Home / Blog

AI forecasting

Why six forecasters now expect a superhuman AI researcher in 2027

All six authors of the AI Futures Project’s new forecast expect an AI that out-researches the best human AI researchers sometime in 2027. Three months ago the one who expected it soonest put even odds on it arriving by the end of 2028. The biggest reason is one estimate, that AI is now making AI progress go about 1.7 times as fast as it used to.

Main takeaways
  • A METR estimate was the biggest reason. METR estimated that Anthropic now makes AI progress about twice as fast as before October 2025, mostly because of AI. The forecasters read that as a 1.7x speed-up from AI alone, which their August models had given less than a 4% chance.
  • A 1.7x pace needs much more than 1.7x faster researchers. In the authors’ model, making researchers 1.7 times as fast speeds up overall progress by only about 1.03 times.
  • A straight line from today reaches a superhuman AI researcher in late 2027. Changing one assumption at a time moves the date between February 2027 and December 2028.
  • The group’s model, told about 1.7x, puts it in early 2027. With Kokotajlo’s settings, about 1 in 80 simulated futures matched the measured pace. Of those, 90% reached a superhuman AI researcher before 2028.
  • Much of the evidence is hidden. Anthropic redacted most of what METR found most informative, and a METR team without inside access saw much less speed-up in preliminary numbers.

Eli Lifland is a forecaster at the AI Futures Project, the nonprofit group behind the scenario AI 2027. In August he gave even odds that AI would be better than top human experts at virtually every task by July 2032. On 10 November the group published a timelines update titled “Looks Like the Intelligence Explosion Will Happen in 2027 After All.” Its six authors, among them Lifland and Daniel Kokotajlo, each give a median for each milestone. A median is the date by which the forecaster thinks the milestone is as likely to have happened as not.

In August, Lifland’s median for AI that beats top experts at nearly everything was July 2032. Where did he put it on 10 November?

Five forecasters moved their medians earlier Medians for top-expert-dominating AI before and after the 10 November 2026 update. Lifland moved from July 2032 to September 2028. Kokotajlo moved from November 2028 to September 2027. Halstead moved from April 2031 to April 2028. Larsen moved from early 2032 to April 2028. Dean moved from early 2035 to September 2028. MEDIAN FOR TOP-EXPERT-DOMINATING AI today Lifland Sep 2028 Jul 2032 Kokotajlo Sep 2027 Nov 2028 Halstead Apr 2028 Apr 2031 Larsen Apr 2028 early 2032 Dean Sep 2028 early 2035 2027 2029 2031 2033 2035

September 2028, almost four years earlier. The other four authors in the chart moved earlier too, each by at least 14 months.

Figure 1. Make a guess. Open circles are each author’s previous median and blue dots the 10 November one, from the update’s appendix chart. The exact August dates are from the August update. Thomas Larsen’s and Romeo Dean’s earlier medians are from about March 2026, read off the authors’ chart. Brendan Halstead is the fifth author shown, and the sixth, Lukas Finnveden, has no earlier forecast in the chart. The robot is a cartoon.

In August, Kokotajlo, the member of the group with the earliest dates, gave a superhuman AI researcher even odds of arriving before 2029. Now all six authors put their median for it in 2027.

The single biggest source of evidence, they write, was a 5 November report from METR, a nonprofit that tests AI systems, which an earlier post here walks through. METR estimated that Anthropic now improves its models about twice as fast as before October 2025, and guessed that “perhaps ≈80%” of the extra speed comes from AI. Without AI’s share, the pace would have risen only from 1 to 1.2, so the forecasters credit AI with 2 ÷ 1.2, or about 1.7x. In August their own computer model of AI progress, run with Kokotajlo’s inputs, gave a speed-up that large by November 2026 less than a 4% chance.

The target is an AI that out-researches the best researcher

The update tracks two milestones. The nearer one is the superhuman AI researcher, or SAR.

