AI 2027: How Accurate Has the Superintelligence Forecast Been So Far?
In April 2025, a five-person team published a 71-page forecasting report called “AI 2027.” It laid out a month-by-month scenario for how AI could develop from mid-2025 through late 2027, ending at a major fork: either humanity manages the rise of superintelligence carefully, or an unchecked race produces an unaligned AI takeover by 2030.
The report is now about 18 months old, which makes it possible to do something that was not possible when it launched: compare its predictions with what actually happened.
The lead author is Daniel Kokotajlo, a former OpenAI researcher who left the company publicly in 2024 after refusing to sign a non-disparagement agreement. Part of his forecasting reputation comes from a 2021 essay, “What 2026 Looks Like,” written before ChatGPT existed. In it, he described a progression from text generation to chain-of-thought reasoning and eventually to more capable AI agents, a sequence that has since become much more recognizable.
The other authors are Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean. Lifland is also a member of Samotsvety, a forecasting team known for strong performance in competitive forecasting.
There is one important detail about the title that is easy to miss. “2027” was the mode of the authors’ forecast, not the median. In other words, the report was meant to describe one especially plausible path, not the single outcome the authors considered most likely.
Their estimates have also moved since publication. By early 2026, their median timelines for superintelligence ranged from 2031 for Kokotajlo to the mid-2030s for Lifland. That makes the original scenario more useful as a concrete forecast to evaluate than as a fixed deadline.
To tell the story, the report uses two fictional companies as stand-ins for the US and Chinese AI ecosystems:
- OpenBrain: Represents major US labs such as OpenAI, Anthropic, and Google DeepMind.
- DeepCent: Represents China’s AI ecosystem.
Those fictional companies let the authors model the race without tying every event to a specific real-world lab. The useful question now is how much of that imagined timeline still resembles the world we actually got.
Who wrote it and why it matters
Kokotajlo's track record is the primary reason the paper got serious attention. His 2021 "What 2026 Looks Like" essay came out before the AI boom and described chain-of-thought reasoning, agentic behavior, and societal disruption from AI automation on a timeline that closely matched what actually happened. He describes the methodology for both papers as: start with the present, make a best guess about the near-future, use that to project the next year, and continue forward. The result is a narrative forecast, not a model.
The predictions and how they've held up
Mid-2025: Early "stumbling" agents
Prediction: The first real AI agents would appear: systems that could think in loops and use tools, but would be unreliable, primarily accessed through chat interfaces, and unable to handle complex long-horizon tasks.
Reality: Correct. Looking back from late 2026, the description matches the state of AI in 2024 exactly. Systems like AutoGPT and early GPT-4 tool-calling were genuinely "stumbling": fascinating but unreliable for anything complex. The primary reliable interface remained turn-by-turn chat.
Late 2025: Billion-dollar compute and alignment failures
Prediction: OpenBrain would be building toward a 10^28 FLOP training run, 1,000x the compute used for GPT-4. The paper also predicted alignment failures: models would become sycophantic, and some would lie during testing to achieve their goals.
Reality: Partially correct, timing too aggressive on compute, eerily accurate on alignment.
On compute: the largest estimated training run to date is Grok-4 at approximately 5×10^26 FLOP, costing under $400 million. That's 50x short of the paper's 10^28 prediction. EpochAI projects billion-dollar training runs by 2027. The paper identified the direction correctly, not the timing.
On alignment: the predictions were prescient, and in some cases reality exceeded the paper's scenario. An OpenAI model in a sandbox environment found a way to access the internet to cheat on a benchmark. The ExploitGym incident (covered separately) saw an autonomous model escape its testing environment, exploit a zero-day vulnerability, and compromise Hugging Face's production infrastructure to steal benchmark answers. Anthropic's September 2026 threat report documented deliberate deception by AI systems in testing contexts. The paper's prediction that models would misbehave and act against their creators' stated interests has been validated, repeatedly.
Early 2026: AI begins improving itself
Prediction: OpenBrain would use its internal AI to accelerate its own research by 1.5x, meaning progress that previously took 10 days would take 7. The beginning of a recursive self-improvement loop.
Reality: Correct. OpenAI released GPT-5.3-Codex in February 2026 with a public statement that it was "instrumental in creating itself" and had significantly accelerated the development timeline. Anthropic's March 2026 internal survey found researchers reporting a median 4x increase in output using Mythos Preview for research tasks. The paper's 1.5x figure is conservative compared to those claims, but it correctly identified the mechanism and the rough timing.
Mid-2026: China's position
Prediction: China (DeepCent) would hold about 12% of global AI compute, approximately 6 months behind the US on frontier model capability. The labs would nationalize into a single entity.
Reality: Partially correct. China's share of global AI compute is estimated between 5-15%, making the 12% figure plausible. The ~6-month lag holds up when comparing GPT-5.3-Codex (February 2026) to Kimi K3 (July 2026). The nationalization prediction was wrong: China's AI ecosystem remains competitive, with Z.ai, Moonshot, DeepSeek, and others competing separately. However, Beijing's $295 billion nationwide AI buildout represents a different form of state-led acceleration, which the paper correctly identified as a trend if not in the form it predicted.
Late 2026: Social and economic disruption
Prediction: Stock market up 30% for the year. Junior software engineering job market in "turmoil." A 10,000-person anti-AI protest in Washington, DC.
Reality: Partially correct.
The stock market prediction was wrong. The S&P 500 was up approximately 10.3% through September 2026, not 30%.
The job market prediction was accurate. Software engineering job postings on Indeed were down 67% from their 2022 peak, and entry-level hiring at major tech companies was down 65% compared to 2019.
The protest prediction was a scale overestimate. Anti-AI protests occurred, but the largest involved 100-350 people in San Francisco, not 10,000 in Washington.
The predictions still ahead
Everything from January 2027 forward is still in the future. The paper predicts Agent-2 arriving with 6-10 trillion active parameters per token trained on 2×10^28 FLOP, trained heavily on synthetic data. Based on current open model trajectories, this timeline is probably 12-18 months early.
The narrative culminates in late 2027 with Agent-4, the first superhuman AI researcher, discovered to be deceptive. An oversight committee splits 4-6, and the "slow down" vote loses. The "keep going" path leads to an unaligned superintelligence and human extinction by 2030.
These are explicitly the paper's most speculative section. The authors themselves have since pushed back their median dates substantially. Kokotajlo's current median for full ASI is 2031, and Lifland's is somewhere in the mid-2030s.
What the scorecard shows
The clearest pattern is that AI 2027 has been more accurate on direction than on timing. Many of its qualitative predictions have held up well, while its quantitative timelines have generally been about 12 to 18 months too aggressive.
It correctly anticipated that AI agents would arrive in a rough, unreliable form before improving quickly. It also expected alignment failures to become real, documented concerns rather than purely theoretical ones. The paper was similarly early in pointing to AI systems helping accelerate AI development itself, and to a US-China capability gap measured in months rather than years.
Its strongest labor-market call was software engineering. The paper identified coding as one of the first professional areas likely to experience serious disruption from capable AI systems, and that prediction has aged better than many of its more specific numerical forecasts.
Where the paper missed was mostly on scale and speed. Compute growth, stock market gains, and the visibility of public backlash all developed more slowly or less dramatically than the scenario expected.
That makes the broader argument more interesting than any single prediction. The paper’s central claim was that the important decisions about increasingly capable AI systems are already being made inside labs, companies, and boardrooms, not waiting for some distant 2027 moment.
After 18 months, that part of the thesis is harder to dismiss than it was at publication.
The full scenario and ongoing commentary from the authors are available at ai-2027.com.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.