Which AI 2027 Predictions Came True?
The Scorecard So Far
Of the 53 predictions we track from the AI 2027 scenario, 17 have been confirmed or are running ahead of schedule as of May 2026. That’s roughly a 32% share of the tracker — but the story is more nuanced than a single number suggests.
Most confirmed predictions are qualitative (the essay described dynamics and trends correctly), while the quantitative predictions (specific benchmarks and financial targets) are more mixed. Here’s the full picture.
Confirmed Predictions (14)
These have clearly materialized within the predicted timeframe:
Infrastructure & Economics
- Massive AI infrastructure investment continues — Hyperscaler AI capex exceeded $200B in 2025, consistent with the scenario’s projections
- Massive datacenter buildouts continue — Multiple GW-scale facilities announced and under construction globally
- Agents struggle with long-horizon tasks — METR evaluations and real-world usage confirm agents fail at multi-day autonomous tasks
Agents & Capabilities
- Unreliable but useful AI agents emerge — ChatGPT agent (July 2025), Claude computer use, and similar products match the scenario description closely
- Computer-using agents marketed as ‘personal assistants’ — Multiple products launched with this framing
- Best AI agents cost hundreds of dollars per month — ChatGPT Pro ($200/mo), Claude Max subscriptions confirmed this price range
- Coding agents provide significant real-world value — Claude Code and related tools have become significant revenue-generating software products
- OSWorld benchmark reaches 65% by mid-2025 — the benchmark target was exceeded by the predicted period
Research & Training
- AI companies focus on AI-for-AI-research — All major labs have internal AI research acceleration programs
- Models shift to continuous/iterative training — GPT-4o → GPT-5 → GPT-5.1 reflect continuous model updates rather than discrete training runs
Geopolitics & Governance
- Export controls impact Chinese AI compute — US export restrictions measurably constrained China’s access to frontier chips
- Department of Defense scales up AI lab contracting — DOD contracts with Anthropic, OpenAI, and others expanded significantly
- Continued skepticism from academics and journalists — Despite rapid progress, mainstream skepticism about near-term AGI persists
Competition
- Gap between top US labs narrows to 0-2 months — The scenario predicted 3-9 month gaps; reality shows even closer competition between labs
Ahead of Schedule (3)
These are happening faster than AI 2027 predicted:
- METR time horizon doubles every 4 months — recent METR-style agent horizon evidence has moved at least as fast as the scenario’s expected pace.
- AI reaches near-best-human hacking capability — advanced AI cyber capability signals appeared ahead of the early-2027 timeline.
- Stock market impact and public backlash from AI job displacement — labor-market disruption concerns emerged earlier and more prominently than the 2026-2027 timeframe suggested in the scenario.
Behind Schedule (4)
These are moving slower than predicted — important for an honest assessment:
- SWE-bench-Verified score reaches 85% — AI 2027 predicted 85% by mid-2025; later self-reported scores may cross the numerical threshold, but comparable independent verification remains important and the target was late.
- Leading AI company reaches $3T valuation — valuations are high, but still behind the trajectory toward the end-2026 target.
- Stock market rises 30% in 2026 — market performance has not matched this aggressive prediction.
- Leading Chinese AI lab ~6 months behind US frontier — the gap appears larger than predicted, partly because Chinese labs face compute constraints.
→ On Track (10)
Progressing roughly as predicted, not yet fully confirmed:
- AI provides substantial bioweapon design help — Anthropic upgraded bio capability assessment to ASL-3; capabilities are emerging as predicted
- Global AI capex reaches $1 trillion cumulative — investment trajectory aligns with this milestone
- Chinese domestic AI chips 3 years behind US-Taiwan — gap assessment remains consistent
- AI R&D progress multiplier reaches 1.5× — coding assistance is relevant, while broader R&D uplift remains uncertain
- Frontier model variant 10× cheaper released — cost reductions are directionally consistent with the scenario, with exact comparability still tracked
- Public unaware of best AI capabilities — public/private capability gaps remain plausible and partially evidenced
- AI scores 85% on Cybench — current evidence supports the trajectory, though benchmark interpretation remains specialized
- Leading AI company reaches $45B annual revenue — frontier-lab revenue growth is tracking toward the scenario’s target
- Leading AI company reaches $40B compute costs, 6GW power, $200B capex — infrastructure and spending trajectories point in the predicted direction
- RE-Bench score reaches 1.3 — early benchmark evidence supports continued monitoring
Emerging (13) and Not Yet Testable (9)
An additional 22 predictions are either showing early signals (emerging) or haven’t reached their predicted timeframe yet. These include the most dramatic predictions in the scenario — autonomous AI researchers, superintelligence, geopolitical crises — most of which target late 2026 through 2027.
For the complete list, see our full prediction tracker.
The Honest Summary
What the scenario got right: Several qualitative claims about 2025 AI development have held up: agent emergence, coding-tool adoption, infrastructure buildout, institutional dynamics, and the narrowing gap between US labs are all strongly represented in the confirmed or ahead categories.
What’s behind: Raw compute scaling and some financial milestones. The scenario may have overweighted brute-force scaling relative to algorithmic and architectural improvements.
What’s most interesting: Agent capabilities (METR time horizons) are improving faster than predicted, while compute scaling is slower. This suggests AI 2027 may have gotten the mechanism slightly wrong but the trajectory roughly right.
Bottom line: As of May 2026, the AI 2027 scenario remains directionally relevant but uneven. The strongest public evidence supports continued close tracking, while the most dramatic takeoff and alignment-crisis claims remain unresolved.
Read more: