Indicators

Incidents per year in the AI Incident Database

AI harms that were publicly reported and recorded in the AI Incident Database, counted by year of the event. It is the only public longitudinal record of observed rather than projected harm. It does NOT answer whether there are more AI harms: it answers how much AI harm is being documented, which mixes real harm, media attention and the editorial team's throughput. If it rises, the first hypothesis is more attention, not more harm. The AIID itself warns that incident IDs are added after the underlying events occur and that the lag varies considerably, and that concentrations by country or company partly reflect what was investigated and processed in that period.

Figure 7.1 · incidentes · proxy indicator · annual cadence · Rising means more risk
Source [31] AI Incident Database — Database Snapshots ↗

FrontierMath tier 4 saturation

Best cumulative score on FrontierMath tier 4, the research-level problem set designed to resist. What it measures is not mathematics: it is how long an exam lasts before it stops discriminating. If a benchmark built to be impossible saturates in twenty months, the community's ability to measure lags the capability it is trying to measure, and that lag is the risk: safety-policy thresholds rest on evaluations that age faster than they are written. Series from Epoch AI's own runs; one series, one evaluator.

Figure 7.1 · % de acierto · direct measurement · irregular cadence · Rising means more risk
Source [372] Epoch AI Benchmarking Hub — benchmarks.csv ↗

GPQA Diamond saturation

Best cumulative score on GPQA Diamond, the PhD-level science question set designed to resist internet search. It is effectively saturated: the last four highs fall within each other's standard errors, meaning the exam no longer discriminates between frontier models. A saturated benchmark is not a fact about capability, it is an exhausted instrument, and quoting it anyway creates the illusion of a ceiling. Series from Epoch AI's own runs; one series, one evaluator.

Figure 7.1 · % de acierto · direct measurement · irregular cadence · Rising means more risk
Source [372] Epoch AI Benchmarking Hub — benchmarks.csv ↗

SWE-bench Verified saturation

Best cumulative score on SWE-bench Verified, which measures resolution of real software repository issues. It is the only member of the saturation family that still discriminates between frontier models, and therefore the most useful to track. Beware of comparing figures across evaluators: the 2026 AI Index states it rose from 60% to nearly 100% in a single year, while Epoch's own runs top out at 83.5%. Neither is lying: scaffolding, token budgets and scoring criteria differ. Site rule: one series, one evaluator.

Figure 7.1 · % de acierto · direct measurement · irregular cadence · Rising means more risk
Source [372] Epoch AI Benchmarking Hub — benchmarks.csv ↗

AI adoption by US firms (Census BTOS)

Share of US firms reporting they used artificial intelligence in any business function in the previous two weeks, per the Census Bureau's Business Trends and Outlook Survey. It is the best public source in this section because it is an official probability survey with explicit reference and publication dates. It serves as the denominator for much of the rest: an incident in a country where 22% of firms use AI does not mean the same as where 5% do. As of 3 May 2026 the breakdown showed 39.7% in Information, 33.9% in Finance and Insurance and about 14% in Retail Trade; by size, 37% among firms with 250 or more employees against under 20% for those with fewer than 20.

Figure 7.1 · % de empresas · direct measurement · monthly cadence · Rising means more risk
Source [225] Business Trends and Outlook Survey — National estimates (National.xlsx) ↗

Largest known training compute per year

The largest training compute published or estimated for a model in each year, per Epoch AI's notable models database. Compute is the input that best predicts capability and the only one currently regulated with a numeric threshold: the 1e25 and 1e26 FLOP marks in the EU AI Act and US executive orders are written in this unit. NOTE: it measures DISCLOSURE, not compute. 2024 appears below 2023 and that did not happen: it reflects that the model with published compute that year was an open Meta model while closed frontier labs reported nothing. Read it as 'largest known compute', never as 'largest compute', and the gap between the two widens every year labs stop reporting.

Figure 7.1 · FLOP · proxy indicator · annual cadence · log scale · Rising means more risk
Source [377] Notable AI Models — base de datos descargable ↗

Threshold · 1 × 10²⁶: Highest regulatory threshold written in FLOP. Only three known models exceeded it, all in 2025. Pre-training efficiency improves about 3x a year, so a fixed threshold admits a more capable model every year: it does not age, it melts.

