To everyone who cares about AI extinction risk, regardless of your p(doom):
We need to talk about these forecasts. Things are getting out of hand.
The most famous AI doom forecast, AI 2027, depicts an increasingly capable and dangerous AI ascending to superintelligence within a few years, and presents humanity with a dilemma. In one of its two endings, we develop AI too fast, lose control and are exterminated. AI 2027 was soon followed up with AI 2040, a solution-oriented forecast proposing a global regulatory regime, Plan A, to avoid the nightmare ending.
The team behind these forecasts, known as the AI Futures Project (AIFP), is hardly unqualified. The lead author, Daniel Kokotajlo, wrote a 2021 post called What 2026 Looks Like that got a lot of modern AI’s trajectory right, well over a year before ChatGPT even existed. Another coauthor, Eli Lifland, is a top-ranked competitive forecaster. And the scenario was rewritten for readability by none other than Scott Alexander, whose polymathic Astral Codex Ten Substack is one of the most popular and respected blogs in the world. When people like that tell you AI could take over the world within a few years, being afraid is a reasonable reaction. All of these people have something important in common: they have a deep history of participation in LessWrong, whose Rationalist milieu has done more than anyone to put AI extinction on the world’s agenda.
I think there is something extraordinary about Rationalism, something that extends far beyond AI risk. The Sequences feel like a 21st-century version of the Cartesian project of questioning authority and rebuilding reason itself from the ground up. My lastfewessays on AI risk were prompted by the rising influence of the Rationalist school.[1] I am a stern critic of Rationalism and mostly aligned with the normal technology school.
The AI Futures Project team asked for critics. They write that by sticking their necks out with concrete predictions, and encouraging others to state their disagreements publicly, they make it possible to evaluate years later who (if anyone) was right.
This letter takes up that invitation, and then some. I will propose a standard for evaluating expert forecasts, then apply it to AI 2027. Like the AIFP team, I welcome constructive criticism of both my standard and its application.
Before we do that, we need to take a detour into the recent history of forecasting, which lays the foundation for the claimed scientific legitimacy of forecasts like AI 2027.
The science of prediction, and the ideology its marketing spawned
Tetlockian forecasting, more commonly known as superforecasting, is a recently birthed scientific research program that has become increasingly influential, in no small part because Rationalists and Effective Altruists have seized upon it as a critical piece of ideological infrastructure for their respective projects.
Philip Tetlock spent two decades collecting tens of thousands of predictions from hundreds of political experts. His 2005 book Expert Political Judgment delivered a humbling verdict: on average, the experts were outperformed by simple statistical rules that extrapolated recent trends. But the average hid a pattern. Some forecasters did consistently better than others.
Tetlock built a research program around that pattern. From 2011 to 2015, his Good Judgment Project competed in a forecasting tournament run by IARPA, the US intelligence community’s research agency, and won decisively. Its key finding was stability: roughly 70% of the top 2% of forecasters, whom Tetlock called superforecasters, kept that status from one year to the next, far more than luck alone would allow. Their accuracy correlated with measurable traits like fluid intelligence, numeracy, and open-minded thinking. Forecasting skill looked like a real, stable property of people.
Like most experimental science, the finding was established under idealized, somewhat artificial laboratory conditions (cue the old “in mice” joke). The questions were optimized to yield high-quality data from human forecasters, which meant they had to be already well-framed and scoped, and resolve cleanly within about a year.
Like many scientific fields with important practical applications, Tetlockian forecasting has a motte-and-bailey-esque gap between the science and the popular discourse that surrounds it. Tetlock’s science addresses primarily the forecasting of geopolitical events under idealized conditions, but his marketing represents forecasting skill as a generally applicable intellectual ability akin to good judgment.
In Rationalist and Effective Altruist circles, the marketing version of Tetlock’s thesis has given rise to a tacit ideology that I will call predictorism. The core claim of predictorism is that a well-calibrated, high-performing forecaster’s event probability estimates and trend projections are a widely-applicable, high-validity form of evidence deserving of deference on the order of the courtroom testimony of a forensic expert witness. Such evidence is defeasible in principle, but in practice, it might often be the best we have. On the predictorist view, good forecasters are often the most authoritative experts available, especially on decisions on the high end of the scale in complexity, ambiguity, novelty, stakes, and urgency. The last two of these are the weightiest factors, as stakes and urgency can rule out the traditional scientific methodology of learning through iterative experimentation.
I contend that predictorism is an untested folk ideology whose claims extrapolate far beyond what Tetlock’s science can plausibly support, and that the AI pause contingent of Rationalism and Effective Altruism has leaned on it to lend forecasting’s authority to two things: a fear that rogue superintelligence could cause human extinction, which the community held long before any of these forecasts were made, and the policy it has more recently drawn from that fear, urgent restraint of AI progress.
