Responsible AI is not keeping pace with AI capability. Documented incidents rose 55% in a single year.
That is a headline finding of Stanford HAI's AI Index Report 2026 (third revision, dated June 29, 2026), which measures the state of AI across nine chapters. The Responsible AI chapter carries the numbers that belong in front of executives.
What it says
The AI Incident Database, which tracks documented cases where AI systems caused or nearly caused harm, recorded 362 incidents in 2025, up from 233 the year before. The pattern repeats inside organizations: in the report's survey of business leaders, half of those reporting incidents experienced three to five of them, up from 30% a year earlier.
Confidence is moving the other way. In the same survey, organizations rating their incident response as excellent fell from 28% to 18%. Experience is supposed to build confidence. Here it is doing the opposite.
The context makes it heavier. 88% of organizations in the McKinsey survey cited by the report use AI in at least one business function. Responsible AI maturity averages 2.3 on a four-point scale. And safety that holds under normal use degrades under attack: in jailbreak testing, nearly every model's safety score dropped, some by a full tier.
Why it matters
The report describes organizations deploying faster than they can operate. That gap does not stay internal. It surfaces in customer security reviews, vendor risk assessments, and diligence processes, where incident history and response capability are now standard questions.
What shapes governance practice is also shifting. ISO/IEC 42001 and the NIST AI RMF entered the most-cited regulatory influences this year. Regulators and customers are converging on management systems and controls, not policy documents.
One more signal: model providers grew less transparent in 2025, with average transparency scores dropping from 58 to 40. What vendors will not disclose becomes risk the deploying organization absorbs.
Key takeaways
- Incident response capability is now a differentiator. Repeat incidents are becoming the norm among affected organizations, and most rate their own response as mediocre.
- Safety testing under normal conditions is not evidence of production readiness. Adversarial testing is where the real posture shows.
- Vendor transparency is declining, so contractual and architectural controls have to compensate for what disclosure no longer covers.
The solution is operational: controls enforced where the AI runs, adversarial testing as a deployment gate, incident response built for AI failure modes, and architecture that covers what vendors no longer disclose.
Source:
· Stanford HAI — Artificial Intelligence Index Report June 2026 — hai.stanford.edu, June 2026
Most organizations can show an AI policy. Could yours show which control failed during its last AI incident?