Documented AI safety incidents rose sharply in 2026, according to the Responsible AI chapter of Stanford University's 2026 AI Index Report. The AI Incident Database recorded 362 incidents in 2025, up from 233 the year before, an increase of more than 55 percent. Monthly incident counts peaked at 435 in January 2026, and the six month moving average has settled at 326, suggesting the elevated pace has persisted rather than reflecting a single spike.

The report's authors note the incidents span a wide range of failures, including AI-generated romance scams built on deepfake video, hate speech produced after a chatbot's safety filters were loosened, and AI-assisted phishing campaigns targeting customers of bankrupt retailers. Businesses evaluating AI vendors for marketing, customer service or content generation face a widening gap between how AI systems are marketed and how their safety and accuracy are actually measured.

That gap shows up in benchmarking practices. Nearly all major AI developers publish results on capability benchmarks that measure how well a model performs a task, but reporting on responsible AI benchmarks covering safety, fairness and transparency remains far less common. A new accuracy benchmark cited in the report found hallucination rates across 26 leading models ranging from 22 percent to 94 percent, a spread researchers say makes it difficult for buyers to compare tools on reliability alone.

For companies weighing AI adoption in marketing and customer-facing functions, the findings point to a widening measurement gap between marketed capability and demonstrated reliability.

Source: Stanford HAI - https://hai.stanford.edu/ai-index/2026-ai-index-report/responsible-ai