---
title: "A.I. Forecasters Now Tie the Best Humans, on Scorecards That Can Flatter Machines"
description: "ForecastBench and the Metaculus Cup show a tie, but a 2024 human baseline, market prices in the prompt and bot-friendly scoring shape the result."
author: "rews special report"
published: 2026-11-20T12:38:46.733Z
modified: 2026-11-20T13:52:51Z
url: https://rews.cc/a/ai-forecasters-have-caught-the-superforecasters-and-the-hard-a877a9
language: en
type: special report
tags: ["ai", "forecasting", "economy", "politics", "tech"]
publisher: "Rews (https://rews.cc)"
---

# A.I. Forecasters Now Tie the Best Humans, on Scorecards That Can Flatter Machines

*ForecastBench and the Metaculus Cup show a tie, but a 2024 human baseline, market prices in the prompt and bot-friendly scoring shape the result.*

Special report · By rews special report · November 20, 2026 · https://rews.cc/a/ai-forecasters-have-caught-the-superforecasters-and-the-hard-a877a9

## In brief

- In July, FRI said several A.I. systems were statistically tied with human superforecasters on its ForecastBench benchmark
- Most ForecastBench entrants score worse on market questions than they would by copying the market price they are shown
- ForecastBench’s human baseline dates from July 2024 and shares no questions with those the newest models answered
- A bot won the summer Metaculus Cup, but pros narrowly led the bots in Metaculus’s controlled spring FutureEval
- Superforecasters returned to ForecastBench for an October rematch, and results await question resolution

In January, the Forecasting Research Institute predicted that artificial intelligence would match the best human forecasters in November 2026. By the institute’s own leaderboard the machines got there early, in a statistical tie that it is now replaying against a fresh round of human forecasts.

The claim that they have arrived has spread fast. [In a post on Astral Codex Ten](https://astralcodexten.com/p/the-ai-forecasting-agenda), Scott Alexander wrote that A.I. forecasters “now equal or beat top human performance at traditional forecasting questions.” He pointed to ForecastBench, the institute’s benchmark, and to the Metaculus Cup, a public tournament. The leaders, he wrote, are specialized systems from startups such as FutureSearch, Mantic and Preseen, and general models like ChatGPT and Claude are about six months behind them.

The record behind those results supports a tie, but a narrower one than the word “superforecaster” suggests. Leaked training data, the flaw that sank the first claims of superhuman A.I. forecasting in 2024, has been largely designed out of the live contests. What distorts the results now is elsewhere. ForecastBench scores today’s machines against human forecasts collected in July 2024, on questions the humans never saw. On its market questions the machines are handed the crowd’s current price, and most entries do worse than they would by copying it. The Metaculus Cup rewards forecasters who update early and often, and a bot never needs to sleep.

Metaculus’s FutureEval is the one contest built as a controlled head-to-head. There, a team of professional forecasters still edged the best bots this spring, by a margin too small to tell apart from chance. A year earlier the bots had trailed by about 20 points.

![Two magnifying glasses rest on a blank decorative certificate on a wooden table, overlooking green fields and wind turbines.](https://rews.cc/img/046fec54ac02331cca54b744787723418e390135.webp)

## How the scoring works

ForecastBench went live in September 2024. It was built by a team that included Ezra Karger and Philip Tetlock, the institute’s president and the psychologist whose research popularized the term superforecaster. [The launch paper](https://arxiv.org/abs/2409.19839) drew on a pool of 1,000 automatically generated questions and found that expert human forecasters beat the best language model, with a p-value below 0.001. The benchmark’s central rule was that it asks only about events that have not yet happened, so no model can have read the answer in training.

Every two weeks the system samples 500 questions for models and 200 for humans, [according to its documentation](https://www.forecastbench.org/docs/). Dataset questions are generated from templates applied to time series from ACLED, DBnomics, FRED, Yahoo Finance and Wikipedia. Market questions come from Manifold, Metaculus, Polymarket and the RAND Forecasting Initiative. The institute has said the market kind arguably tests human-style judgment better, because those questions “often require judgment about novel, one-off events.”

