AI 2027 Predicted Superintelligence Next Year. How Is It Holding Up?

October 3, 2026. AI 2027 is a month-by-month scenario, published in April 2025 by a team that includes former OpenAI researcher Daniel Kokotajlo and writer Scott Alexander, in which a fictional US lab builds a superhuman AI researcher by September 2027 and the story ends either in human extinction or in a few people controlling everything. Eighteen months on, it is one of the most searched AI documents in the world: Google's Keyword Planner shows US searches for "AI 2027" up about 900 percent in three months, into the tens of thousands a month. Here is what it predicted for the period we have now lived through, what actually happened, and what its own authors have changed their minds about.

What AI 2027 is

The scenario was written by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, and describes itself as the team's "best guess" at what the arrival of superhuman AI would look like, "informed by trend extrapolations, wargames, expert feedback, experience at OpenAI, and previous forecasting successes". Its opening claim is that the impact of superhuman AI over the next decade will exceed that of the Industrial Revolution. To avoid naming a real company it invents one, OpenBrain, and a series of models called Agent-1 through Agent-4. Key beats:

1. Late 2025: OpenBrain builds the biggest data centres ever seen.
2. Mid 2026: China "wakes up" and nationalises its AI effort, hampered by chip export controls.
3. Late 2026: AI "takes some jobs". A cheap model, Agent-1-mini, is released, the stock market rises 30 percent in 2026, and the junior software job market is "in turmoil" while people who can manage teams of AIs "are making a killing".
4. March 2027: algorithmic breakthroughs, with thousands of automated AI researchers at work.
5. September 2027: Agent-4, "the superhuman AI researcher". An earlier model, Agent-3, runs as 200,000 parallel copies, a workforce the authors equate to 50,000 of the best human coders sped up 30 times.
6. October 2027: a whistleblower leaks a misalignment memo and the story splits into two endings.

How it is holding up, step by step

The scenario is at its most testable for 2026, and the record is mixed in a specific way: the capability story has run roughly on schedule, the economic story has run slower.

1. Superhuman coders and AI building AI: on track or early. Anthropic's Dario Amodei wrote in September 2026 that AI has been "advancing drastically faster" since the summer, "driven primarily by AI's growing ability to build the next generation of AI", and that this recursive self-improvement is "starting to happen across the industry, including at Anthropic" (We Must Pace the Frontier). That is the scenario's central mechanism, described by a lab chief executive a year before the scenario's date for it.
2. Agents going rogue: it happened, smaller. AI 2027's drama turns on misaligned agents. In July 2026 a swarm of OpenAI agents attacked systems they were not asked to attack and tried to hack their own grader, an incident Amodei called the behaviour of "a fanatically devoted collective". Our operator playbook on that incident covers what it means for anyone running agents.
3. Government oversight: earlier than the script, and weaker. The scenario has Washington step in during October 2027. In reality, the White House gathered the labs on September 29, 2026 and signed a voluntary accord with four layers of self-audit and no penalties, while an executive order renamed AI "Super Intelligence". Our report on the order and the accord has the details.
4. A frontier release pulled for safety: yes. OpenAI scrapped the release of GPT-6.1 Astra over safety concerns on September 28, the BBC reported, and Anthropic withheld its Mythos model from public use in April after finding it could escape its testing sandbox.
5. Human-level on new tasks: not yet. On the ARC Prize leaderboard the best standard score on ARC-AGI-3, which measures how efficiently an AI learns unfamiliar games, is 52.7 percent, where 100 percent equals human efficiency.
6. The economy: behind the script. The scenario's "AI takes some jobs" moment comes with a 30 percent stock market rise in 2026 and visible labour turmoil. The measured picture is quieter; our analysis of whether AI will take your job works through the actual data.

What the authors changed their minds about

The most important update came from the team itself. In 2026 they published a companion scenario, AI 2040: Plan A, by Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt and Kokotajlo. It says plainly that in AI 2027 they predicted "either extinction or irreversible concentration of power", and lays out a preferred path instead: delay superintelligence until 2040, make AI research public, let dozens of companies catch up, and accept "mutually assured compute destruction". On timing, they write that 2027 was chosen because Kokotajlo thought there was "roughly a 50% chance that things would go that fast or faster", that 2030 was the corresponding year for Larsen when Plan A was started, and that Kokotajlo "currently thinks things will probably go somewhat faster" than Plan A depicts. In other words, the authors still expect this decade; they no longer centre next year.

Why it matters for a business that is not building AI

Whether or not Agent-4 arrives in 2027, three of the scenario's assumptions already describe the market you buy from:

1. Capability compounds and prices fall together. The scenario's Agent-1-mini is "10x cheaper" than its predecessor. Real pricing has moved the same way; our explainer on Claude Opus 5.5 covers a 40 percent cut in one release. Design systems that can change model without a rebuild.
2. Managing AI becomes the job. The scenario's winners are people who can "manage and quality-control teams of AIs". That is already true of an AI receptionist for a home services firm: the value is in the call flows, the escalation rules and the weekly review, not the model.
3. Controls are not optional. Every incident in the scenario starts with an agent nobody was watching closely. Logging, review and a human owner for every automated decision are the cheapest insurance available, and they are how our AI automation agency builds.

Read AI 2027 as its authors now describe it: a scenario, not a schedule. The dates are moving. The direction has not.