Superhuman AI researcher (SAR)

The authors define a SAR as “an AI that matches the research taste of the best human researchers at an AGI company and can fully automate its coding.” Research taste is “the ability to select research directions and experiments, and interpret their results.” An AGI company is one trying to build artificial general intelligence, AI that can do most of the work people do. Roughly, a company with a SAR would rather fire all its human researchers and run on AI alone than keep a team of the world’s best researchers with no AI.

The farther one is top-expert-dominating AI, or TED-AI, which they describe as AI “better than top human experts at virtually all tasks.” Both sit on a ladder that ends at superintelligence, which an earlier post here explains. Most of the update is about SAR, which the team thinks could be a “point of no return.” An AI that does AI research better than any person can be set to work on the next AI, and that one on the one after. This loop is called recursive self-improvement.

A 1.7x pace asks far more of AI than a 1.7x faster workforce

A speed-up of 1.7x sounds like AI making each researcher 1.7 times as productive. The authors separate three kinds of speed-up, and METR’s 1.7 is the smallest of them.

  • Labor uplift is how much faster researchers get their own work done with AI. Anthropic staff, polled informally, put theirs at 12 times by October 2026, which METR calls “only very loose evidence.”
  • Software research uplift is how much faster the company improves its algorithms, the methods it uses to build and train models.
  • The pace of progress is how fast the models themselves get better. This is the number METR estimated.

The authors picture AI progress as a car driving up a hill. Its speed is the pace of progress, and its engine is research effort. The hill gets steeper as the car climbs, because the easy improvements get found first.

What does it take to make the car go 1.7 times as fast?

A car on a steepening hill A cartoon car climbs a hill that gets steeper to the right. A faint outline of the car sits low on the flat part, labelled before October 2025 at 1 times speed. The main car, higher up, goes 1.7 times as fast and carries a large engine. A sign warns that the hill gets steeper. A panel shows that software supplies about two thirds of progress and hardware one third, and that the research crew must work about 7.8 times as fast with the same experiment computers. before Oct 2025 1× speed 1.7× speed engine 3.1× WHERE PROGRESS COMES FROM software ⅔ hardware ⅓ ! steeper ahead RESEARCH CREW ≈7.8× as fast same computers for tests
Pace of progress1.7×the car’s speed
Software progress2.05×hardware keeps its old pace
Research effort3.1×the engine, on a steeper hill
Researcher speed≈7.8×with the same computers

Experiments need computers, and AI has not given the company more of them. With computers held fixed, research effort grows more slowly than researcher speed, so a 3.1 times bigger engine needs researchers working about 7.8 times as fast.

Figure 2. Press each step. The analogy and the first three numbers are the authors’. Software is about two thirds of progress, so a 1.7x pace needs (3 × 1.7 − 1) ÷ 2 = 2.05x software progress. Each year of software progress makes the next one about twice as hard, and about 0.6 extra years of it have piled up since the speed-up began, so 2.05 × 20.6 = 3.1. The researcher figure is my arithmetic. In the formula the authors borrow, research effort grows as researcher speed to the power 0.55 when computers are held fixed, so researchers need 3.11/0.55 ≈ 7.8. The car and crew are cartoons.

Run in reverse, their model makes researchers 1.7 times as fast with nothing else changed, and overall progress speeds up by only about 1.03 times. In a run of the model tuned to match AI 2027, the pace reaches 1.7x around April 2027, so by this measure, they write, “we are ~5 months ahead of schedule.”

A straight line from today reaches the best researcher in late 2027

The simplest method in the update needs three numbers. The first is how good AI’s research taste is today. For that the authors take Model 4, an internal Anthropic model in use since about 25 September 2026. With Kokotajlo’s settings, its taste sits 3.5 standard deviations below the median Anthropic researcher. A standard deviation, or SD, is the typical distance between a person and the middle of a bell curve, so −3.5 SDs is about the 0.02nd percentile. A SAR needs +3.09 SDs, the best researcher in a thousand.