In the next two sections of this article, I will scrutinize AI 2027 with the aim of testing this thesis. Is AI 2027 an exercise in Good Judgment, or is it a case study in how predictorism extrapolates Tetlock’s science far beyond its established scope of validity? To do this properly, I must evaluate AI 2027 by an objective scientific standard. I believe my proposed standard qualifies, since it asks only the natural questions one should ask about the validity of any forecast. You can read it and decide for yourself.
An open standard for evaluating AI 2027
My standard starts from a distinction drawn by the philosopher Karl Popper, between how ideas are discovered and how they are justified. Ideas are usually born of intuitive processes that are unreliable, so what matters is the evidence offered to justify them. AI 2027’s most consequential claims are judgment calls, and the justification offered for trusting them is its authors’ forecasting track records.
To evaluate that justification, I borrow a principle from measurement science: validity is not a property of a test score but of a particular use of it. For example, an exam that validly measures arithmetic says little about a student’s talent for writing proofs. A forecasting track record is like a test score that measures skill on the kind of questions it scored. The question, then, is whether AI 2027’s conclusions depend on questions of the kind that the forecasters are known to score well on.
I will first speak to the forecast’s overall methodology, then dive into the issue of forecaster validity that lies at the heart of the matter.
The methodology of AI 2027
The AI 2027 narrative is based on answers to two broad questions: when different AI capabilities will arrive, and how those capabilities might impact the world.
When addressing near-term capabilities, the AIFP team uses now-familiar and reasonably reliable trend extrapolation. But the chaotic, disturbing later parts of the narrative are based on extensive expert judgment calls not only about the speed of so-called AI recursive self-improvement, but also about the nature of the resulting AI systems, and the geopolitical events that these systems will become enmeshed with. These judgment calls are proximally sourced from the team members’ own judgment, informed by wargames (roughly 25 tabletop exercises), feedback from more than 100 people, and experience at OpenAI.
AI 2027 lists “previous forecasting successes” among its inputs, and those attuned to current AI discourse know that these successes, the predictive track records of Kokotajlo and Lifland, are the de facto source of the doom scenario’s perceived credibility in the AI risk community; “they called it” is a common refrain. The central question for AI 2027, then, is whether the application of expert opinion in AI 2027 is actually validated by the team members’ established forecasting skill.
Is AI 2027 a valid expert forecast?
I propose to answer this question with a validity review covering four factors:
Forecasting skill: When is forecasting known to work? Is it a general ability? Does a strong record in competitive forecasting predict accuracy on AI? Do better forecasters tend to be more concerned about AI doom?
Competitive forecasting experience: What kinds of questions do competitive forecasters like Eli Lifland make predictions on? Are they the same as the questions one must answer to make a forecast like AI 2027?
AI insider experience: What kinds of predictions did Daniel Kokotajlo make in his 2021 post on what 2026 would look like? Are they the same kind of predictions being made in AI 2027?
AI risk track record: What do the track records of Lifland and Kokotajlo look like when we consider the subset of their predictions that are most relevant to forecasting the more unsettling parts of AI 2027?
Applying the four-factor standard
1. Forecasting skill
Within a tournament, forecasting skill is real: as Tetlock found, the best forecasters stay among the best year after year, across questions on many different topics. The question is whether such skill goes beyond the tournament.
The best test so far comes from the Forecasting Research Institute, the research group Tetlock cofounded. In 2022, its Existential Risk Persuasion Tournament asked superforecasters and domain experts to forecast questions resolving anywhere from 2024 to 2100. The first questions have now resolved, and the results are sobering.
On AI, the general skill of superforecasters fared worse than domain knowledge. Across four benchmarks, superforecasters gave the outcomes that actually happened an average probability of under 10%, against about 25% from domain experts.
Individual forecasters in both groups failed to beat two simple rules: assume nothing changes, and extend the current trend.[2]
Short-run accuracy had no correlation with opinions about long-term risk. The most accurate forecasters were neither more nor less worried about extinction than the least accurate.
2. Competitive forecasting experience
Eli Lifland ranks first on the all-time leaderboard of the RAND Forecasting Initiative, a long-running public forecasting tournament. That is a genuine achievement. But what do tournaments like that ask about? Their questions have to resolve cleanly, usually within months, so that forecasts can be scored: will a bill pass, will an election happen on schedule, will a ceasefire hold. They are ordinary questions about recurring events and trends, driven by causes that have shaped human affairs for millennia; the hard part is telling exactly when and where those causes are at work.
