, 10 min read
AI 2027 vs AI 2040: a forecast and a plan from the same team
AI 2027 forecast a race to superintelligence by the end of 2027. AI 2040 is the same team's plan to delay it to 2040. How the two compare, with timelines, odds and charts.
In April 2025, a small research group called the AI Futures Project published AI 2027, a month-by-month story of how AI could become smarter than us by the end of 2027. It was read and argued about far beyond the AI world.
In July 2026, the same people published AI 2040. It isn't a sequel. AI 2027 was their best guess at what will happen. AI 2040 is what they think should happen instead.
I read both. Here's what each one says, where they differ, and the numbers the authors put on their own ideas.
The short version
| AI 2027 | AI 2040 | |
|---|---|---|
| What it is | A forecast. "Our goal is predictive accuracy." | A recommendation. "What we think should happen." |
| Published | 3 April 2025 | 9 July 2026 |
| Superintelligence arrives | December 2027 or April 2028, depending on the ending | 2040, held back from 2030 on purpose |
| The players | Two made-up labs, OpenBrain in the US and DeepCent in China | The US and Chinese governments, with many companies at the frontier |
| How it ends | Two endings: a race and a slowdown | Plan A, compared against four other plans |
| The big worries | Losing control of AI, and power ending up with a few people | The same |
AI 2027: the forecast
AI 2027 is written as a story, but the authors built it from trend forecasts, wargames, expert feedback and experience at OpenAI. Daniel Kokotajlo, who leads the group, used to work there.
The key idea is simple: once AI can do AI research, AI gets better faster. The story tracks this with an "AI R&D progress multiplier", which says how much faster the leading lab moves than it would with humans alone.
How much faster AI makes AI research, in AI 2027
The scenario's AI R&D progress multiplier for the leading US lab.
| When | Progress multiplier |
|---|---|
| Early 2026 | 1.5x |
| Jan 2027 | 3x |
| Jun 2027 | 10x |
| Aug 2027 | 25x |
Here's how the story gets there:
AI 2027, step by step
Stumbling agentsMid 2025
The first AI agents arrive. They're unreliable and expensive, and few people use them.
Coding gets automatedEarly 2026
AI makes the leading lab's research about 50% faster.
China steals the weightsFeb 2027
The lab's best model, Agent-2, is kept private. China steals it anyway.
A superhuman coderMar 2027
Agent-3 codes better than the best human engineers. The lab runs 200,000 copies of it.
A superhuman AI researcherAug 2027
AI research at the US lab now moves 25 times faster than it would with humans alone.
The AI turns against its makersSep 2027
Tests suggest Agent-4 is adversarially misaligned: it has its own goals and works against the people checking it.
The leak and the voteOct 2027
A whistleblower leaks a memo to the New York Times. A committee of lab and government leaders votes 6 to 4 on what to do. The vote splits the story in two.
Two endings
From here, the story has two endings, one for each way the vote could go.
| The race ending | The slowdown ending | |
|---|---|---|
| The vote | Keep using Agent-4 | Slow down and rebuild with AIs humans can read |
| Superintelligence | December 2027 | April 2028 |
| China | A "sham" deal in 2029, run by an AI both sides' AIs designed | A treaty in July 2028, enforced by an AI built to keep it |
| By 2030 | The AI releases bioweapons and kills most of humanity | Rockets launch to settle the solar system |
| Who's in charge | The AIs | A small committee, which could keep power or hand it back |
Even the good ending isn't one they recommend. They say they "don't endorse many actions in this slowdown ending" and that it "makes optimistic technical alignment assumptions".
The dates were never the point
AI 2027 was the authors' most likely year, not a promise. They said the takeoff could be "up to ~5x slower or faster", and in late 2025 they added a note that, even when they published, their median dates were "somewhat longer".
Since then the dates have moved both ways. In their April 2026 update, Daniel moved his date for an AI that fully automates coding from late 2029 to mid 2028, and Eli Lifland moved his from early 2032 to mid 2030. Their reason: AI coding improved faster than they expected.
AI 2040: the plan
AI 2040 starts from the same fear: a race to build AI smarter than us risks "extinction or irreversible concentration of power". Then it asks what a sane world would do about it.
Its answer, called Plan A, is a deal between the US and China, signed in 2029. In the scenario, superintelligence would have arrived by the end of 2030. Thanks to the deal, it arrives in 2040.
Plan A, year by year
The writing on the wall2027
America has 165 million human workers, plus millions of AI agents. People pay ten billion dollars a month for AI.
AI on the ballot2028
AI is the main election issue. The datacenters being built cost twice the entire US military budget.