Companies with a published frontier safety framework

How many AI companies impose capability thresholds on themselves with written consequences and publish them. It is the voluntary-governance indicator: it measures whether the promise exists, not whether it is kept. The count alone misleads, which is why it carries a threshold: at the May 2024 AI Seoul Summit sixteen companies committed to publishing and four more joined later; twelve published. The distance between committed and published is a better indicator than the count, because it measures failure to keep an explicit commitment and admits no optimistic reading. The twelve as of the December 2025 report are Anthropic, OpenAI, Google DeepMind, Magic, Naver, Meta, G42, Cohere, Microsoft, Amazon, xAI and NVIDIA.

Figure 7.1 · empresas · direct measurement · irregular cadence · Falling means more risk
Source [708] Common Elements of Frontier AI Safety Policies ↗

Threshold · 20: Companies that committed to publishing a frontier safety framework: sixteen at the May 2024 AI Seoul Summit plus four that joined later. The gap with those that actually published is the finding.

Task time horizon (METR)

Length of tasks a model completes with 50% success, measured in the time of an expert human with no prior context. It is the capability indicator with the longest comparable series and the best public proxy for sustained autonomy: not how fast an agent writes code, but how many hours it chains decisions without correction. METR rejects four readings: it does not measure how long an agent can act autonomously but how hard the task is; tasks are software, machine learning and cybersecurity, not 'everything intellectual'; the human reference is a newcomer to the project, not a professional with the context in mind; and real work is rarely algorithmically scorable.

Figure 7.1 · horas · direct measurement · irregular cadence · log scale · Rising means more risk
Source [711] METR Time Horizon 1.1 — benchmark_results_1_1.yaml ↗

Threshold · 16: Ceiling declared by METR. Above 16 hours, measurements are unreliable with their current task suite.

AI adoption measured by transactions (Ramp AI Index)

Share of firms in Ramp's panel paying for AI products, measured from transaction data across more than 70,000 US companies on its corporate card and payments system. Its advantage over a survey is that it is not self-reported: a company paying an AI subscription is using it. This is a contrast series, not the primary one: Ramp is a fintech selling to the same firms it measures, and its panel skews toward young, technology-focused, fast-growing companies. The 56% in its panel against the 22.4% in the Census probability sample is not a contradiction: it is the difference between a panel and a representative sample. Dates stamp the month of each monthly release.

Figure 7.1 · % de empresas del panel · proxy indicator · monthly cadence · Rising means more risk · Out of date · Last measured: 1 Aug 2026
Source [885] Ramp AI Index ↗

Generative AI services registered in China

Cumulative number of generative AI services that completed registration with the Chinese regulator, mandatory for those with public opinion attributes or social mobilisation capacity. It is an administrative count, not a measure of capability or safety: it says how many services entered the system, and nothing about how safe the models behind them are. It is published because it is the only hard number on regulatory scale that the Chinese system produces, and because since July 2026 it also covers models running on the phone itself.

Figure 7.1 · servicios · proxy indicator · irregular cadence · Rising means more risk
Source [185] 国家互联网信息办公室关于发布2025年生成式人工智能服务已备案信息的公告 ↗

Youth employment decline in AI-exposed occupations

Annualized employment change for women aged 22 to 25 in the occupations most exposed to AI —the subgroup falling the most; men of the same age fall less—, computed from the baseline and measured on real ADP payroll data by the Stanford Digital Economy Lab's Canaries in the Coal Mine dashboard. It is the indicator closest to economic harm measured in people rather than in intention surveys: a balanced sample of 25,000 firms and 4.6 million workers across more than 730 occupations, with a November 2022 baseline, the month ChatGPT launched. THE COUNTERARGUMENT COMES FROM THE AUTHORS THEMSELVES: there is no economy-wide displacement. The effect is real but concentrated in one group, and with the broadest set of controls the declines only become statistically significant in 2024. A dashboard showing a large gap in one group and zero in the aggregate is exactly what should be published.

Figure 7.1 · % anual desde noviembre de 2022 · direct measurement · monthly cadence · Falling means more risk
Source [991] Canaries Dashboard — AI Economic Indicators ↗