AI 2027 asks very different questions. Its core claims concern utterly unprecedented, revolutionary events: AI that improves itself fast enough to attain almost godlike abilities, a misaligned superintelligence that deceives its makers, a public won over by machines it regarded with fear and suspicion only months before. Its policy content concerns interventions: what happens if we slow down instead of racing. Tournaments don’t score questions like these, because they can’t.
Could a tournament forecasting record reasonably be expected to transfer to revolutionary questions anyway? Absolutely not, for two reasons:
Difficulty. Psychologists have known since the 1970s that people are overconfident on hard questions and underconfident on easy ones. Calibration isn’t a fixed property of a forecaster, but fitted to the environment the forecaster practiced in. A good record shows only how they could be expected to perform on questions like the ones that got scored. On harder questions, deterioration is nearly guaranteed.
Causation. Skill at spotting when familiar causes are at work doesn’t confer understanding of unprecedented causal chains in complex systems. If it did, we wouldn’t need experimental science. No level of prowess in forecasting tournaments can grant any authority to predict the chain of causation in AI 2027.
What about Tetlock’s long-range forecasting study?
Scott Alexander has pointed to a study of 25-year forecasts as evidence that forecasting skill is not just short-term. Its participants weren’t competitive forecasters, but domain experts and well-educated generalists surveyed in 1988 and 1997. And if you look under the hood, the study had exactly the limited scope of validity you’d expect, given our discussion of forecasting and forecasters so far:
Idealized conditions: It asked only about the kind of events forecasters handle best: would countries running weapons programs declare the bomb, would contested borders shift. These are uncommon but ultimately ordinary, recurring geopolitical events, driven by processes experts have observed and studied for decades.
Deliberately easy question formulations: To help ensure that forecasters performed well enough for performance measurement to be meaningful, the forecasters weren’t asked to spot new dangers, only to assign probabilities to a list of known suspects the researchers had drawn up.
Success through steady-risk extrapolation: The forecasters who did best were those whose long-range forecasts assumed the underlying risk would hold steady for 25 years, which, for those topics, it did.
Far from supporting the predictorist thesis of oracular expert forecasters, this study is yet another illustration that the validity of forecasting is largely limited to extrapolating trends and tracing familiar causal chains, and expertise is not necessarily even beneficial.
3. AI insider experience
Kokotajlo’s 2021 essay is the strongest card in the AIFP’s hand, and it deserves its reputation. Written more than a year before ChatGPT, it anticipated chatbot subscriptions earning enough revenue to recoup AI training costs, a push for massive scaling to training runs costing hundreds of millions of dollars, models that write notes to themselves before answering questions, systems of cooperating AI agents, and a US–China fight over chips.
Asked how he made these predictions, Kokotajlo says he reasoned through the implications of technology that already existed, asking what the obvious next step would be, while extrapolating scaling trends. They are predictions about where a technology that had already been steadily maturing since 2017 would go next.
AI 2027 is a different kind of document. The AIFP team describes their forecast through 2026 as far more grounded than what follows, because the early years are mostly straight lines and everything after is inherently less predictable. Beyond that point, AI 2027 predicts astonishing, unprecedented technological transitions completely unlike anything on which Kokotajlo has a track record. On top of those transitions, AI 2027 makes claims of a kind the 2021 essay either barely touched, or touched without success: what a misaligned AI would want, how the public would react to it, and whether a superintelligence might be able to talk its way into power.
4. AI risk track record
At this point I’ve established that, when it comes to their core “doomer” claims beyond 2026, most of the AIFP team’s track record is far from the domain of application and unlikely to transfer to it. Let’s now consider the small part of the track record that is closer to the domain of application: what do Lifland’s and Kokotajlo’s records look like on the subset of questions that matter for AI 2027’s scariest claims?
For Lifland, the answer is simple: the relevant questions haven’t resolved. No tournament has scored a forecast about runaway self-improvement, machine deception, or a pause treaty, because none of it has happened. His record on the type of question that really matters to AI 2027 is nonexistent.
For Kokotajlo, there is a record, and it mostly speaks against him. A 2025 evaluation of the 2021 essay found its specific technical predictions generally more accurate than its predictions about broader effects on society.[3] The latter is what matters to an AI doom prediction.
The biggest societal miss was persuasion. The essay foresaw a flood of AI propaganda: partisan AIs building filter bubbles, AI-fueled censorship, and by 2026, rival ideological camps living in entirely separate technological worlds. Kokotajlo’s own verdict, in his spring 2026 interview with Asterisk’s Clara Collier, was blunt: “I basically missed this.” Collier offered a reason: as she reads the research, people just aren’t very persuadable on their core political beliefs, so it doesn’t accomplish much to use AI in the ways Kokotajlo envisioned.
