Metaculus hosts an AI 2027 Tournament with 19 forecasting questions derived directly from the AI 2027 scenario. The tournament launched in June 2025 and tracks crowd predictions on technology, economics, politics, and societal impact through 2028 and beyond.

This page compares the AI 2027 scenario’s claims, the Metaculus crowd forecasts, and our independent tracker assessments.

How to Read This Comparison

Different measurement systems: AI 2027 makes narrative claims (“X will happen by Y”). Metaculus uses probability percentages or numerical estimates. Our tracker uses status labels (confirmed, on-track, behind, etc.). We bridge these below, but they’re fundamentally different — probabilities capture uncertainty; status labels capture current evidence.

Rough mapping:

Our StatusApproximate Probability Equivalent
Confirmed>90% (already happened)
Ahead>80% and faster than expected
→ On Track60–80%
Emerging30–60%
Behind20–50% (direction right, timing wrong)
Not Yet TestableNo probability meaningful yet

Side-by-Side: Key Tournament Questions

Technology & Capabilities

Neuralese Recurrence

SourceAssessment
AI 2027Non-text reasoning (“neuralese”) deployed in early 2027 — a key capability unlock
Metaculus crowdMedian: ~Jan 2028 for public release (Q38598)
Our trackerEmerging (conf. 0.35) — predicted early 2027
Jan Wehner forecastOct 2026 (more aggressive than crowd)

Analysis: The crowd is more conservative than the scenario by about a year. Our tracker agrees this is still early — no frontier lab has shipped neuralese recurrence publicly. Meta published research in 2024, but production deployment remains uncertain.

VideoGameBench Performance

SourceAssessment
AI 2027Implies rapid capability gains across diverse tasks
Metaculus crowdMedian: ~58% by Jan 2028 (range: 34%–83%) (Q38609)
Our trackerNot directly tracked — closest: agent autonomy predictions
Jan Wehner forecast80% (more aggressive than crowd)

Analysis: Starting from <1% in mid-2025, the crowd expects strong but incomplete progress. This aligns with the broader pattern: capabilities improve dramatically but don’t hit the most aggressive targets.

AI-Authored Scientific Papers

SourceAssessment
AI 2027AI increasingly contributes to research, approaching researcher parity
Metaculus crowd~78% chance of AI-authored paper at NeurIPS/ICML/ICLR before 2028 (Q estimated)
Our trackerAI for AI research confirmed (conf. 0.85)
Jan Wehner forecast84%

Analysis: Broad agreement. AI is already contributing meaningfully to research. The specific milestone of a published AI-authored paper is more about institutional acceptance than capability.

OpenAI Self-Improvement Risk Levels

SourceAssessment
AI 2027AI models gain enough capability to contribute to AI R&D, creating recursive loops
Metaculus crowd”High” risk level: median ~May 2027; “Critical”: much later (Q estimated)
Our tracker→ On Track for 1.5× R&D multiplier (conf. 0.70)
Jan Wehner forecast”High” by May 2027; “Critical” by July 2032

Analysis: This is a crucial question for the takeoff timeline. The crowd and our tracker broadly agree that meaningful AI self-improvement is approaching but not yet at crisis levels. The AI 2027 scenario is more aggressive on timing.


Economics & Compute

Total AI Compute Capacity (Jan 2028)

SourceAssessment
AI 2027Massive compute scaling — trillions in infrastructure investment
Metaculus crowdActively forecasted (Q38419) — specific numerical estimate
Our trackerInfrastructure investment confirmed (conf. 1.0); 10²⁸ FLOP training run behind (conf. 0.70)

Analysis: No disagreement that compute is scaling enormously. The debate is about how fast — and whether the largest training runs match the scenario’s timeline.

Compute Concentration

SourceAssessment
AI 2027”OpenBrain” (based on OpenAI/Google) dominates; others 3–9 months behind
Metaculus crowdHighest % of compute held by one org on Jan 2028 (Q38393)
Our trackerLab gap narrower than predicted (0–2 months, not 3–9)

Analysis: The scenario predicted more concentration than we’re seeing. Multiple labs remain competitive. The Metaculus question will help quantify how concentrated compute actually becomes.