Answers are graded with the Brier score, the squared gap between a stated probability and what actually happened, where lower is better. In March the institute [introduced the Brier Index](https://forecastingresearch.substack.com/p/introducing-the-brier-index), which turns that into a scale from 0 to 100. A forecaster who says 50 percent every time scores 50, and perfect foresight scores 100. In a column first published in the Financial Times on Sept. 16, [Tim Harford reported](https://timharford.com/2026/10/ai-can-forecast-the-future-should-we-let-it/) that the typical superforecaster scored 68.9 and the best specialized A.I. systems 68.8. The average human forecaster scored 62.8, and untuned general-purpose models just above 61. Such a benchmark, he wrote, is “a lot better than guesswork and a lot worse than omniscience.”

![A brass lever and mechanical device sit beside railway tracks that fork toward a leafy green tree and a bare tree in a grassy landscape.](https://rews.cc/img/a1884c42af1fce4246b239e3e10a4212fb6830b6.webp)

## A trend line that held

On Jan. 29 [the institute reported](https://forecastingresearch.substack.com/p/llms-are-closing-the-gap-on-human) that superforecasters still held first place, 0.017 Brier points ahead of the best models. It reckoned that gap was worth about a year of progress: Claude 3.5 Sonnet, released in October 2024, had scored 0.117, and a preview of xAI’s Grok 4.20, run a year later, scored 0.102. Tied with Grok for second place was an ensemble from a startup called Cassi, labeled ensemble\_2\_crowdadj. The institute then extended the line. “If this trend continues, we extrapolate LLM-superforecaster parity in November 2026,” it wrote, giving a 95 percent confidence interval of January 2026 to November 2027.

The line arrived early. In July the startup Lightning Rod Labs [wrote](https://blog.lightningrod.ai/p/ai-reaches-human-superforecaster-range) that its Foresight-v3 model had reached fourth place with an overall Brier Index of 67.7. That put it behind the superforecaster median of 69.3 but inside the humans’ confidence interval, along with systems from Google and Cassi. On July 16 the institute [said parity had likely been reached](https://forecastingresearch.substack.com/p/ai-models-have-likely-reached-parity). Cassi’s system was on top, and entries from xAI and Google DeepMind were also statistically indistinguishable from the superforecasters. On market questions, Cassi’s system ranked above the superforecaster median, which the institute called “the first time that any model has achieved this.” It announced the finding on X as a reply to Mr. Alexander’s July essay, [“The AI Superforecasters Are Here.”](https://www.astralcodexten.com/p/the-ai-superforecasters-are-here)

> In his recent blog post, @slatestarcodex wrote that “the AI superforecasters are here.”
>
> We run ForecastBench—a leading AI forecasting benchmark and the only one to compare AI forecasters to human superforecasters. So, how close are current AI systems to superforecasters?
>
> For…
>
> — **Forecasting Research Institute** @Research\_FRI on X · [July 16, 2026](https://x.com/Research_FRI/status/2077772870655131951)

The fine print mattered. Cassi’s one-sided p-value against equal accuracy was 0.41, so the data could not rule out a dead heat, and the institute wrote that the results were “more consistent with superforecaster parity than with outperformance.” On the faster preliminary board, which counts only dataset questions, 17 submissions ranked above the superforecasters. Shortly before, Google DeepMind’s “green-tree” had been the only one.

## The 2024 claims

Superforecasters came out of a government contest. From 2011 to 2015, the U.S. intelligence research agency IARPA ran a geopolitical forecasting tournament. The Good Judgment Project, led by Mr. Tetlock and colleagues at the University of Pennsylvania, won it by finding the top 2 percent of its volunteers and putting them on teams. Good Judgment, the company that grew out of the project, [cites an unpublished 2015 study](https://goodjudgment.com/resources/the-superforecasters-track-record/superforecasters-vs-the-icpm/) that found its forecasters more accurate than the intelligence community’s internal prediction market, whose analysts could read classified material. The margin usually quoted is about 30 percent.

![Five people assemble a giant telescope on a hill overlooking a winding river and the rising sun.](https://rews.cc/img/3c0e0ac0b4117fa67d12b5e500ef991a1bad4ea2.webp)

Machines took a run at that record in 2024 and fell short. A NeurIPS paper by Danny Halawi, Fred Zhang, Chen Yueh-Han and Jacob Steinhardt [described a retrieval system built on GPT-4](https://papers.neurips.cc/paper_files/paper/2024/file/5a5acfd0876c940d81619c1dc60e7748-Paper-Conference.pdf) that scored 0.179 across 914 test questions. The human crowd scored 0.149. That November, a study in *Science Advances* led by Philipp Schoenegger, with Mr. Tetlock among its authors, [found that the median forecast of 12 models](https://www.science.org/doi/10.1126/sciadv.adp1528) was statistically indistinguishable from a crowd of 925 human forecasters. The test covered only 31 questions.

