An Approach to Understanding and Addressing AI Harms

As AI capabilities rapidly advance, Anthropic has shared an evolved approach to assessing and mitigating a broad range of AI Harms, spanning from catastrophic scenarios like biological threats to child safety, misinformation, and fraud.

This approach complements the Responsible Scaling Policy (RSP), which focuses on catastrophic risks, and is designed to manage harm from a more comprehensive perspective. The criteria for assessing harm are categorized into the following 5 dimensions.

• Physical Impact: Effects on physical health and well-being
• Psychological Impact: Effects on mental health and cognitive function
• Economic Impact: Financial consequences and property considerations
• Societal Impact: Effects on communities, institutions, and shared systems
• Individual Autonomy Impact: Effects on individual decision-making and freedom

For each dimension, factors such as likelihood of occurrence, scale, scope of impact, persistence, causality, technical contribution, and mitigability are comprehensively reviewed. Depending on the severity of the risk, this involves establishing a Usage Policy, conducting Evaluations including Red Teaming and adversarial testing, employing misuse detection technology, and applying strong Enforcement measures such as account suspension in parallel.

For example, in the case of the Computer Use feature, where models interact with computer interfaces, risks such as fraud through financial software or phishing and influence campaigns via communication tools are proactively identified, and monitoring and enforcement standards tailored to these risks are designed and applied.