Geopolitics & Governance

US-China AI Agreement

SourceAssessment
AI 2027No meaningful cooperation — instead an accelerating race
Metaculus crowd~4–5% chance of formal US-China agreement before 2029 (Q estimated)
Our trackerExport controls confirmed impacting Chinese compute (conf. 0.80)
Jan Wehner forecast4%

Analysis: Strong agreement across all sources. A formal US-China AI agreement is extremely unlikely. The competitive dynamic AI 2027 described is playing out.

AI as “Most Important Problem” (Gallup)

SourceAssessment
AI 2027Growing public concern, anti-AI protests, political salience
Metaculus crowdMedian ~19% by Jan 2028 (Q38407)
Our trackerAnti-AI protest emerging (conf. 0.50); Labor market impact ahead of predicted
Jan Wehner forecast3% (dramatically lower than crowd)

Analysis: This is one of the biggest disagreements within the forecasting community. The Metaculus crowd seems very aggressive at 19% — for reference, “inflation” peaked at ~6% on Gallup. Jan Wehner’s 3% seems more calibrated to historical polling patterns. Our tracker sees growing public concern but hasn’t committed to a specific percentage.


Category Scorecard

CategoryAI 2027 ScenarioMetaculus Crowd LeanOur Tracker Assessment
Coding automationVery aggressive (superhuman by Mar 2027)Moderate — expects progress but slowerMixed — agents confirmed, benchmarks behind
AI research accelerationCentral claim — recursive improvement by 2027Moderate timeline (~2027–2028 for key milestones)On track for 1.5× but timing uncertain
Compute scalingMassive and rapidScaling confirmed, specifics debatedConfirmed but largest runs behind
Revenue/valuation$55B revenue, $2.5T valuation by end 2026Not directly tracked in tournamentBehind on valuation, emerging on revenue
Public concernHigh — protests, political salienceHigh — 19% “most important problem”Emerging signals, possibly ahead
Geopolitical raceIntense US-China competition, no cooperationStrong agreement — 4–5% chance of agreementConfirmed competition dynamic
Neuralese / novel architecturesKey unlock in early 2027Conservative — median ~2028Emerging, low confidence
Safety / alignmentIncreasingly urgent, partially ignoredModerate concern acknowledgedOn track — ASL-3 upgrades, model specs

Who Is More Aggressive?

QuestionMost AggressiveLeast Aggressive
Overall AI progress paceAI 2027 scenarioMetaculus crowd
Public concern about AIMetaculus crowd (19%)Jan Wehner (3%)
Neuralese deploymentAI 2027 (early 2027)Metaculus crowd (~2028)
US-China cooperationAll agree: very unlikely
Coding automationAI 2027Our tracker (benchmarks behind)
AI self-improvement riskAI 2027Metaculus crowd (more gradual)

Where All Three Sources Agree

These are the highest-confidence areas — when the scenario, crowd wisdom, and our evidence-based tracker all point the same direction:

  1. AI coding tools are transforming software development — universally acknowledged
  2. US-China AI cooperation is extremely unlikely — all sources below 5%
  3. Compute infrastructure is scaling massively — no disagreement
  4. AI agents exist but reliability remains a challenge — all confirm
  5. AI is increasingly contributing to AI research — direction agreed, pace debated

Where They Disagree Most

  1. Public salience of AI — 3% to 19% range on Gallup question
  2. Neuralese timeline — 1+ year gap between scenario and crowd
  3. Speed of coding automation — scenario’s SWE-bench targets missed
  4. Self-improvement timeline — scenario expects earlier recursive loops than crowd

Methodology Note

Metaculus forecasts are live crowd predictions that update continuously. The values cited here were gathered in March 2026 from search results, the Jan Wehner analysis (August 2025), and Metaculus question descriptions. For the latest crowd forecasts, visit the Metaculus AI 2027 Tournament directly.

Our tracker assessments are evidence-based judgments updated as new information arrives. See our Methodology page for how we assign status labels and confidence scores.


Related:

The Metaculus tournament originally had 16 questions (as of August 2025) and has since expanded to 19. Not all questions map directly to our tracked predictions.