The “AI 2027” Scenario: How realistic is it?
We recorded 40 claims from this post. 18 were checked against evidence. 22 can’t be checked. 11 more are held. Read the post at its source.
Claims from this post (40)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: UnresolvedSaid
“The US has done almost nothing legislatively about the risks of AI (aside from deepfake porn).”
What we tested: The US has passed almost no legislation addressing AI risks, apart from deepfake porn.
Assessment: Can’t be checkedSaid
“And, as I will argue at the end, stoking fear, uncertainty and doubt may actually ultimately be harming the (noble!) cause of AI safety, rather than helping it.”
What we tested: Stoking fear, uncertainty and doubt about AI may be harming, rather than helping, the cause of AI safety.
Assessment: UnresolvedSaid
“And so far, anyway, the impact on the labor market of Generative AI has been modest; ditto for productivity in any measurable statistics.”
What we tested: So far, generative AI's impact on the labor market and on measurable productivity has been modest.
Assessment: UnresolvedSaid
“Cell phones and the internet were pretty big, but it’s not obvious their impact was greater than the industrial revolution, and so far AI’s impact has (despite all the PR) been far less than either cell phones or the internet.”
What we tested: So far, AI's impact has been far less than that of cell phones or the internet.
Assessment: Can’t be checkedSaid
“But there is a vastly higher probability that the future won’t transpire as described; it might not go anything at all like what they describe.”
What we tested: The future of AI is far more likely not to unfold as the AI 2027 scenario describes than to unfold that way.
Assessment: Can’t be checkedSaid
“We kinda sorta have stumbling agents now, (though I don’t think there is a single agent yet that any reader here would trust to actually act on their behalf without monitoring).”
What we tested: We kinda sorta have stumbling AI agents now, and Marcus does not think there is yet a single agent that any of his readers would trust to act on their behalf without monitoring.
Assessment: ContradictedSaid
“Google Duplex, a widely hyped and now mostly forgotten system for “Accomplishing Real-World Tasks Over the Phone” from 2018 still hasn’t fully materialized, exactly as Ernest Davis and I projected at the time.”
What we tested: Google Duplex, hyped in 2018, still has not fully materialized.
What happened: Google reported over one million bookings spanning restaurant reservations and salon appointments, alongside millions of business-information updates. This establishes substantial realization of the original core tasks; the automation figure covers calls generally, not a separately measured appointment success rate. (Google, )
Assessment: UnresolvedSaid
“Remember how we were told in 2012 that we would all have driverless cars by 2017? That hasn’t come to pass, either, except in about 10 of the world’s 20,000 cities.”
What we tested: Driverless cars, promised for everyone by 2017, operate in only about 10 of the world's 20,000 cities.
Assessment: SupportedSaid
“Remember how we were told in 2023 that hallucinations would be solved in a matter of months? They are still here.”
What we tested: Despite 2023 promises that hallucinations would soon be solved, they persist as of May 2025.
What happened: OpenAI's newest reasoning models in April 2025 hallucinated on 33% (o3) and 48% (o4-mini) of PersonQA questions, higher than the older o1. Hallucinations were plainly still present a month before the post. (OpenAI, )
Assessment: Can’t be checkedSaid
“In reality, the chance that we will have reliable AI agents that truly advance AI research by the end of the year is small.”
What we tested: There is only a small chance that reliable AI agents that truly advance AI research will exist by the end of 2025.
Assessment: Can’t be checkedSaid
“And some of what they describe (like AI’s going rogue, completely beyond human control) might never happen.”
What we tested: AI going rogue and completely beyond human control might never happen.
Assessment: UnresolvedSaid
“Overall, then, the AI 2027 scenario almost certainly underestimates how much time we have to prepare for general intelligence, by years if not decades.”
What we tested: General intelligence will almost certainly arrive years, if not decades, later than the AI 2027 scenario projects.
Assessment: UnresolvedSaid
“Agent-1 is an extension of what the big companies are showing demos of now. It doesn’t actually exist, except in demo form.”
What we tested: Agents like the scenario's Agent-1 exist only in demo form from the big companies.
Assessment: Can’t be checkedSaid
“In all likelihood the hallucinations and boneheaded errors that have bedeviled LLMs since their introduction will plague agents (which are after all built on LLMs), too.”
What we tested: AI agents, being built on LLMs, will likely be plagued by the same hallucinations and boneheaded errors as LLMs.
Assessment: Can’t be checkedSaid
“Far more likely is that Agent-1 type systems will be in a kind of “just a demo, still being debugged” phase for years.”
What we tested: Agent-1-type AI agents will likely remain in a 'just a demo, still being debugged' phase for years.