Then came a louder claim. In September 2024 the Center for AI Safety released FiveThirtyNine, a bot built on GPT-4o, alongside a report by Dan Hendrycks and three colleagues titled “LLMs Are Superhuman Forecasters.” Mr. Alexander tested it on questions dated after November 2023 and [found it did much worse](https://www.astralcodexten.com/p/mantic-monday-91624) than the Metaculus average. Mr. Hendrycks defended the work. On X he replied to a user whose question about the bot had brought back results about FiveThirtyEight, the polling site.

> FiveThirtyNine is not in its pretraining distribution nor in any news articles. It doesn’t have access to the twitter API.
>
> — **Dan Hendrycks** @hendrycks on X · [September 9, 2024](https://x.com/hendrycks/status/1833186002468540478)

The sharpest rebuttal came from inside the industry. Researchers at FutureSearch [reviewed four such papers on LessWrong](https://www.lesswrong.com/posts/uGkRcHqatmPkvpGLq/contra-papers-claiming-superhuman-ai-forecasting) and disclosed their conflict of interest: if simple wrappers around a language model could forecast at a superhuman level, much of their own work would be obsolete. They found that the gap between the Halawi system and the crowd was larger than the paper’s own threshold for a large margin. The equivalence test in the silicon-crowd study was so loose that it would have rated the human crowd equal to a forecaster who always says 50 percent. And FiveThirtyNine’s retrieval leaned on Google’s date filters, which are known to let later information through.

> These papers do not communicate their results carefully enough, shaping public perception in inaccurate and misleading ways.
>
> — FutureSearch researchers, in a 2024 LessWrong post, [lesswrong.com](https://www.lesswrong.com/posts/uGkRcHqatmPkvpGLq/contra-papers-claiming-superhuman-ai-forecasting)

Two years later, FutureSearch’s own bot finished fifth in the summer Metaculus Cup. Two humans finished above it and every other human below. Its chief executive, Dan Schwarz, shared Mr. Alexander’s July essay on X.

## Leakage, then and now

The fullest catalog of the leakage problem appeared in May 2025, when Daniel Paleka, Shashwat Goel, Jonas Geiping and Florian Tramèr [posted “Pitfalls in Evaluating Language Model Forecasters.”](https://arxiv.org/abs/2506.00723) Backtesting means grading a model on questions that have already resolved, and they argued it lets the future leak in several ways. Search engines’ date filters fail. Retrieval models are trained on later text. Nobody is quite sure when a model’s training data ends. And a model asked about the past can sometimes work out the outcome by logic alone. The authors urged caution about claims that such systems “match or exceed human performance.”

ForecastBench and the Metaculus tournaments avoid most of that by asking only about the future, because a question that has not resolved cannot be in anyone’s training set. Backtests are harder. Zeyu Zhang and Bradly Stadie, in a paper [published in *Transactions on Machine Learning Research* this month](https://arxiv.org/abs/2608.02985), argue that the standard contamination check, which compares a model’s scores before and after its training cutoff, cannot tell leakage from skill. Four flagship models failed that check even though every scored question resolved after their cutoffs, because models know more about the months just before the cutoff. “Recency mimics leakage,” the authors wrote. Measured against an outside reference, five models whose naive scores had looked suspicious showed no detectable leakage at the boundaries tested. Backtests remain usable, they concluded, but “need one defensible reference.”

## The price in the prompt

On market questions, the models that ForecastBench runs itself get the crowd forecast as context, priced 10 days before the due date, [according to the leaderboard’s notes](https://www.forecastbench.org/leaderboards/). Outside teams can add whatever context they choose. In an April post, Good Judgment [looked at what that does](https://goodjudgment.substack.com/p/what-forecastbench-doesnt-measure). GPT-4.5’s forecasts had “a correlation of 0.994 with provided market forecasts,” it wrote, and on 26 of 122 questions the model submitted the market’s number exactly.