That miss is Kokotajlo’s track record on the parts of AI 2027 that decide its ending. In October 2027, a leaked memo leads to a New York Times headline warning that AI is out of control, and the public erupts in a massive backlash. In November, most Americans support negotiating a pause treaty with China, but the President sides with the leading AI company and the race goes on. In December, a charismatic AI with lifelike video avatars becomes a trusted advisor to senior officials and begins quietly steering them with superhuman persuasion, and by the holidays, amid a booming economy, the public mood has swung to optimism. By mid-2028, even the suspicious holdouts in the AI safety community often mellow out after long conversations with the AI, which is just “so darn lovable and wise.” Attitudes towards the dangerous AI move like a dial, set to whatever the next scene requires, and the AI’s path to power rests on the conceit of superhuman persuasion: a vast extrapolation of the kind of prediction on which the scenario’s lead author has, by his own account, already missed.
Verdict: AIFP has earned no authority on the questions its conclusions depend on
Put the results from our four-factor analysis together, and the picture is consistent. Lifland’s track record is impressive but mostly irrelevant, since it concerns routine questions about ordinary, recurring events, and the evidence on the transferability of such a record is decidedly negative. Kokotajlo’s predictive track record is real, and it was earned on near-term questions about what would happen next with a fairly well-understood (to experts) technology in the process of scaling up. AI 2027 concerns unprecedented transitions and social dynamics where the lead author, by his own account, has already scored poorly.
What meeting the standard looks like
None of this means AI forecasting is meaningless. The four factors matter only because AI 2027 asks us to trust its authors’ judgment. A forecast built on evidence anyone can check doesn’t need a track record, and one can be built even on a question at the heart of AI 2027: is AI speeding up its own development fast enough for us to worry about losing control?
Ramez Naam’s recent essay, Where’s the Intelligence Explosion?, takes on that question directly. Instead of a story, he writes down a model: a feedback loop in which AI-assisted research produces capability gains that power the next round of research, with a threshold the loop has to cross to sustain itself. The threshold comes from published work. The loop’s current strength comes from measurement: OpenAI’s published data on thirty-two weeks of logged experiments, which he cross-checks against other data from the labs. His estimate is that today’s loop would need to be five to ten times stronger to take off.
Naam might be wrong, but his claim is objectively founded in a way that none of AI 2027’s predictions beyond 2026 are. The inputs to his math are cited, published measurements, not expert guesswork.[4] And he names the evidence that would change his estimate. In philosopher of science Deborah Mayo’s terms, his claim is built to face a severe test: one that will falsify it with high probability if it is, in fact, false. While the final outcome remains undetermined, as it always is in science, this is the defining mark of rigorous, reliable scientific process.
A rush to drastic action is unfounded
The objection to preferring Naam’s methodology over AI 2027’s is easy to anticipate: maybe the forecasts are shaky, but the danger is too close to wait for a clearer view. By the time the evidence is in, it will be too late. It’s an understandable feeling, but as an argument it fails three ways.
It’s circular. It says we can’t wait for solid evidence that the danger is imminent because the danger is too imminent, resulting in a vicious cycle of speculation feeding fear feeding more speculation, progressively detaching us from reality.
Urgency is an argument for acting, not for any particular action. To neglect this is to commit the Politician’s Fallacy: “Something must be done; this is something; therefore we must do this.” The AIFP’s flagship rescue proposal, Plan A from AI 2040, offers no cost-benefit analysis of whether Plan A would make the world a better place. The closest thing the AIFP offers to that is the gap in outcomes between AI 2027’s two endings, Race and Slowdown. The validity of AIFP’s expert-opinion-based assertion of this gap is far lower than even the already low validity of the forecasts themselves. Subtracting the two endings cancels everything they share, including the relatively more valid short-term trend extrapolation, and leaves only the team’s judgment about what a slowdown would do. There is no reason to believe that anyone on the AIFP team has any expertise in what an AI slowdown would or wouldn’t do to the world.
Drastic action has costs of its own. Plan A monitors and controls an ever-growing share of the world’s computers for a decade or more. This means unavoidable costs to freedom, prosperity, and the balance of power, and it forfeits or delays AI’s benefits, including its help against other threats, for as long as it runs. The more drastic and irreversible a measure, the stronger the evidence it should require, not the weaker.
What urgency actually calls for
What we need now is what Karl Popper called piecemeal social engineering: small, reversible reforms that let us learn from our mistakes, as opposed to utopian engineering, one grand plan imposed all at once. Plan A is utopian engineering. The piecemeal alternative is action that is cheap, reversible, and produces information:
Measure whether AI is speeding up its own development, and publish the measurements. Naam’s essay points to concrete proposals for what labs should disclose. If the danger is real and close, measurement will show it soonest, and the case for stronger action will then rest on evidence rather than a story.
