A Closer Look at the Existing Risks of Generative AI: Mapping the Who, What, and How of Real-World Incidents

Most harms come from how systems are used, not how they're built: The analysis of 499 real-world incidents found that the majority of reported problems stem from "use-related issues" rather than failures during design and development, suggesting that responsible deployment and user education are critical to preventing harm.

Harms extend far beyond the direct users: A key finding is that incidents "bring harm to parties beyond the end user(s) of the Generative AI system at fault," meaning that bystanders, communities, and society can suffer consequences even when the person using the AI system faces no direct negative impact.

Generative AI risks look different from traditional AI risks: The paper emphasizes that "the landscape of Generative AI harms is distinct from that of traditional AI," indicating that lessons learned from previous AI systems may not fully apply and new risk frameworks are needed specifically for generative models.

Non-technical solutions deserve more attention: The authors call for "prioritization of non-technical risk and harm mitigation strategies, including public disclosures and education," suggesting that policy, transparency, and user awareness may be as important as technical safeguards in reducing real-world harms.

Regulatory approaches need careful calibration: The paper recommends "careful regulatory stances," implying that overly rigid or poorly designed regulations could be counterproductive, and that policymakers should base rules on evidence from actual incidents rather than theoretical concerns alone.