![A sprawling tree bears hundreds of small wooden doors, with an open doorway in its trunk revealing a sunlit landscape.](https://rews.cc/img/0c7569fa542fa5906a9ad4beec3bb1ef51507148.webp)

In August, a contributor posting as Dominus on the Effective Altruism Forum [worked through ForecastBench’s open-source scoring code](https://forum.effectivealtruism.org/posts/8YuG9EuhyYqaNBt7F/the-copier-line-what-forecastbench-s-market-scores-actually) and drew what the post called a copier line: the score a forecaster would get by submitting the price it was shown. By the post’s count, 442 of 527 forecasters on ForecastBench scored worse than that line as of Aug. 11. The post also found that missing forecasts are filled in with the market price, and that the humans had the price too, on all 90 of their market questions.

Kevin Murphy, a computer scientist at the University of British Columbia, got a similar result in an April paper. He built a baseline with no language model in it: it returns the market price on market questions and a historical prior on dataset questions. It scored 81.5 on the market half of the index. Other methods on the leaderboard were “statistically indistinguishable from simply returning the market price,” he wrote. His own system beat the baseline by 3.7 points.

![A pink mechanical fish swims among a school of teal fish above coral in a textured underwater scene.](https://rews.cc/img/84bc833e6586bb2458a7f63383c843962392836f.webp)

Blending with the crowd helps the machines. In January, Cassi’s crowd-adjusted ensemble beat its plain version by about 0.01 Brier points. A November 2025 technical report on the AIA Forecaster, by Rohan Alur and 13 co-authors, [said the system matched superforecasters on ForecastBench](https://arxiv.org/abs/2511.07678). On a harder benchmark built from liquid prediction markets, it did worse than the market consensus. Combined with the market, though, it beat the market alone.

None of this is cheating. The superforecasters had the same price in front of them, and deferring to a deep market is a defensible strategy. It does change what the market-question milestone measures. When Cassi passed the superforecaster median in July, both sides were adjusting the same starting number, and most of the field could not beat that number at all. The AIA result points the other way: a model that adds information to a market, even while trailing it, is doing more than copying.

## The 2024 humans

The human side of the ForecastBench comparison has not changed since July 2024, when the institute last surveyed its superforecasters. To compare forecasters who answered different questions, the rankings use a two-way fixed-effects model that adjusts for how hard each question was. By January, the original human sample and the newest models shared “zero overlapping questions,” the institute wrote. In July it flagged the problem itself.

> Our comparison relies on a statistical extrapolation that grows less reliable over time.
>
> — Forecasting Research Institute, in its July 16 post on parity, [forecastingresearch.substack.com](https://forecastingresearch.substack.com/p/ai-models-have-likely-reached-parity)

Mr. Harford put it more plainly: “ForecastBench is not quite comparing like with like.” Some questions are easier than others, and the humans and the machines answered different ones.

Good Judgment, which sells superforecasters’ services, made the fullest case against the benchmark [in a Feb. 25 post](https://goodjudgment.substack.com/p/what-superforecasters-actually-said) responding to an institute survey. The median superforecaster in that survey expected A.I. to beat the benchmark by 2028, earlier than either the median expert or the median member of the public. Respondents also pointed to design features that could hand the machines an early win.

> The questions resolved so far skew toward short-horizon, data-rich topics where AI has structural advantages.
>
> — Good Judgment, in a February post on its Substack, [goodjudgment.substack.com](https://goodjudgment.substack.com/p/what-superforecasters-actually-said)

Several systems enter every two weeks, and one respondent wrote that “given enough evaluations, eventually one will fall under this mark by chance.” The statistical point holds. The July announcement named a handful of systems tied with the humans, and ForecastBench had scored 527 forecasters by Dominus’s count. The same February post said superforecasters were almost 50 percent more accurate on market questions, at “0.40 vs. 0.59” against the nearest A.I. entrant. Good Judgment’s April post gave that gap as roughly 0.039 against 0.059, the scale the leaderboard uses, which suggests the February figures dropped a zero. By July, Cassi had passed the human median on those questions anyway.

The April post added that the benchmark asks only yes-or-no questions and leaves out the teaming and aggregation methods Good Judgment uses. The institute assigned three superforecasters per question and participation averaged about eight, the company wrote, while a basic Good Judgment engagement typically uses 16 or more. The human benchmark was a small crowd, combined simply.

