a hiring opportunity for whoever is paying attention.

Practical version: an SME that can absorb two junior hires and pair them with AI tooling and a defined process is buying, at a discount, the capacity its larger competitors have frozen. The constraint on your business was probably never headcount cost. It was throughput. This is the cheapest throughput available in a decade.

The honest limitations of this scorecard

Nine rows of a scenario that runs to December 2027 is a partial grade, and the scenario's central claims — the superhuman coder in March 2027, superintelligence by December — remain untested and cannot be graded from here. Three further caveats worth stating plainly:

• Definitions move the verdicts. "Global AI capex" has no agreed boundary. Include chips, power and non-hyperscalers and $1T for 2026 is arguable; restrict it to listed data-centre operators and it is not. Row 5 could be scored either way, which is why we scored it PARTIAL rather than picking a side.
• Productivity measurement is genuinely unsettled. METR's 1.18x and the industry's self-reported 1.7x are not measuring the same thing on the same population. We weighted the controlled measurement, and a reasonable analyst could weight it differently.
• Half a year is left in 2026. Rows 5, 6 and 9 can still move. A strong Q4 in the markets or a $200B capex announcement changes two of them.

The AI 2027 Tracker maintains a running, prediction-by-prediction status board and is the right place to check whether any of the above has shifted since publication. Our contribution here is not the tracking — it is the split between belief-driven and capability-driven predictions, and what that split implies for the operating decisions of a business with fewer than 500 people.

Frequently Asked Questions

Is the AI 2027 scenario accurate so far?

Partially. Of nine checkable 2026 predictions, four landed, two were directionally right but materially off, and three missed. Its forecasts about money, hiring and geopolitics have held up well. Its forecasts about actual machine capability — a 10^28 FLOP training run, a 1.5x AI research multiplier — are 10-100x short, and its author has since moved his own median for recursive self-improvement from about 2028 to 2030.

Did AI 2027 correctly predict the junior developer job collapse?

It predicted the outcome but not the cause. Entry-level hiring is down about 73% and new graduates fell from 32% of Big Tech hires to 7%, which is at least as severe as the scenario described. But analyses attribute only around 10% of the collapse to AI, with budget cuts and interest rates carrying the rest. Postings for junior roles actually rose 47% while hiring into them fell 73% — employers are filling junior openings with experienced engineers, which is a buyer's-market story rather than an automation story.

Should a small business plan for AGI by 2027?

No. Plan for the measured multiplier, which is roughly 1.2x on tasks you can specify precisely, and treat larger gains as a training outcome rather than a purchase. The capability predictions are the ones slipping, while the cost and labour-market effects are already here. Build for cheaper inference and a loose commitment structure, not for a step-change on a fixed date.

What is the single most useful takeaway from grading AI 2027?

That the economic consequences of superintelligence arrived before the superintelligence. Roughly $700B of annual capital expenditure and a frozen junior hiring market are responses to an expectation, not to a delivered capability. Businesses that understand the difference can act on the mispricing — most visibly in hiring — instead of reacting to the narrative.

How much faster does AI actually make developers in 2026?

The best controlled measurement, from METR, is an 18% speedup for experienced developers in early 2026. The same study measured a 19% slowdown a year earlier, so the 37-point swing reflects a learning curve rather than a change in the tools. Industry self-reports cluster nearer 1.7x, and AI now writes somewhere between 27% and 50% of production code depending on the survey.