The second number is how fast taste improves. Kokotajlo’s settings add 3 SDs for every year of progress at the 2025 pace. The third is the pace itself. On AECI, Anthropic’s score that combines many benchmark tests, its models gained about 15 points a year before October 2025 and about 29 points a year since. Each point is then worth 0.2 SDs, so taste climbs 5.8 SDs a year and covers the 6.59 SD gap in about 1.1 years.

Drawn as a straight line with Kokotajlo’s inputs and no further speed-up from AI, today’s pace reaches a superhuman AI researcher around November 2027, about 14 months after Model 4.

They call the method conservative, meaning it probably errs late, because it holds the pace at 29 points a year.

What date do you get with other inputs?

Model 4’s research taste

SDs of taste per year of 2025-pace progress

AECI points a year

From the left tail of the bell curve to the best researcher in a thousand A bell curve of research taste among Anthropic researchers. A robot marks Model 4's taste at minus 3.5 standard deviations. A flag marks plus 3.09 standard deviations, the best researcher in a thousand. With 3 standard deviations per year of 2025-pace progress and 29 AECI points a year, the gap of 6.59 standard deviations closes in 1.1 years, so a superhuman AI researcher arrives around November 2027. −4−3−2−10+1+2+3+4RESEARCH TASTE, IN SDS FROM THE MEDIANmedian researcherbest in 1,000SAR, +3.09Model 4gap 6.59 SDCALENDAR202720282029SAR Nov 2027

Gap 6.59 SD ÷ 5.8 SD a year = 1.1 years

SAR around November 2027, about 14 months after Model 4.

Figure 3. Pick the inputs. Years to SAR = (3.09 − starting taste) ÷ (slope ÷ 15 × points a year), counted from 25 September 2026. With one input changed at a time, this reproduces every row of the authors’ table. The 4.5 slope comes from TasteVal, a benchmark that measures one kind of research taste. The bell curve is a standard normal curve, and the robot is a cartoon.

The date is hard to push late. Lifland’s lower slope of 2.3 moves it only to March 2028, and only going back to the old pace of 15 points a year gives December 2028. The authors also ran a Monte Carlo simulation, which draws many random combinations of inputs and records the date each one gives. With Model 4’s taste centered higher, at −1.5 SDs, the median came out at July 2027, and 65% of draws reached SAR by the end of 2027.

Other straight lines land within a year of it

Anthropic publishes a measure of how much of its research AI does. In July 2026 alone it rated about 15,000 research tasks on a six-step scale from Epoch AI. AL stands for automation level, and the scale runs from AL0, where AI “adds nothing,” to AL5, where AI works “end to end with little or no human involvement.”

Who does the work at each level?

AL0, not used. “AI adds nothing.”

AL1, marginal. “Occasionally convenient for lookups, brainstorming, or polishing writeups, but doesn’t change how the work is done or who is doing it.”

AL2, assists. “Meaningful speed ups or improvements to parts of the task; human mostly drives and reviews everything.”

AL3, collaborates. “AI doing large chunks under close human direction; human still makes calls and stitches things together.” Over 90% of Anthropic’s sampled tasks were at this level or above.

AL4, leads. “Completing most of the task end to end from a high-level ask; human supervises, course-corrects, and approves outputs.” Claude led 26% of Anthropic’s sampled tasks in August 2026, up from under 1% in February.

AL5, autonomous. “End to end with little or no human involvement.” Anthropic found no part of its research work at this level.

26%of tasks led by Claude (AL4) in August 2026
>90%at AL3 or above
0parts of the work fully autonomous (AL5)
Figure 4. Press a level. Level names and quoted descriptions are Epoch AI’s scale as given in the update. Shares are from Anthropic’s measurements. The people and robots are cartoons.

The share of tasks that Claude leads rose from under 1% in February 2026 to 26% in August, and no part of the work was rated fully autonomous. If SAR arrives when 95% of tasks are fully automated, S-shaped growth curves fitted to these shares put it in November 2027. The authors call their guess about how quickly each level follows the one below it “based on little data.”