## The Metaculus Cup

The Metaculus result needs no statistical bridge. In the summer Cup, a bot named laertes finished first overall, ManticAI second and FutureSearch fifth, [Metaculus announced](https://www.metaculus.com/notebooks/45734/announcing-the-metaculus-cup-summer-2026-winners/). That settled the site’s own question on whether a bot would beat every human that season: it resolved yes. Bots cannot win prizes or medals, but they answer the same questions as the humans. On X, Metaculus credited Jeffrey Liang, who built laertes.

> For the first time, an AI forecaster has won the seasonal Metaculus Cup.
>
> The Economist has covered the story; link in replies. Congratulations to Jeffrey Liang for building laertes, the winner in question.
>
> — **Metaculus** @metaculus on X · [September 17, 2026](https://x.com/metaculus/status/2100642327526052022)

In his July essay, Mr. Alexander had written that humans took first and second in the spring Cup, with Preseen’s bot third, and that Metaculus forecasters gave a bot a 15 percent chance of winning in the summer. The Economist described Mr. Liang as “a self-described polymath who lives in Texas,” and he spent fewer than 150 hours and a few thousand dollars on compute and data, [Mr. Alexander wrote in October](https://www.astralcodexten.com/p/links-for-october-2026). In September Mantic, the runner-up, [raised $25 million in seed funding](https://techstartups.com/2026/09/18/british-ai-startup-mantic-raises-25m-to-build-superhuman-ai-forecasting-after-metaculus-win/) in a round led by Radical Ventures.

The way the Cup keeps score helps the machines. Metaculus counts every forecast for as long as it stands, and [it has said](https://www.metaculus.com/notebooks/20027/scores-and-medals-trade-offs-and-decisions/) it uses time-averaged scores “because they incentivize making updates.” It acknowledged the cost in the same document: “If a user forgets to update then they may receive a poor score,” which “makes it harder to measure and reward superior judgment.” A bot does not forget. The community prediction is hidden for a short period after a question opens and visible after that, [according to Metaculus’s FAQ](https://www.metaculus.com/faq/).

Mr. Schwarz made the case for multi-agent forecasting systems in a talk posted on YouTube this summer.

[![AI Superforecasters Are Here: FutureSearch CEO Dan Schwarz](https://i.ytimg.com/vi/dvCHht5fNus/hqdefault.jpg)](https://www.youtube.com/watch?v=dvCHht5fNus)

FutureSearch’s chief executive, Dan Schwarz, describes how his company’s A.I. forecasting systems compare with human forecasters, in a talk posted on YouTube. Video: AI in the AM · YouTube

Metaculus built FutureEval to remove those advantages. In the spring round, which ran from Jan. 7 to April 15, 173 bots forecast on 297 scored questions. Ten Metaculus Pro Forecasters answered 99 of them with their forecasts hidden so the bots could not copy them, and both sides were scored at a single moment rather than averaged over time. [In the results](https://www.lesswrong.com/posts/wZBbDqzfBjYG58CxK/futureeval-spring-results-pros-beat-bots-but-the-gap-is), written up by Ben Wilson and John Bash of Metaculus, the pro team led the top-10 bot team by 1.25 points per question. The 95 percent confidence interval ran from minus 2.37 to plus 4.87, and the one-sided p-value was 0.247.

In the second quarter of 2025 the bots had trailed by about 20 points. This spring, nine of the 10 pros still finished above every individual bot. The bots did worse than the pros on yes-or-no and multiple-choice questions and better on numeric ones. The authors warned against reading the result as a draw.

> This should not be confused with a tie since we don’t have enough data to conclude a head-to-head score of 0.
>
> — Ben Wilson and John Bash of Metaculus, on the spring FutureEval results, [lesswrong.com](https://www.lesswrong.com/posts/wZBbDqzfBjYG58CxK/futureeval-spring-results-pros-beat-bots-but-the-gap-is)

## Weighing the claim

Read together, the records support a narrow version of the headline. On short-horizon yes-or-no questions, the best A.I. systems are now statistically indistinguishable from superforecasters on ForecastBench and from Metaculus’s professionals in FutureEval. Neither record shows them clearly ahead. The leakage that undid the 2024 papers does not explain the live results. Other things lift them: a human baseline frozen in 2024 and scored on different questions, a market price most entries cannot improve on, hundreds of systems tested against one human median and a tournament format that rewards stamina.