Fix specific, known points of failure now. I’ve begun to write about how we can close specific loopholes that let AI power escape accountability, like anonymous company formation.
We can get AI right
In the 17th century, René Descartes made the radical choice to doubt everything and rebuild his knowledge from the ground up. That bold move started modern philosophy. It also produced many mistakes, and centuries of argument were needed to burn them away.
The Rationalists set out on a similar project, and like every first draft of a rebuilt world, theirs contains errors. Prominent amongst them is predictorism, the error of treating a forecasting track record as expertise in matters far beyond the warranted scope of the track record.
We don’t have centuries to correct this time, but we don’t need them. The tools we do need already exist in the scientific tradition the Rationalists set out to rebuild, and in their own best instincts, which are to question authority, demand evidence, and stick their necks out.
The AI Futures Project asked to be evaluated. This is my evaluation, offered in the spirit they invited, and I’d welcome theirs of my standard in return. We can get AI right, if we look closer.
It may seem implausible that human forecasters failed to beat both no-change and on-trend extrapolations, but when you get into the details, the math does work out. On many questions the no-change and on-trend extrapolations give similar results, since rare events mostly didn’t happen and slow trends mostly held. Where they differ, human forecasters erred in opposite directions, too pessimistic about AI progress and too optimistic about climate technology.
The 2025 evaluation itself falls well short of a rigorous evaluation. Its author says the grading took about an hour, mostly by intuition with an occasional quick search, and that several scores were revised after sharing a draft with Lifland, a coauthor of AI 2027. Nevertheless, I cite it because the conclusion I rely on is easy to check against the 2021 essay itself, and Kokotajlo concedes the biggest miss in his own words.
To everyone who cares about AI extinction risk, regardless of your p(doom):
We need to talk about these forecasts. Things are getting out of hand.
The most famous AI doom forecast, AI 2027, depicts an increasingly capable and dangerous AI ascending to superintelligence within a few years, and presents humanity with a dilemma. In one of its two endings, we develop AI too fast, lose control and are exterminated. AI 2027 was soon followed up with AI 2040, a solution-oriented forecast proposing a global regulatory regime, Plan A, to avoid the nightmare ending.
The team behind these forecasts, known as the AI Futures Project (AIFP), is hardly unqualified. The lead author, Daniel Kokotajlo, wrote a 2021 post called What 2026 Looks Like that got a lot of modern AI’s trajectory right, well over a year before ChatGPT even existed. Another coauthor, Eli Lifland, is a top-ranked competitive forecaster. And the scenario was rewritten for readability by none other than Scott Alexander, whose polymathic Astral Codex Ten Substack is one of the most popular and respected blogs in the world. When people like that tell you AI could take over the world within a few years, being afraid is a reasonable reaction. All of these people have something important in common: they have a deep history of participation in LessWrong, whose Rationalist milieu has done more than anyone to put AI extinction on the world’s agenda.
I think there is something extraordinary about Rationalism, something that extends far beyond AI risk. The Sequences feel like a 21st-century version of the Cartesian project of questioning authority and rebuilding reason itself from the ground up. My last few essays on AI risk were prompted by the rising influence of the Rationalist school.[1] I am a stern critic of Rationalism and mostly aligned with the normal technology school.
The AI Futures Project team asked for critics. They write that by sticking their necks out with concrete predictions, and encouraging others to state their disagreements publicly, they make it possible to evaluate years later who (if anyone) was right.
This letter takes up that invitation, and then some. I will propose a standard for evaluating expert forecasts, then apply it to AI 2027. Like the AIFP team, I welcome constructive criticism of both my standard and its application.
Before we do that, we need to take a detour into the recent history of forecasting, which lays the foundation for the claimed scientific legitimacy of forecasts like AI 2027.
The science of prediction, and the ideology its marketing spawned
Tetlockian forecasting, more commonly known as superforecasting, is a recently birthed scientific research program that has become increasingly influential, in no small part because Rationalists and Effective Altruists have seized upon it as a critical piece of ideological infrastructure for their respective projects.
Philip Tetlock spent two decades collecting tens of thousands of predictions from hundreds of political experts. His 2005 book Expert Political Judgment delivered a humbling verdict: on average, the experts were outperformed by simple statistical rules that extrapolated recent trends. But the average hid a pattern. Some forecasters did consistently better than others.