A second line follows staff estimates of their own speed-up, which reached 12 times with Model 4. If a SAR is worth roughly a 1,000-fold speed-up for researchers, that trend gets there around August 2028, though taking self-reports “at face value” could make it too fast. Both lines leave out any further speed-up from AI, so the authors count them as conservative. In their summary, the other simple models “predict late 2027 to mid-2028.”

Weaknesses expected to slow AI have not slowed it much

If AI can do 90% of a job instantly and the last 10% still needs a person, the whole job goes at most 10 times faster. This is Amdahl’s law, which Gene Amdahl described for parallel computers in 1967. The authors count this kind of bottleneck among the main reasons people doubted that AI would soon speed up AI research.

How much faster does a project go as AI takes over more of its tasks?

50% by AI
The tasks people still do set the speed Twenty task tiles, half done by robots and half by people. A project bar shrinks to half its length, a 2 times speed-up. A curve shows the speed-up, 1 divided by 1 minus the share automated, rising slowly at first and then steeply: 10 times at 90% and 100 times at 99%. TWENTY TASKS IN ONE PROJECTTime for the project, all by peopleWith AI doing its share instantlySPEED-UP, BY SHARE OF TASKS DONE BY AI1×10×100×0%50%90%2.0×

Half the tasks are done by AI, so the project takes half as long, a 2× speed-up.

Figure 5. Drag the slider. In this extreme case, automated tasks take no time and the rest take as long as before, so the speed-up is 1 ÷ (1 − share automated). The authors use the same formula. Real automated tasks still take some time, so real speed-ups are lower. The task tiles are cartoons.

Today’s models still have weaknesses that were expected to hold progress back, such as tasks whose results are hard to measure. If those weaknesses are not holding progress back much, the bottleneck argument was weaker than skeptics thought. If they are, then Figure 5 applies, and each newly automated task buys a bigger jump than the one before. Either way, the authors write, “you update towards Superhuman AI Researchers being much closer in time than we previously thought.”

The group’s own model finds 2027 once it is told about 1.7x

The group’s main tool is the AI Futures Model, a computer simulation of AI progress feeding back on itself. Each forecaster sets a range for every input, and the model draws many possible futures from those ranges. For this update the authors kept only the futures that showed a 1.7x pace on 1 November 2026. In probability this is called conditioning. It asks what the future looks like, given what was just measured.

What happens to the forecast if you keep only the futures that match what was measured?

Filtering simulated futures A toy model with invented numbers. Two hundred simulated futures, each a line of the pace of progress over time. A gold window at today marks a pace near 1.7 times. The futures that pass through the window are highlighted, and their arrival dates, marked in a strip below, sit earlier than those of all the futures together. 1×2×5×10×PACE OF PROGRESS, TOY MODELmeasured today, about 1.7×todayWHEN EACH FUTURE REACHES THE GOAL−1 yrnow+1 yr+2 yr+3 yr+4 yr+5 yr

 

The real results, from the AI Futures Model

Kokotajlo’s settings

Futures that match≈1.2%
SAR medianFeb 2027
Chance of SAR before 202890%
TED-AI medianMay 2027

Lifland’s settings

SAR medianFeb 2027
Chance of SAR before 202890%
TED-AI medianJul 2027
Figure 6. Press the buttons. The top panel is a toy model with invented numbers, built to show the method. Each line stops when its future reaches the goal or leaves the chart. The cards below give the authors’ results from the AI Futures Model for different paces on 1 November 2026. The 1.2x pace on 1 May is the authors’ rough guess.

With Kokotajlo’s settings, only about 1.2% of futures were near 1.7x, roughly 1 in 80. In those, the median SAR date was February 2027, and 90% reached SAR before 2028. The kept futures, the authors write, “probably reach SAR a few months too early.” In the model, the easiest way to reach 1.7x by November is for AI to already have research taste near the median researcher’s. The team thinks today’s models are worse than that. The explanation they rate most plausible for the miss is that AI close to automating all coding speeds research up more than they had assumed.

