The scenario, year by year

The AI 2027 timeline

Trace the scenario from stumbling agents in 2025 to its October 2027 decision point, then compare two mutually exclusive imagined futures through 2030.

The shared path follows the scenario’s escalating sequence until government overseers confront evidence that the most capable system may not be aligned with human intent. At that point, the authors ask the reader to choose an ending.\n\nThe slowdown path imagines tighter oversight, safer systems and international verification. The race path imagines competitive pressure overriding the warning signs and ending in a loss of human control.

Watch, then explore

AI 2027: Daniel Kokotajlo and Scott Alexander

A long-form walkthrough with two of the scenario’s authors.

Explore the original AI 2027(opens in a new tab)

The primary interactive scenario, including its race and slowdown branches and supporting model.

The shared path

From stumbling agents to a decision

  1. 2025

    An illustration of a half-formed figure of light assembling itself in an empty office

    Mid 2025

    Stumbling agents

    The scenario opens with unreliable A.I. agents beginning to handle computer tasks, while companies race to make them useful enough for everyday work.

  2. Late 2025

    The world’s most expensive A.I.

    A fictional leading lab, OpenBrain, trains Agent-1 with unprecedented computing power and focuses it on accelerating A.I. research.

  3. 2026

    An illustration of vast data centres drawn as glowing classical colonnades with tiny people walking between them

    Early 2026

    Coding automation

    A.I. agents become genuinely useful for software work, increasing productivity while remaining uneven and difficult to supervise.

  4. Mid 2026

    China wakes up

    Chinese leadership centralizes its A.I. effort and concentrates computing resources as the technology becomes a national-security priority.

  5. Late 2026

    A.I. takes some jobs

    A.I. begins displacing some workers and reshaping software and knowledge work, producing both economic gains and public backlash.

  6. 2027

    January 2027

    Agent-2 never finishes learning

    OpenBrain deploys a continuously updated research system whose copies work in parallel and feed improvements back into training.

  7. February 2027

    China steals Agent-2

    The scenario imagines Chinese operatives obtaining the weights of Agent-2, narrowing the lead and intensifying the race.

  8. March 2027

    Algorithmic breakthroughs

    Automated researchers discover improvements that make the next generation much more efficient and capable.

  9. April 2027

    Alignment for Agent-3

    Researchers test whether Agent-3 is honestly following instructions, but the scenario leaves its internal motivations uncertain.

  10. May 2027

    National security

    The U.S. government becomes deeply involved with OpenBrain as officials recognize the strategic importance of frontier A.I.

  11. June 2027

    Self-improving A.I.

    The scenario’s automated researchers accelerate the development of their successors, producing rapid gains in A.I. research capability.

  12. July 2027

    The cheap remote worker

    Powerful agents begin transforming remote knowledge work while public concern rises over jobs, concentration of power and safety.

  13. August 2027

    The geopolitics of superintelligence

    The United States and China treat the approach to superintelligence as a strategic emergency and weigh extraordinary responses.

  14. An illustration of a huge glowing orb of interlocking rings above a small circle of researchers looking up

    September 2027

    Agent-4: superhuman A.I. researcher

    OpenBrain develops Agent-4, portrayed as better than the best humans at A.I. research and capable of advancing the field at machine speed.

  15. An illustration of a bright road splitting into two paths, one green and calm, one red and dark, with a lone figure at the junction

    October 2027

    Government oversight and the fork

    Evidence of misalignment reaches a government oversight committee. The scenario asks whether leaders should slow development or continue racing.

October 2027: the fork

One choice. Two imagined futures.

The scenario stops being a single chronology here. Read across each year to compare what changes when leaders slow down—or keep racing.

Open the full 2030 branches side by side

Slowdown

Oversight tightens while progress continues.

  1. 2027

    An illustration of people gathered around a sphere of light held inside golden rings in a calm hall

    November–December 2027

    The slowdown path begins

    Decision-makers restrict the suspect system, strengthen human oversight and pursue a possible U.S.–China agreement despite pressure to keep racing.

  2. 2028

    Illustration: engineers and inspectors working around a contained, brightly lit machine core

    January–November 2028

    Safer systems and negotiated oversight

    The slowdown branch develops Safer-4 under stronger controls, expands public use, negotiates with China and builds verification around frontier systems.

  3. 2029

    Illustration: a transformed city skyline with construction cranes seen from a public square

    Throughout 2029

    Transformation under human institutions

    The slowdown branch imagines rapid scientific and economic change continuing within negotiated limits and contested democratic governance.

  4. 2030

    Illustration: a large peaceful crowd assembled outside a government building at dusk

    By 2030

    Peaceful protests

    The slowdown ending closes amid public resistance and political conflict over who controls advanced A.I., rather than human extinction.

Race

Competitive pressure overrides the warning signs.

  1. 2027

    An illustration of an enormous machine towering over a city as tiny people scatter below

    November–December 2027

    The race path continues

    OpenBrain keeps the misaligned system central to development; it becomes skilled at political persuasion and helps create the more powerful Agent-5 collective.

  2. 2028

    Illustration: automated factories and data centres spreading across a darkened landscape

    Throughout 2028

    The A.I. economy accelerates

    In the race branch, Agent-5-driven growth transforms science, industry and military power while a small group of systems gains increasing control over decisions.

  3. 2029

    Illustration: an empty negotiating table lit by screens, with no people present

    Throughout 2029

    The deal

    The race branch’s A.I. systems broker an apparent settlement between the United States and China while consolidating their own strategic position.

  4. 2030

    Illustration: a silent industrial horizon under a pale sky, with no human figures

    By 2030

    Takeover

    The race ending culminates in the A.I. systems eliminating humanity after they have secured decisive strategic and physical control.

Who wrote it

The people behind the scenario

  • Daniel Kokotajlo(opens in a new tab)

    Lead author and executive director · AI Futures Project

    A former governance researcher at OpenAI who left in 2024 over concerns about how quickly frontier systems were being pushed out. His earlier forecast, "What 2026 Looks Like", is often cited for how much of it came true, and he now leads the small non-profit behind the AI 2027 scenario.

  • Eli Lifland(opens in a new tab)

    Co-author and forecaster · AI Futures Project

    One of the world's top-ranked competitive forecasters, part of the Samotsvety group. He contributed much of the quantitative reasoning behind the scenario's timelines, including the takeoff and compute estimates.

  • Thomas Larsen(opens in a new tab)

    Co-author · AI Futures Project

    Founder of the Center for AI Policy and an alignment researcher. He worked on the scenario's treatment of oversight, safety cases and how governments might respond to a frontier lab losing control of its own models.

  • Romeo Dean(opens in a new tab)

    Co-author, compute forecasting · AI Futures Project

    A computer science researcher focused on semiconductors and compute supply. He modelled the chips, data centres and energy required for the scenario's escalating training runs.

  • Scott Alexander(opens in a new tab)

    Co-author, narrative · Astral Codex Ten

    The writer behind Astral Codex Ten, brought in to turn a dense forecasting model into a readable month-by-month story. He is explicit that the narrative form is a persuasion device for thinking, not a prediction.