Assessment: UnresolvedSaid
“The authors wisely don’t want to put all their eggs in the basket of pretraining scaling (which used to fuel most people’s AGI fantasies), because by now they know perfectly well that pretraining scaling has hit a rough patch.”
What we tested: Pretraining scaling has hit a rough patch.
Assessment: Can’t be checkedSaid
“A tiny bit smarter each day is not enough for the kind of quantum leaps that are implied.”
What we tested: Small daily incremental gains in LLMs are not enough to produce the quantum leaps the AI 2027 scenario implies.
Assessment: Can’t be checkedSaid
“First of all, LLMs like GPT-4 have been getting smarter every day since OpenAI first showed GPT-4 to Bill Gates in the summer of 2022, but we still don’t have GPT-5 level intelligence, let alone Agent-2-level.”
What we tested: Despite incremental daily improvement since 2022, LLMs still have not reached GPT-5-level intelligence.
Assessment: UnresolvedSaid
“A big part of the issue— which AI 2027 never faces — is that we simply don’t know how to create synthetic data for many other domains, like what might happen if India and Pakistan go to war using drones or what the effects of 30% tariff might be on the global economy.”
What we tested: We do not know how to create useful synthetic training data for many open-ended, unverifiable domains.
Assessment: Can’t be checkedSaid
“Breakthroughs in “neuralese recurrence” (whatever that is) and memory (which real-world people have been working on for decades, with relatively little success) and “iterated distillation and amplification” comes straight from the kind of writing that gave us Star Trek’s “positronic brain”.”
What we tested: The scenario's breakthroughs in neuralese recurrence, memory (which people have worked on for decades with relatively little success), and iterated distillation and amplification come straight from the kind of writing that gave us Star Trek's positronic brain.
Assessment: Can’t be checkedSaid
“Another unsatisfying move is the jump from having agents that predict internet text (which is basically what all current text-based LLMs do) to agents that have basic personalities and drives – which is entirely speculative and something that no actual LLM has.”
What we tested: No actual LLM has basic personalities and drives; current LLMs basically predict internet text.
Assessment: UnresolvedSaid
“In reality, producing text in response to instructions is already a requirement of ALL current models, and has been since the original GPT, but that pressure has not led any of those models to develop personality or drives in any deep sense whatsoever.”
What we tested: Producing text in response to instructions has been a requirement of all current models since the original GPT, but that pressure has not led any of them to develop personality or drives in any deep sense whatsoever.
Assessment: Can’t be checkedSaid
“The fictional force is strong in this team, but taken as an actual piece of forecasting, the scenario is unconvincing.”
What we tested: Taken as an actual piece of forecasting, the AI 2027 scenario is unconvincing.
Assessment: Can’t be checkedSaid
“such as those coding machines taking on human personalities and drives (which might happen never or not soon), and choosing to enslave us (which some thinkers like Steve Pinker argue is deeply unlikely, ever, and which I think is possible but very far from certain).”
What we tested: AI taking on human-like drives might never happen or not soon, and AI choosing to enslave humans is possible but very far from certain.
Assessment: Can’t be checkedSaid
“All of which is to say the central scenario in the story is a worst-case scenario that is unlikely to happen soon - if ever —even based on the data presented in the author’s own appendix.”
What we tested: The AI 2027 worst-case scenario is unlikely to happen soon, if ever.
Assessment: Can’t be checkedSaid
“And you could worry about that even if you don’t expect machines to go rogue, since bad human actors could certainly use further advances in intelligence to cause mayhem, if those advances are left unchecked.”
What we tested: If advances in AI are left unchecked, bad human actors could use them to cause mayhem.
Assessment: Can’t be checkedSaid
“The authors are trying to call attention to a real problem: we are not sufficiently prepared for what happens when and if “superhuman intelligence” arrives, especially not if it is misaligned with humans.”
What we tested: We are not sufficiently prepared for the arrival of superhuman intelligence, especially if it is misaligned.
Assessment: Can’t be checkedSaid
“We certainly aren’t prepared now, and if we don’t get moving, we won’t be prepared in 10, either.”
What we tested: We are not prepared for AGI now, and unless we act, we will not be prepared in ten years either.
Assessment: Can’t be checkedSaid
“But here’s the thing - projects like AI 2027 are probably having a paradoxical effect. They are written with the intention of slowing down an arms race to build a technology that we can’t control, but what they are actually doing is speeding up that very arms race, in two different ways.”
What we tested: Projects like AI 2027 are probably speeding up the AI arms race they intend to slow.
Assessment: UnresolvedSaid
“Their stories about this are, in my view, greatly flawed, but having outside groups with science fiction chops writing stuff like this distracts away from those flaws, and gives more power to the very companies trying hardest to race towards AGI.”
