Our approach to understanding and addressing AI harms

As AI capabilities rapidly advance, understanding and addressing the full spectrum of potential impacts becomes increasingly important. Today, we're sharing insights into our evolving approach to assessing and mitigating various harms that could result from our systems, ranging from catastrophic scenarios like biological threats to critical concerns like child safety, disinformation and fraud.

Why is this approach important? As models continue to evolve, we need more comprehensive ways to think about and manage their potential impacts. We believe that considering different types of harms in a structured way helps us better understand the challenges ahead and informs our thinking about responsible AI development.

Our approach complements our Responsible Scaling Policy (RSP) , which focuses specifically on catastrophic risks. Identifying and addressing the full range of potential impacts requires a broader perspective. That's why we've built out a more comprehensive framework to assess harm that we can then proportionately manage and mitigate.

For each dimension, we consider factors like likelihood, scale, affected populations, duration, causality, technology contribution, and mitigation feasibility. This helps us understand the real-world significance of different potential impacts.

Depending on harm type and severity, we address and manage risks through a variety of policies and practices including developing and maintaining a comprehensive Usage Policy , conducting evaluations (including red teaming and adversarial testing) before and after launch, sophisticated detection techniques to spot misuse and abuse, and robust enforcement ranging from prompt modifications to account blocking. This perspective helps us balance multiple considerations: addressing harms with proportionate safeguards while maintaining the helpfulness and functionality of our systems in everyday use cases. We’re excited to share more about this work in the near future.

*Important note : This approach is still evolving. We're sharing our current thinking while acknowledging it will continue to develop as we learn more. We welcome collaboration from across the AI ecosystem as we work to make these systems benefit humanity.

We've developed an approach that helps our teams communicate clearly, make well-reasoned decisions, and develop targeted solutions for both known and emergent harms. This approach is designed to be both principled and adaptable to keep up with the evolving AI landscape. We examine potential AI impacts across multiple baseline dimensions, with room to grow and expand over time:

When exploring new capabilities or features, we examine how they might introduce additional considerations across different harm dimensions. For example: