Which AI 2027 Predictions Came True?
What has happened, and what remains unresolved
Several developments described in AI 2027 are visible in the public record: useful but unreliable agents, premium agent subscriptions, growing infrastructure investment and AI used in AI research. More specific claims about complete research automation, numerical financial targets and geopolitical implementation need different evidence.
The live prediction list contains the current status counts. A count of confirmed dossiers is not a calibrated accuracy score: the claims have different scopes, deadlines and degrees of dependence, and this selection is not exhaustive.
Developments with direct support
- Coding agents: about 70% of agent users in the 2025 survey reported shorter task times; METR’s early-2025 controlled study found a 19% slowdown. The tracker rates practical usefulness in some professional workflows Confirmed (85%), not indispensability for most programmers.
- Useful but unreliable agents: products and documented projects show usefulness; evaluations and incidents also show consequential failures. These findings do not supply a universal reliability rate.
- Premium agent pricing: historical subscription tiers support the price observation. Price alone does not identify the best agent.
- Infrastructure investment and data-centre buildouts: disclosures establish substantial activity, with plans distinguished from completed capacity.
- AI used for AI research: named organizations document contributions. This does not resolve how much overall research progress was accelerated.
- Defense contracting: four July 14, 2025 awards with $200M ceilings each precede the late-2026 window. Ceilings are not expenditure and contracts are not completed operational integration.
A milestone can be achieved late
OSWorld and SWE-bench illustrate why the benchmark version, evaluation setup and observation date matter. Crossing a threshold after the original window does not make the original timing correct. Vendor and independently standardized results also need distinct labels.
Each dossier explains the numerical outcome separately from timing and the strength of its evidence. Later progress does not erase an earlier missed window.
Promising evidence can leave the original claim open
The revenue assessment records $16.286B in primary, unaudited H1 recognized revenue and a separately reported $16.8B for Q3. Its September reconstruction’s roughly $47.9B annual case is illustrative, not realized revenue or confirmation of the $45B test. Cyber capability remains Ahead (85%): Astra’s reported 86 of 226 FrontierCyber tasks versus Sol’s 34 sit alongside separate hardened-target failures. Near-best-human reliability and thousands-of-copies effectiveness remain unresolved.
The METR dossier records a 105-day doubling fit for 2024-onward models at 50% success. Its separate 129-day fit starts in 2023; neither establishes years-long tasks at 80% reliability. For China-US comparisons, Leading Chinese AI lab ~6 months behind US frontier remains On Track (75%): the dossier’s June 30, 2026 ECI comparison reconstructs a gap of about 6.6 to 7.4 months, with benchmark uncertainty and internal capabilities unresolved. Chinese domestic AI chips 3 years behind US-Taiwan remains On Track (80%): Epoch’s 3-to-4-year per-chip lag estimate supports the scenario’s approximately three-year claim, while production and deployment require separate comparisons. The reality guide gives the dated primary-source figures and their limits.
The same discipline applies to research acceleration. Increasing automation matters, but accepted research progress needs a credible comparison with work without AI.
How to read an unresolved assessment
An unresolved claim may have encouraging precursor evidence, insufficient measurement or limited public visibility. It does not automatically mean failure. A future deadline does not prevent assessing a qualifying early result, and a missed deadline should remain visible even when a later result is confirmed.
Read the full comparison with reality · Browse all dossiers · See the evidence standards