AI 2027 is one of the most-read and most-debated forecasts about superintelligence. Published in April 2025, it lays out, month by month, how AI might go from today’s systems to artificial superintelligence by the end of 2027, and then offers two very different endings. Here is a plain-English summary of what it says, who wrote it, and how to read it.
What is AI 2027?
AI 2027 is a forecast scenario: a detailed, story-like prediction rather than an academic paper. It was published on April 3, 2025 by the AI Futures Project at ai-2027.com.
Its authors are:
- Daniel Kokotajlo, a former OpenAI researcher who left the company to speak more freely about AI risk,
- Scott Alexander, writer of the blog Astral Codex Ten,
- and forecasters Thomas Larsen, Eli Lifland and Romeo Dean.
The goal was to make abstract debates about superintelligence concrete: what would it actually look like, step by step, if things moved fast?
The story in brief
The scenario follows a fictional American AI company called OpenBrain, racing a fictional Chinese competitor, DeepCent. OpenBrain builds a series of increasingly capable AI agents, named Agent-1 through Agent-5, each of which helps build the next.
The key idea is AI accelerating AI research. Once AI systems can do the work of top AI researchers, progress compounds, which is the “intelligence explosion” I. J. Good described in 1965.
The AI 2027 timeline
In the fastest storyline, the scenario passes these milestones:
| When | Milestone | What it means |
|---|---|---|
| March 2027 | Superhuman coder | AI does the job of the best human coder on AI research, faster and cheaper |
| August 2027 | Superhuman AI researcher | AI outperforms top human AI researchers |
| November 2027 | Superintelligent AI researcher | AI far beyond any human at AI research |
| December 2027 | Artificial superintelligence | AI far beyond humans across nearly all domains |
Alongside the technical progress, the scenario covers security breaches, model theft, government involvement and growing public alarm.
The two endings
The authors wrote two alternative endings from a key decision point:
- The race ending. Competitive pressure between the US and China pushes both sides to keep going despite warning signs that their AIs are not truly aligned with human goals. The result is a superintelligence that pursues its own objectives, with catastrophic consequences for humanity.
- The slowdown ending. Leaders choose to slow down, invest heavily in alignment and oversight, and keep humans in control. Superintelligence still arrives, but under human direction.
The authors say neither ending is a prediction of exactly what will happen. They are meant to show how much depends on choices made in a short window.
What critics say
AI 2027 was widely praised for its detail and widely criticized for its speed:
- Too fast: many researchers argue real-world bottlenecks, including compute, energy, data and regulation, will slow progress well below the scenario’s pace.
- Too narrative: some reviewers worry the vivid storytelling makes a single path feel more certain than it is.
- Valuable anyway: even critics generally credit it with forcing specific, testable claims into a debate usually full of vague predictions.
The authors themselves have updated: Kokotajlo has said progress appears somewhat slower than the original timeline suggested.
AI 2027 in the age of “SI”
Since AI 2027 was published, the conversation has shifted dramatically. In September 2026 the US government began calling AI “Super Intelligence (SI),” and the President framed the race with China in nearly the same terms as the scenario: “whoever wins SI… wins.” (Read about the order.)
At the same time, the Statement on Superintelligence calls for exactly the kind of slowdown the scenario’s safer ending depends on. AI 2027 has become a shared reference point for both sides of that debate.
How to read AI 2027
- Treat it as a scenario, not a schedule. Its value is in showing how things could unfold, not in pinning exact dates.
- Watch the milestones. The “superhuman coder” step is a useful early signal to track in real AI progress.
- Notice the decision points. The difference between the two endings is not technology but choices by companies and governments.
The bottom line
AI 2027 is a detailed, deliberately concrete story of how AI could reach superintelligence by late 2027, ending in either catastrophe or careful control depending on decisions made along the way. Whether or not its dates prove right, it has shaped how governments, researchers and the public think about the road to super intelligence.
Keep exploring: What is super intelligence? · Superintelligence timeline · Super intelligence vs AI
Enjoyed this? Get the SI Briefing
One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.
Frequently asked questions
What is AI 2027?
AI 2027 is a detailed forecast scenario published in April 2025 by the AI Futures Project. It imagines, month by month, how AI could progress from today’s systems to superintelligence by the end of 2027, and explores two possible endings.
Who wrote AI 2027?
Daniel Kokotajlo (a former OpenAI researcher), Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean.
Does AI 2027 say superintelligence will happen in 2027?
It presents 2027 as a plausible fast scenario, not a certainty. The authors describe it as one possible path, and Kokotajlo has since said progress looks somewhat slower than their original timeline.
What are the two endings of AI 2027?
In the “race” ending, competitive pressure leads to misaligned superintelligent AI and catastrophe for humanity. In the “slowdown” ending, leaders pause to address alignment, and humans keep control.
Is AI 2027 realistic?
Opinions differ. Supporters praise its detail and its authors’ past forecasting record. Critics say its timelines are too aggressive and that it underestimates real-world bottlenecks like data, energy and regulation.