Stanford's AI Index: Safety Benchmarks Lag as Incidents Surge

Every year, Stanford's Human-Centered AI institute publishes its AI Index, a systematic look at where the technology actually stands. The 2026 edition arrived this month, and it contains the kind of sober accounting that tends to get overshadowed by product announcements.

The safety numbers are not encouraging. Documented AI incidents rose to 362 in 2025, up from 233 the year before, a 55 percent increase in a single year. At the same time, the Foundation Model Transparency Index fell from 58 to 40 points. The most capable models, the report finds, are also the ones disclosing the least about how they were built and trained.

There is a fundamental tension the report names directly: improving one safety dimension often degrades another. Making a model more resistant to harmful outputs can reduce its accuracy on legitimate tasks. No clean solution to this trade-off has emerged, and the field is not moving fast enough on finding one.

On workforce effects, the report moves beyond the projections that have dominated previous editions. Employment of software developers aged 22 to 25 has dropped nearly 20 percent since 2024. The report describes this as AI job displacement having shifted from prediction to reality. That is a significant shift in framing from the hedged language of earlier years.

The geopolitical picture has also changed. Chinese and US AI models have traded top positions on major benchmarks since early 2025. Anthropic's current benchmark lead stands at just 2.7 percentage points. The US still dominates in patent impact and flagship model production, but the margin has narrowed considerably.

International AI talent migration to the US is down 89 percent since 2017, with 80 percent of that decline occurring in just the past year. Whether this reflects immigration policy, competing AI hubs abroad, or researchers choosing to stay home, the data does not say clearly. For a country that has treated talent attraction as a core AI strategy, the number is hard to dismiss.

On the environmental side, AI data center power capacity has reached 29.6 gigawatts globally, roughly the peak demand of New York City. Training Grok 4 produced approximately 72,816 tons of CO2, equivalent to annual emissions from around 17,000 vehicles.

Consumer adoption, by contrast, has accelerated faster than any previous consumer technology. Generative AI reached 53 percent global adoption within three years, a faster trajectory than the PC or the internet. The estimated value to US consumers reached $172 billion annually by early 2026.

The full report is available at Stanford HAI.