What we tested: OpenAI's and Anthropic's AGI-imminence stories are greatly flawed, and outside scenarios like AI 2027 distract from those flaws and empower those companies.
Assessment: UnresolvedSaid
“AI 2027 isn’t slowing them down; it’s putting wind (and money, and political power) in their sails.”
What we tested: AI 2027 is giving AI companies like OpenAI and Anthropic more money and political power rather than slowing them.
Assessment: UnresolvedSaid
“It’s also encouraging the world to make short-term choices about AI (e.g., making plans around export controls when China will inevitably catch up) rather than longer-term choices (investing in research to develop safer, more alignable approaches to AI).”
What we tested: China will inevitably catch up in AI despite export controls.
Assessment: UnresolvedSaid
“It’s also encouraging the world to make short-term choices about AI (e.g., making plans around export controls when China will inevitably catch up) rather than longer-term choices (investing in research to develop safer, more alignable approaches to AI).”
What we tested: AI 2027 is encouraging the world to make short-term choices about AI, such as planning around export controls when China will inevitably catch up, rather than longer-term choices such as investing in research on safer, more alignable approaches to AI.
Assessment: UnresolvedSaid
“The second issue is that by writing scenarios that trade so heavily on China-US conflict, they are feeding the worst fears of hawks, both in the US and China — escalating how much money and power both sides will give to companies racing as fast as possible, and reduce investments in efforts to mitigate the risks of building AI that we are scarcely able to control.”
What we tested: US-China conflict framing in AI scenarios feeds hawks on both sides, increasing money and power for racing AI companies and reducing investment in AI risk mitigation.
Assessment: UnresolvedSaid
“Tall tales about the imminence of AGI aren’t slowing down the AI race dynamic that the authors of AI-2027 want to mitigate; they are speeding that very dynamic up.”
What we tested: Stories claiming AGI is imminent are speeding up, not slowing, the AI race dynamic.
Assessment: UnresolvedSaid
“The all too common strategy of “scare everyone with the thought of superintelligence” is backfiring, because it’s making it seem that superintelligence is inevitable, and unstoppable, and achievable only by the companies we have today.”
What we tested: Scaring people about superintelligence backfires because it makes superintelligence seem inevitable, unstoppable, and achievable only by today's companies.
Assessment: Can’t be checkedSaid
“Sam Altman is laughing his way all the way to the bank.”
What we tested: Sam Altman is laughing all the way to the bank.
Assessment: Can’t be checkedSaid
“But international collaboration, such as a massive global effort, ala CERN for AI, such as I suggested in 2017, focused on AI safety, might be one way to avoid the very perils they describe.”
What we tested: International collaboration, such as a massive global CERN-like effort for AI focused on AI safety, might be one way to avoid the perils AI 2027 describes.
Assessment: Can’t be checkedSaid
“Such an effort might be particularly likely to succeed it were focused on developing altogether new, more transparent approaches to AI, in contrast to the current fad of trying to pull alignment out of black box LLMs, which has thus far yielded little fruit.”
What we tested: An international AI safety effort might be especially likely to succeed if it focused on new, more transparent approaches to AI rather than on aligning black-box LLMs.
Assessment: Can’t be checkedSaid
“If “great” AI research agents don’t arrive by the end of 2025, everything in the rest of the essay gets moved back.”
What we tested: If 'great' AI research agents do not arrive by the end of 2025, everything later in the AI 2027 scenario's timeline is pushed back.
Held from this post (11)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The two reviews read the claim differently
- The two reviews read the claim differently
- The two reviews read the claim differently
- The record does not meet the current rules. Our reading, not assessed: Marcus seriously doubts that by the end of 2025 AI agents will both be great at helping with AI research and have resolved the error tendencies of mid-2025 prototype agents.
- The record does not meet the current rules. Our reading, not assessed: AI agents are very unlikely to be contributing fundamentally new ideas to AI research by the end of 2025.
- The two reviews reached different assessments. Our reading, not assessed: Efforts to achieve alignment in black-box LLMs have so far yielded little fruit.
- The record does not meet the current rules. Our reading, not assessed: The chance that within the next decade we will all have been replaced, in a bloodless coup, by domesticated human-like animals living in glorified cages is indistinguishable from zero.
- The record does not meet the current rules. Our reading, not assessed: AGI is highly unlikely to arrive within three years, though it cannot be ruled out within ten.
- The two reviews reached different assessments. Our reading, not assessed: Companies churning out synthetic math problems, especially since last fall, have produced considerable improvements on math and coding, but those gains have been far from universal.
- The two reviews read the claim differently
- The two reviews read the claim differently