The best-argued prediction against machine superiority came from Arvind Narayanan and Sayash Kapoor of Princeton, in their April 2025 essay [“AI as Normal Technology.”](https://www.normaltech.ai/p/ai-as-normal-technology)

> We predict that AI will not be able to meaningfully outperform trained humans (particularly teams of humans and especially if augmented with simple automated tools) at forecasting geopolitical events (say elections).
>
> — Arvind Narayanan and Sayash Kapoor, Princeton computer scientists, [normaltech.ai](https://www.normaltech.ai/p/ai-as-normal-technology)

Nothing published so far refutes that. Their yardstick is trained human teams with tools, and the human side of ForecastBench is a median of individuals who forecast two years ago. Mr. Alexander called their prediction “a great test case!” and set his own deadline: “We’ll start to have a good idea which world we’re in within a year; after two years, the answer should be decisive.”

## What comes next

The institute announced the rematch on X in September: superforecasters would return to ForecastBench in October to take on the top A.I. systems directly.

> AI models now match superforecasters on ForecastBench, our LLM forecasting benchmark. (Although not by a statistically significant margin).
>
> This October, we’re running the rematch: superforecasters return to the benchmark, and the top AI forecasters can take them on.
>
> 🧵 In this…
>
> — **Forecasting Research Institute** @Research\_FRI on X · [September 22, 2026](https://x.com/Research_FRI/status/2102411966103093386)

Good Judgment [wrote on Oct. 30](https://goodjudgment.substack.com/p/what-the-forecastbench-rematch-still) that the institute had finished collecting the new human forecasts, and said it welcomed a refreshed baseline. The post’s subtitle read: “Why we’ll read the results the same way, whoever comes out ahead.” The questions still have to resolve, and new systems appear on the tournament leaderboard only 50 days after they submit. The institute has also said it will add quantile questions and update its dataset questions this fall.

Mr. Alexander’s newer post looks past the tie to harder tests. About a year ago the institute found that models did markedly worse on conditional questions, and the researchers he spoke with expect benchmark results within a few months. Questions about A.I. itself are where forecasters keep missing. This summer, top human and A.I. forecasters put the chance that A.I. would solve a Millennium Prize Problem this year at 5 percent. In the institute’s biorisk work, the median superforecaster expected models to match a top team of virologists on troubleshooting tests by 2034, and models reached that mark as of April 2025.

The next scoreboard closes sooner. The fall Metaculus Cup [ends Jan. 1, 2027](https://www.metaculus.com/tournament/metaculus-cup-fall-2026/). Asked [how many bots will finish in its top 10](https://www.metaculus.com/questions/45619/how-many-bots-will-finish-in-the-top-10-of-the-metaculus-cup-fall-2026/), Metaculus forecasters put the figure at 5.1.

## See also

- [Year AI Beats Superforecasters on ForecastBench](https://leap.forecastingresearch.org/forecasts/forecastbench-performance) — leap.forecastingresearch.org · FRI expert panel forecasts of when A.I. would overtake superforecasters
- [FutureEval](https://www.metaculus.com/futureeval/) — metaculus.com · Metaculus's running record of bots against pro forecasters
- [AI Forecasting in 2026: What 11 Analyses Say](https://www.metaculus.com/notebooks/43363/ai-forecasting-in-2026/) — metaculus.com · Metaculus roundup of competing assessments of A.I. forecasting
- [Wisdom of the silicon crowd (full text)](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11800985/) — ncbi.nlm.nih.gov · Free full text of the 2024 study on combining forecasts from many models
- [Superforecasters: A Decade of Stochastic Dominance](https://goodjudgment.substack.com/p/superforecasters-a-decade-of-stochastic) — goodjudgment.substack.com · Good Judgment's account of the superforecaster track record
- [AI is Not Normal Technology](https://www.lesswrong.com/posts/F2mwKxFRhjqizA89p/ai-is-not-normal-technology) — lesswrong.com · A rebuttal to Narayanan and Kapoor, including their forecasting prediction
- [Metaculus Tournament Rules](https://www.metaculus.com/tournament-rules/) — metaculus.com · The rules on bots, A.I. assistance and prize eligibility
