Typology of Risks of Generative Text-to-Image Models

The paper identifies 22 distinct types of risks associated with text-to-image models like DALL-E and Midjourney, organized across six stakeholder groups, indicating that the risks from these systems are more diverse and complex than commonly recognized.

Significant knowledge gaps exist in how these risks are currently understood and addressed, suggesting that despite some risks receiving attention, many remain inadequately studied or treated by developers and policymakers.

The taxonomy spans risks ranging from data bias to malicious use, meaning responsible AI governance must address both systemic issues (like biased training data) and intentional harms (like creating deceptive or harmful content).

The paper emphasizes that different stakeholder groups face different risks, implying that a one-size-fits-all approach to responsible AI development is insufficient—protections must be tailored to address specific vulnerabilities of creators, users, subjects depicted in images, and society broadly.

By mapping previously overlooked risks, the paper aims to guide future research and governance initiatives, suggesting that the field currently lacks comprehensive frameworks for responsible development and that policymakers need better evidence to create effective safeguards.