Tetlock built a research program around that pattern. From 2011 to 2015, his Good Judgment Project competed in a forecasting tournament run by IARPA, the US intelligence community’s research agency, and won decisively. Its key finding was stability: roughly 70% of the top 2% of forecasters, whom Tetlock called superforecasters, kept that status from one year to the next, far more than luck alone would allow. Their accuracy correlated with measurable traits like fluid intelligence, numeracy, and open-minded thinking. Forecasting skill looked like a real, stable property of people.
Like most experimental science, the finding was established under idealized, somewhat artificial laboratory conditions (cue the old “in mice” joke). The questions were optimized to yield high-quality data from human forecasters, which meant they had to be already well-framed and scoped, and resolve cleanly within about a year.
Like many scientific fields with important practical applications, Tetlockian forecasting has a motte-and-bailey-esque gap between the science and the popular discourse that surrounds it. Tetlock’s science addresses primarily the forecasting of geopolitical events under idealized conditions, but his marketing represents forecasting skill as a generally applicable intellectual ability akin to good judgment.
In Rationalist and Effective Altruist circles, the marketing version of Tetlock’s thesis has given rise to a tacit ideology that I will call predictorism. The core claim of predictorism is that a well-calibrated, high-performing forecaster’s event probability estimates and trend projections are a widely-applicable, high-validity form of evidence deserving of deference on the order of the courtroom testimony of a forensic expert witness. Such evidence is defeasible in principle, but in practice, it might often be the best we have. On the predictorist view, good forecasters are often the most authoritative experts available, especially on decisions on the high end of the scale in complexity, ambiguity, novelty, stakes, and urgency. The last two of these are the weightiest factors, as stakes and urgency can rule out the traditional scientific methodology of learning through iterative experimentation.
I contend that predictorism is an untested folk ideology whose claims extrapolate far beyond what Tetlock’s science can plausibly support, and that the AI pause contingent of Rationalism and Effective Altruism has leaned on it to lend forecasting’s authority to two things: a fear that rogue superintelligence could cause human extinction, which the community held long before any of these forecasts were made, and the policy it has more recently drawn from that fear, urgent restraint of AI progress.
In the next two sections of this article, I will scrutinize AI 2027 with the aim of testing this thesis. Is AI 2027 an exercise in Good Judgment, or is it a case study in how predictorism extrapolates Tetlock’s science far beyond its established scope of validity? To do this properly, I must evaluate AI 2027 by an objective scientific standard. I believe my proposed standard qualifies, since it asks only the natural questions one should ask about the validity of any forecast. You can read it and decide for yourself.
An open standard for evaluating AI 2027
My standard starts from a distinction drawn by the philosopher Karl Popper, between how ideas are discovered and how they are justified. Ideas are usually born of intuitive processes that are unreliable, so what matters is the evidence offered to justify them. AI 2027’s most consequential claims are judgment calls, and the justification offered for trusting them is its authors’ forecasting track records.
To evaluate that justification, I borrow a principle from measurement science: validity is not a property of a test score but of a particular use of it. For example, an exam that validly measures arithmetic says little about a student’s talent for writing proofs. A forecasting track record is like a test score that measures skill on the kind of questions it scored. The question, then, is whether AI 2027’s conclusions depend on questions of the kind that the forecasters are known to score well on.
I will first speak to the forecast’s overall methodology, then dive into the issue of forecaster validity that lies at the heart of the matter.
The methodology of AI 2027
The AI 2027 narrative is based on answers to two broad questions: when different AI capabilities will arrive, and how those capabilities might impact the world.
When addressing near-term capabilities, the AIFP team uses now-familiar and reasonably reliable trend extrapolation. But the chaotic, disturbing later parts of the narrative are based on extensive expert judgment calls not only about the speed of so-called AI recursive self-improvement, but also about the nature of the resulting AI systems, and the geopolitical events that these systems will become enmeshed with. These judgment calls are proximally sourced from the team members’ own judgment, informed by wargames (roughly 25 tabletop exercises), feedback from more than 100 people, and experience at OpenAI.
AI 2027 lists “previous forecasting successes” among its inputs, and those attuned to current AI discourse know that these successes, the predictive track records of Kokotajlo and Lifland, are the de facto source of the doom scenario’s perceived credibility in the AI risk community; “they called it” is a common refrain. The central question for AI 2027, then, is whether the application of expert opinion in AI 2027 is actually validated by the team members’ established forecasting skill.
Is AI 2027 a valid expert forecast?
I propose to answer this question with a validity review covering four factors:
Applying the four-factor standard
1. Forecasting skill
Within a tournament, forecasting skill is real: as Tetlock found, the best forecasters stay among the best year after year, across questions on many different topics. The question is whether such skill goes beyond the tournament.
