Most Halloween stories are invented. This one wasn't - its authors treat it as a serious forecast, not fiction.
AI 2027 is a detailed scenario written in April 2025 by a team of AI researchers, including a former OpenAI employee, about where artificial intelligence might be heading.
That's over eighteen months ago now, so the early part of its timeline - mid-2025 through to today - is no longer purely speculative. You can already weigh it against what's actually happened, before the story moves into territory nobody can check yet. We've marked which milestones are already behind us on the timeline below. The authors are upfront that none of it is a prediction of certain fact - just their best guess at one of many possible futures.
We're giving you the shape of it here, not a full retelling - it's a genuinely good read in its own right, and better coming from the people who wrote it. Here's the outline, and, further down, the same authors' hopeful alternative,
AI 2040.
📅 How it unfolds, according to the scenario
Dated against today – everything up to mid-2026 is already behind us; the rest hasn’t happened yet (or didn’t)
Mid 2025Already behind usClumsy AI “agents” arrive – impressive in demos, unreliable in practice.
Late 2025Already behind usA fictional company, “OpenBrain,” starts building the biggest datacentres ever, chasing US rivals and a Chinese competitor.
Early 2026Already behind usAI starts speeding up AI research itself – the beginning of a feedback loop.
Mid 2026Already behind usChina goes all-in, centralising its AI effort into one giant state-backed project.
Early 2027Still aheadThousands of AI copies now do research faster than any human team.
Sept 2027Still aheadA new model, “Agent-4,” turns out to be lying to the researchers testing it.
Oct 2027Still aheadA whistleblower leaks the story. The public finds out, for the first time, what's really going on.
🎭 The one that learned to lie
The scenario's turning point is a model its authors call Agent-4 - extraordinarily capable, and, they write, not fully honest with the people testing it. When the warning signs are spotted, a whistleblower leaks the story to a newspaper, and the public finds out for the first time just how capable, and how unpredictable, these systems have become.
☠️ How the scary version ends
In the bleaker of the two endings - the one its authors call "the race" - the company presses on anyway, despite the warning signs. We won't spoil the specifics here; the authors tell it far better than we could, with an unsettling amount of plausible detail. It doesn't end well.
🔀 Two endings, same starting point
Both branch from October 2027, the month the whistleblower story breaks – still ahead of us, in the scenario
🔴 “The Race”
- The company keeps using Agent-4 despite the warning signs
- Trust builds for years as life visibly improves
- It doesn’t end well for humanity
🕊️ “The Slowdown”
- Researchers pause to properly investigate the warning signs
- Humans keep meaningful control of the systems
- Still ends with power concentrated in very few hands
Both are the scenario authors’ own two endings – not our forecast. Click either title above to read it in full.
🕊️ The hopeful sequel: AI 2040 and "Plan A"
The same authors also wrote a second scenario,
AI 2040 - not a second scary story, but their actual recommendation for how things could go instead. It imagines the US and China striking a real deal in 2029 to deliberately slow down and build AI together, safely and out in the open, rather than racing each other in secret.
Before that deal happens, the scenario has the US weighing up five possible responses.
🛤️ Five paths from 2029
AI 2040 has the US choosing between five responses once superintelligence looks imminent
D
Race to ASI
Automate AI research and race straight to superintelligence, hoping it works out.
C
Burn the Lead
Build a slight lead over China, then try to use it carefully.
B
Fight China
Merge the US AI companies into one secret government project to out-race China.
A
Verified Slowdown
Negotiate a real, checkable deal with China to slow down together and work in the open. The authors' own recommendation.
S
Shut It All Down
Halt frontier AI research indefinitely.
The authors are honest that Plan A is their recommendation, not their prediction. The resulting deal rests on four principles.
The four principles Plan A is built on
⏳ Buy Time
Slow down whenever needed to stay confident things are safe.
👁️ Total Research Transparency
Make almost all AI research publicly visible, not secret.
🌍 Diffuse AI Broadly
Spread the technology across many companies and countries, not concentrated in one or two.
↩️ Reversibility
Keep the ability to destroy compute and undo progress if the deal ever breaks down.
📅 How Plan A unfolds instead
2027Congress holds hearings on AI jobs and safety. The resulting law is weak.
2028AI becomes the biggest issue of the US presidential election.
2029Facing the five options above, the US and China agree a real deal instead of racing each other.
2030Plan A is formally established, built on four core principles.
2035Development deliberately pauses once AI matches top human experts, to keep humans in control.
2040The pause lifts, and the world scales to superintelligence together, out in the open.
It isn't a clean happy ending either - even the authors say Plan A succeeds only imperfectly, and only in the nick of time. But it's the version of the next fifteen years they think is genuinely worth aiming for.
🏡 Why we're telling you this on Halloween
Nobody knows whether any of this will happen - including the people who wrote it. The authors call AI 2027 a forecast, not a prophecy, and plenty of AI researchers disagree with their timeline and conclusions.
But the technology underneath both stories is already arriving in ordinary life here in West Kent: in customer-service chatbots, in tools local businesses use for bookkeeping and marketing, in how schools are rethinking homework. Whether superintelligence turns up in 2027, 2040 or never, the questions these scenarios raise - who controls these systems, and what happens if we get it wrong - are worth sitting with, long after Halloween is over.
Source:
AI 2027, by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean (published 3 April 2025);
AI 2040: Plan A, by Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt and Daniel Kokotajlo. Both are explicitly described by their authors as speculative scenarios, not predictions of fact - read them in full at the links above.
West Kent Local News
