Claim 05 of 06
Genius for everyone, on demand.
Progress is measured by a constructed index: frontier capability against human expert benchmarks, multiplied by practical daily reach. The index has risen from 0.5 to 19.4 since 2020, against a completion threshold of 100.
Primary measure
Basis of the progress score
The single objective measure assigned to this claim, tracked from the 2020 baseline to the current value. This measure alone determines the progress percentage.
Expert-Parity Access Index
A constructed composite, not a published statistic. The capability component measures frontier model performance against human expert benchmarks and currently stands near 62 percent of expert parity. The reach component measures practical daily availability by internet access, price, and language coverage, and currently stands near 30 percent. The index is their product. Completion requires expert parity available to the entire population.
Supporting measures
Related indicators
Reported alongside the primary measure to show whether it is consistent with the wider evidence. These do not enter the progress score. Bars show relative movement only.
Recent developments
Events since 2025
Developments relevant to this claim, and their effect on the primary measure where there is one.
Completion estimate
Basis for 2034
Starting point, adjustments applied, and the resulting estimate.
This measure is a constructed composite rather than a published statistic. Its two components behave differently and are reported separately.
The capability component stands near 62 percent of expert parity. Frontier models score above 90 percent on GPQA Diamond, where domain PhDs score approximately 65 to 75 percent within their own field. On Humanity's Last Exam the leading model scored 55.5 percent in August 2026 against approximately 90 percent for human domain experts, having gained roughly 30 percentage points in twelve months.
The reach component stands near 30 percent, derived from global internet penetration near 68 percent and the share of that population with affordable frontier-tier access in a supported language.
Capability is not expected to remain the binding constraint. The Metaculus community median for AGI is November 2033, and the remaining capability gap is primarily long-horizon reliability rather than domain knowledge.
The date estimate is therefore governed by reach. Internet penetration is projected near 90 percent in the late 2030s and inference cost has fallen more than two orders of magnitude since 2022. The estimate used here is 2034 for expert parity combined with majority reach. Coverage approaching the full population falls in the 2040s.
Assessment
Arguments for and against
The strongest available case on each side, stated without weighting.
Reasons to expect completion
- Benchmark improvement has been rapid and sustained.Frontier scores on Humanity's Last Exam rose approximately 30 percentage points in twelve months, on a benchmark constructed specifically to resist saturation.
- Expert parity has already been exceeded in several domains.GPQA Diamond, competition mathematics, most competitive programming, and protein structure prediction are all above expert human performance.
- Marginal distribution cost is near zero.Serving an additional user of a trained model is inexpensive relative to training. ChatGPT alone reached approximately one billion monthly active users within four years of launch.
- Language coverage has broadened.Frontier models perform competently across dozens of languages, including several with limited digital training corpora.
- Forecast timelines have shortened substantially.The Metaculus community median for AGI moved from approximately fifty years out in 2020 to November 2033 as of February 2026.
Reasons to expect failure
- Reach is the binding constraint and improves slowly.Approximately one third of world population lacks reliable internet access, and connectivity growth is limited by infrastructure investment rather than by software.
- Availability does not imply use or benefit.Access statistics measure account activity, not the application of expert-level capability to consequential tasks.
- Reliability remains unresolved.Models produce confident errors, and the domains where expert capability is most valuable are those in which a non-expert user cannot verify the output.
- Benchmark saturation limits measurement.Humanity's Last Exam was constructed to be difficult and is being solved rapidly, which constrains how much the score indicates about general capability.
- Frontier capability is concentrated.A small number of laboratories control frontier training. Population-scale access depends on their pricing and deployment decisions.
- Cognitive capability has not historically been the constraint on outcomes.Coordination, institutional capacity, and implementation rather than availability of expertise have limited progress on most large-scale problems.
Scenarios
Most likely paths
Two sequences: the most probable route to completion and the most probable route to failure. These are structured projections, not forecasts, and neither is assigned a probability.
Most likely path to completion
The most probable path requires continuation of current trends rather than a discontinuity.
Capability reaches expert parity on general benchmarks around 2029 at the current rate of improvement. Long-horizon agentic reliability, the principal remaining gap, is addressed through training and evaluation methods over a similar period.
Inference cost continues to fall. By approximately 2031 frontier-class capability runs locally on consumer hardware, removing per-query cost and network dependency as barriers.
Reach then expands faster than capability. The marginal user is added at negligible cost, and distribution proceeds through mobile device penetration, which substantially exceeds fixed broadband penetration in low-income regions.
The largest measurable effects appear in education and in professional services in regions with low practitioner density, where the counterfactual is absence of expertise rather than substitution for an existing professional.
Majority reach combined with expert parity is reached near 2034. Coverage approaching the full population follows in the 2040s as connectivity infrastructure completes.
Most likely path to failure
The most probable failure is in the reach component rather than the capability component.
Capability continues to improve on the observed trajectory. Expert parity on general benchmarks is achieved on approximately the projected schedule.
Access stratifies by price tier. Free and low-cost tiers remain substantially less capable than frontier deployments, and the performance gap between tiers widens rather than narrows, because competitive pressure at the frontier is limited to a small number of providers.
Connectivity growth in low-income regions slows as the remaining unconnected population becomes more expensive to reach, being predominantly rural, low-density, and low-income.
Among users with access, the dominant application is consumption rather than production, following the pattern of prior consumer information technologies. Measured effects on educational attainment and output remain smaller than capability improvements alone would suggest.
Available cognitive capability increases by orders of magnitude while measured outcomes on the other five claims improve at rates closer to their historical trends.
The other claims
Open another claim page
Each claim carries its own colour, applied as the background of its page.