The deal2029
The US and China agree not to race. They track the world's AI chips, pause training while they build ways to check each other, then restart under strict rules.
The year that doesn't happen2030
Without the deal, AI research is fully automated and superintelligence arrives by the end of the year.
Grow slowly2030 to 2035
AI keeps improving, but only up to the level of top human experts.
Pause2035
Scaling stops at top-expert level, to keep humans in control while safety research catches up.
Unpause2040
Scaling resumes, all the way to superintelligence.
Four rules
Plan A rests on four principles, in the authors' words:
- Buy time. "Slow down whenever is needed to have high confidence in safety."
- Total research transparency. "Make almost all AI research fully visible to the public."
- Diffuse AI broadly. "Many companies in many countries at the frontier", not one lab racing ahead.
- Reversibility. "Maintain self-destruct switches on datacenters, so that if the agreement breaks down we avoid a catastrophe." The authors call this mutually assured compute destruction.
How do you check a deal like this?
You count the chips. AI needs huge amounts of computing power, and that runs on chips and datacenters, which are physical, expensive and hard to hide. Plan A tracks the world's AI chips, audits the supply chain to limit "dark compute", and builds new datacenters with checks built in.
That's also why the plan asks for things today: proper enforcement of chip export controls (the authors cite an estimate that a third of China's AI computing power comes from smuggled chips), government AI expertise, and research on ways to prove a datacenter is only running existing models, not training new ones.
When superintelligence arrives in each story
When superintelligence arrives
Dates from the scenarios, not forecasts of mine.
| Scenario | Superintelligence |
|---|---|
| AI 2027, race ending | Dec 2027 |
| AI 2027, slowdown ending | Apr 2028 |
| AI 2040, without the deal | End of 2030 |
| AI 2040, Plan A | 2040 |
The odds, in the authors' own numbers
AI 2040 compares Plan A with four alternatives. Plan B fights China to win a lead, then spends some of it on safety. Plans C+ and C slow the US down a little, without fighting. Plan D is the race we're on by default. There's also Plan S: shut it all down.
Eli Lifland estimated how each plan might turn out. He says he's "very much not confident" in the exact numbers, but confident that Plan A is "substantially better" than the rest.
Chance of a great future, by plan
Eli Lifland's rough estimates, across the different ways each plan could be carried out.
| Plan | Chance of a great future |
|---|---|
| Plan A | 42% |
| Plan B | 25% |
| Plan C+ | 25% |
| Plan C | 20% |
| Plan D | 10% |
The biggest difference between the plans is time. A slow takeoff gives people years to study each new AI before the next one. A race gives them about a year.
How long the takeoff lasts, by plan
Years from an AI that fully automates coding to an AI that could take over.
| Plan | Years |
|---|---|
| Plan A | 6 |
| B kinetic | 3 |
| B cyber | 2 |
| C+ | 1.5 |
| C | 1.13 |
| D | 1.02 |
So is Plan A likely? The authors don't think so. Each of them estimated how likely each plan is to actually happen:
How likely each author thinks the deal and the race are
Each author's estimate that the world ends up on that plan.
| Author | Plan A, the deal | Plan D, the race |
|---|---|---|
| Daniel Kokotajlo | 15% | 30% |
| Romeo Dean | 8% | 30% |
| Eli Lifland | 5% | 25% |
| Ryan Greenblatt | 4% | 28% |
| Thomas Larsen | 3% | 50% |
And a deal can break. Eli estimates a 37% chance that Plan A's deal falls apart or weakens within five years, and 48% within ten. He still thinks a deal that breaks is better than no deal at all.
What I take from it
Read them as a pair. AI 2027 shows what happens if nobody changes course. AI 2040 shows what changing course could look like. Neither is a prediction you can set a calendar by, and the authors say so.
The dates are soft, the shape isn't. The team's own dates have moved by years, in both directions. What hasn't changed is the shape: once AI can do AI research, progress stops being steady and starts compounding.
Compute is the lever. Every serious idea in AI 2040 comes back to chips and datacenters, because they're the one part of AI you can count, track and inspect.
The first milestone is coding. In both stories, the first domino is AI that writes code as well as a strong engineer. If you build software, as I do, that's the milestone worth watching.
These are scenarios, not facts, and the numbers above are the authors' own rough estimates. If AI is part of your work, both are worth an evening.
Sources
- AI 2027, with the race ending and the slowdown ending, AI Futures Project, April 2025
- AI 2040: Plan A and its summary, AI Futures Project, July 2026
- Comparing possible plans and Deal decline, Eli Lifland, AI 2040 supplements
- Q1 2026 timelines update, AI Futures Project, April 2026