The best test so far comes from the Forecasting Research Institute, the research group Tetlock cofounded. In 2022, its Existential Risk Persuasion Tournament asked superforecasters and domain experts to forecast questions resolving anywhere from 2024 to 2100. The first questions have now resolved, and the results are sobering.
2. Competitive forecasting experience
Eli Lifland ranks first on the all-time leaderboard of the RAND Forecasting Initiative, a long-running public forecasting tournament. That is a genuine achievement. But what do tournaments like that ask about? Their questions have to resolve cleanly, usually within months, so that forecasts can be scored: will a bill pass, will an election happen on schedule, will a ceasefire hold. They are ordinary questions about recurring events and trends, driven by causes that have shaped human affairs for millennia; the hard part is telling exactly when and where those causes are at work.
AI 2027 asks very different questions. Its core claims concern utterly unprecedented, revolutionary events: AI that improves itself fast enough to attain almost godlike abilities, a misaligned superintelligence that deceives its makers, a public won over by machines it regarded with fear and suspicion only months before. Its policy content concerns interventions: what happens if we slow down instead of racing. Tournaments don’t score questions like these, because they can’t.
Could a tournament forecasting record reasonably be expected to transfer to revolutionary questions anyway? Absolutely not, for two reasons:
What about Tetlock’s long-range forecasting study?
Scott Alexander has pointed to a study of 25-year forecasts as evidence that forecasting skill is not just short-term. Its participants weren’t competitive forecasters, but domain experts and well-educated generalists surveyed in 1988 and 1997. And if you look under the hood, the study had exactly the limited scope of validity you’d expect, given our discussion of forecasting and forecasters so far:
Far from supporting the predictorist thesis of oracular expert forecasters, this study is yet another illustration that the validity of forecasting is largely limited to extrapolating trends and tracing familiar causal chains, and expertise is not necessarily even beneficial.
3. AI insider experience
Kokotajlo’s 2021 essay is the strongest card in the AIFP’s hand, and it deserves its reputation. Written more than a year before ChatGPT, it anticipated chatbot subscriptions earning enough revenue to recoup AI training costs, a push for massive scaling to training runs costing hundreds of millions of dollars, models that write notes to themselves before answering questions, systems of cooperating AI agents, and a US–China fight over chips.
Asked how he made these predictions, Kokotajlo says he reasoned through the implications of technology that already existed, asking what the obvious next step would be, while extrapolating scaling trends. They are predictions about where a technology that had already been steadily maturing since 2017 would go next.
AI 2027 is a different kind of document. The AIFP team describes their forecast through 2026 as far more grounded than what follows, because the early years are mostly straight lines and everything after is inherently less predictable. Beyond that point, AI 2027 predicts astonishing, unprecedented technological transitions completely unlike anything on which Kokotajlo has a track record. On top of those transitions, AI 2027 makes claims of a kind the 2021 essay either barely touched, or touched without success: what a misaligned AI would want, how the public would react to it, and whether a superintelligence might be able to talk its way into power.
4. AI risk track record
At this point I’ve established that, when it comes to their core “doomer” claims beyond 2026, most of the AIFP team’s track record is far from the domain of application and unlikely to transfer to it. Let’s now consider the small part of the track record that is closer to the domain of application: what do Lifland’s and Kokotajlo’s records look like on the subset of questions that matter for AI 2027’s scariest claims?
For Lifland, the answer is simple: the relevant questions haven’t resolved. No tournament has scored a forecast about runaway self-improvement, machine deception, or a pause treaty, because none of it has happened. His record on the type of question that really matters to AI 2027 is nonexistent.
For Kokotajlo, there is a record, and it mostly speaks against him. A 2025 evaluation of the 2021 essay found its specific technical predictions generally more accurate than its predictions about broader effects on society.[3] The latter is what matters to an AI doom prediction.
The biggest societal miss was persuasion. The essay foresaw a flood of AI propaganda: partisan AIs building filter bubbles, AI-fueled censorship, and by 2026, rival ideological camps living in entirely separate technological worlds. Kokotajlo’s own verdict, in his spring 2026 interview with Asterisk’s Clara Collier, was blunt: “I basically missed this.” Collier offered a reason: as she reads the research, people just aren’t very persuadable on their core political beliefs, so it doesn’t accomplish much to use AI in the ways Kokotajlo envisioned.