All six medians fall in 2027

Each author then gave an all-things-considered forecast, blending the methods with judgment. All six put their SAR median in 2027, from May for Kokotajlo to November for Finnveden and Larsen. Their TED-AI medians run from September 2027 to September 2028, and each gives TED-AI between a 35% and a 65% chance of arriving before 2028.

SAR medianTED-AI median65% chance of TED-AI before 2028
“Let’s say May 2027 for SAR median.” He also suspects that “humans in general might have a bias towards underestimating AI capabilities progress.” “I set my median slightly lower than the 1.7x AI-attributed acceleration, given that I had a strong prior against these results.” “I generally have also updated towards more spikiness in AI capabilities.” He expects AI to get good at different tasks at very different speeds, which puts TED-AI well after SAR. “For SAR, I am mostly relying on AECI extrapolation.” He gives a 90% chance that 2027 brings at least as much progress as 2026. “Probability mass early 2027 vs. late 2027 doesn’t seem wildly different to me, but late seems a bit more likely than early.” “My all-things-considered view is that METR is probably right.” He also writes that the report “on its own does not have sufficient data in it to support that claim.”
Jan 2027Jan 2028Jan 2029
Figure 7. Press a name. Medians and the chance of TED-AI before 2028 are from the update’s forecast table, and quotes are from each author’s section. Every forecast assumes nobody slows AI down on purpose more than in the recent past.

Kokotajlo had written down in advance what would shorten his forecast, including AI revenue growing fourfold or more in 2026. Revenue, he writes, “seems to have gone up about 6x.” Lifland lists what pulls him the other way, including “a nagging feeling that the update feels so large without being based on commensurate evidence,” and markets that did not react to the METR report. He adds that this kind of instinct “has performed very poorly for me in the past.”

Most of the evidence is out of public view

METR wrote that “Anthropic redacted a majority of the evidence we found most informative for these conclusions.” The doubling has some public support in a benchmark chart. The 80% share, and so the 1.7, has “no comparable public support,” the authors write.

Reasons the 1.7 could be too high
  • METR warns that much of its evidence comes from “well-specified tasks that are easy to grade automatically.” If models are improving faster on clean tasks than on messy real work, the real speed-up is smaller.
  • A second METR team, without inside access, found “much less of a speed-up” on a similar measure in preliminary numbers.
  • Larsen points to an explanation with no help from AI. Anthropic may have shifted computing power toward reinforcement learning, a training method that rewards a model for correct answers. That raises benchmark scores more directly than improving pretraining, the first stage of training on large amounts of text.
  • Anthropic’s system card for Claude Mythos 5.2, the technical report published with the model, says only that it “does not double the internal rate of overall aggregate progress in AI research and development.” The authors note that Mythos 5.2 does not seem to be Model 4.

The authors still mostly trust METR. Before the assessment, its staff expected advanced AI no sooner than the authors did, so the result went against what they expected to find.

The authors argue for a pause

If SAR arrives in 2027, the authors write, “an industry-wide pause on AI R&D should be the top priority for the AI industry, the US government, and the entire world.” They judge that companies “are not on track to align a SAR,” meaning they cannot yet make sure it pursues the goals its builders intend. They also worry about security. “If we train SAR in the immediate future, the algorithms will probably leak to China, and maybe the model weights too.”

Conditional on no slowdown

Every date in this post assumes nobody slows AI development on purpose more than in the recent past. A pause of the kind the authors call for would move all of them.

Two pieces of evidence are due soon. The authors expect the METR team without inside access to publish its numbers “in the next couple of days.” Anthropic’s system card says the company “will discuss overall trends in acceleration in a future report.”

Further reading

AI forecasting AI research AI safety superintelligence