Stanford University’s 2026 AI Index shows a technology advancing faster than the systems used to compare its risks. The 423-page report covers models, research, investment and public opinion, but its most consequential finding may be the absence of consistent safety results across frontier systems.
Capability benchmarks are routinely published. Responsible-AI measures covering security, fairness, factuality and human agency are far less complete. Most entries in Stanford’s comparison table are empty, making meaningful external safety comparison almost impossible.
The gap is becoming more serious as documented incidents rise and the performance difference between leading countries narrows.
Safety reporting remains fragmented
Among the frontier models tracked, only Claude Opus 4.5 reports results on more than two responsible-AI benchmarks. Only GPT-5.2 reports StrongREJECT. This does not prove that laboratories perform no internal testing; Stanford notes that red-teaming and alignment work occurs, but rarely through a common public framework.
Without comparable results, buyers cannot easily determine whether one model handles a specific risk better than another. Developers can select the tests and presentation most favourable to their own systems, while safety improvements remain difficult to track over time.
Trade-offs complicate the task. Improving privacy may reduce fairness, or stronger safety constraints may affect accuracy. The field lacks an established method for deciding among those outcomes, and standardised historical data is limited in areas such as explainability.
Incident figures show why this matters. The AI Incident Database recorded 362 cases in 2025, up from 233 in 2024. A broader OECD monitor reached 435 incidents in January 2026, with a six-month moving average of 326.
Organisational readiness is moving backward
A Stanford and McKinsey survey found that organisations rating their incident response as excellent fell from 28% in 2024 to 18% in 2025. “Good” ratings declined from 39% to 24%, while the share experiencing three to five incidents rose from 30% to 50%.
These numbers suggest adoption is expanding faster than response capability. A company may have model evaluations, yet still lack ownership, evidence collection and recovery procedures when a system causes harm.
Common benchmarks cannot replace incident preparation, but they can help organisations select models and target controls. Providers should disclose results across a stable core set, explain test conditions and report limitations instead of presenting one safety score as universal assurance.
Customers should also demand operational evidence: logs, escalation paths, rollback capability and documented changes between model versions. Safety is not only a characteristic of the model; it emerges from the complete deployed system.
Competition and public trust raise the stakes
Stanford finds that the performance gap between leading US and Chinese models has effectively closed. The top position changed several times after early 2025, and Anthropic’s leading model was ahead by only 2.7% in March 2026. The US produced 50 top-tier models in 2025 against China’s 30, while China leads publication volume, citation share and patent grants.
Hardware concentration creates another vulnerability. The US hosts 5,427 data centres, but TSMC fabricates nearly every leading AI chip, primarily through one Taiwanese foundry despite new US production.
Public opinion is equally complex. Globally, 59% say AI’s benefits outweigh drawbacks, while 52% say AI products make them nervous. Experts are far more optimistic about employment: 73% expect a positive effect on work, compared with 23% of the public.
Trust in regulation varies sharply. Only 31% of US respondents trust their government to regulate AI responsibly, against a 54% global average. Across 25 countries, the EU earns more confidence than either the US or China.
The report’s message is not that AI progress should stop. It is that capability competition, deployment and incidents are accelerating without an equally visible safety language. Comparable benchmarks will not solve every trade-off, but their absence prevents informed scrutiny.
Frontier developers increasingly compete on small performance margins. They should be expected to compete just as openly on evidence that their systems can be deployed responsibly.