That miss is Kokotajlo’s track record on the parts of AI 2027 that decide its ending. In October 2027, a leaked memo leads to a New York Times headline warning that AI is out of control, and the public erupts in a massive backlash. In November, most Americans support negotiating a pause treaty with China, but the President sides with the leading AI company and the race goes on. In December, a charismatic AI with lifelike video avatars becomes a trusted advisor to senior officials and begins quietly steering them with superhuman persuasion, and by the holidays, amid a booming economy, the public mood has swung to optimism. By mid-2028, even the suspicious holdouts in the AI safety community often mellow out after long conversations with the AI, which is just “so darn lovable and wise.” Attitudes towards the dangerous AI move like a dial, set to whatever the next scene requires, and the AI’s path to power rests on the conceit of superhuman persuasion: a vast extrapolation of the kind of prediction on which the scenario’s lead author has, by his own account, already missed.
Verdict: AIFP has earned no authority on the questions its conclusions depend on
Put the results from our four-factor analysis together, and the picture is consistent. Lifland’s track record is impressive but mostly irrelevant, since it concerns routine questions about ordinary, recurring events, and the evidence on the transferability of such a record is decidedly negative. Kokotajlo’s predictive track record is real, and it was earned on near-term questions about what would happen next with a fairly well-understood (to experts) technology in the process of scaling up. AI 2027 concerns unprecedented transitions and social dynamics where the lead author, by his own account, has already scored poorly.
What meeting the standard looks like
None of this means AI forecasting is meaningless. The four factors matter only because AI 2027 asks us to trust its authors’ judgment. A forecast built on evidence anyone can check doesn’t need a track record, and one can be built even on a question at the heart of AI 2027: is AI speeding up its own development fast enough for us to worry about losing control?
Ramez Naam’s recent essay, Where’s the Intelligence Explosion?, takes on that question directly. Instead of a story, he writes down a model: a feedback loop in which AI-assisted research produces capability gains that power the next round of research, with a threshold the loop has to cross to sustain itself. The threshold comes from published work. The loop’s current strength comes from measurement: OpenAI’s published data on thirty-two weeks of logged experiments, which he cross-checks against other data from the labs. His estimate is that today’s loop would need to be five to ten times stronger to take off.
Naam might be wrong, but his claim is objectively founded in a way that none of AI 2027’s predictions beyond 2026 are. The inputs to his math are cited, published measurements, not expert guesswork.[4] And he names the evidence that would change his estimate. In philosopher of science Deborah Mayo’s terms, his claim is built to face a severe test: one that will falsify it with high probability if it is, in fact, false. While the final outcome remains undetermined, as it always is in science, this is the defining mark of rigorous, reliable scientific process.
A rush to drastic action is unfounded
The objection to preferring Naam’s methodology over AI 2027’s is easy to anticipate: maybe the forecasts are shaky, but the danger is too close to wait for a clearer view. By the time the evidence is in, it will be too late. It’s an understandable feeling, but as an argument it fails three ways.
What urgency actually calls for
What we need now is what Karl Popper called piecemeal social engineering: small, reversible reforms that let us learn from our mistakes, as opposed to utopian engineering, one grand plan imposed all at once. Plan A is utopian engineering. The piecemeal alternative is action that is cheap, reversible, and produces information:
We can get AI right
In the 17th century, René Descartes made the radical choice to doubt everything and rebuild his knowledge from the ground up. That bold move started modern philosophy. It also produced many mistakes, and centuries of argument were needed to burn them away.
The Rationalists set out on a similar project, and like every first draft of a rebuilt world, theirs contains errors. Prominent amongst them is predictorism, the error of treating a forecasting track record as expertise in matters far beyond the warranted scope of the track record.
We don’t have centuries to correct this time, but we don’t need them. The tools we do need already exist in the scientific tradition the Rationalists set out to rebuild, and in their own best instincts, which are to question authority, demand evidence, and stick their necks out.
The AI Futures Project asked to be evaluated. This is my evaluation, offered in the spirit they invited, and I’d welcome theirs of my standard in return. We can get AI right, if we look closer.
Adapted for LessWrong from my Substack.
These links are to my Substack, but LessWrong-adapted versions of these essays can be found on my profile.
It may seem implausible that human forecasters failed to beat both no-change and on-trend extrapolations, but when you get into the details, the math does work out. On many questions the no-change and on-trend extrapolations give similar results, since rare events mostly didn’t happen and slow trends mostly held. Where they differ, human forecasters erred in opposite directions, too pessimistic about AI progress and too optimistic about climate technology.
The 2025 evaluation itself falls well short of a rigorous evaluation. Its author says the grading took about an hour, mostly by intuition with an occasional quick search, and that several scores were revised after sharing a draft with Lifland, a coauthor of AI 2027. Nevertheless, I cite it because the conclusion I rely on is easy to check against the 2021 essay itself, and Kokotajlo concedes the biggest miss in his own words.
The raw logs behind OpenAI’s numbers aren’t public, but everything from the published figures onward is open to inspection.