Understanding and addressing AI harms

Summary

An evolving framework addresses AI harms beyond catastrophic risks, covering concerns such as biological threats, child safety, disinformation, and fraud. It complements the Responsible Scaling Policy by examining physical, psychological, economic, societal, and individual-autonomy impacts, then assessing likelihood, scale, affected populations, duration, causality, technology contribution, and mitigation feasibility. Policies and practices include a Usage Policy, pre- and post-launch evaluations, red teaming, adversarial testing, misuse detection, and enforcement ranging from prompt modifications to account blocking. Examples include computer-use safeguards for banking platforms and communications tools, plus hierarchical summarization intended to detect harms while maintaining privacy standards, alongside model-response evaluations balancing helpfulness with harmlessness. For Claude 3.7 Sonnet, encouraging safe, helpful responses to ambiguous prompts produced a reported 45% reduction in unnecessary refusals while maintaining safeguards against truly harmful content; the broader approach remains under development.

Takeaways

1. 1

   The framework assesses harms across five dimensions: physical, psychological, economic, societal, and individual autonomy.

2. 2

   Computer-use analysis considers risks involving financial software, banking platforms, communication tools, fraud, manipulation, influence operations, and phishing; hierarchical summarization supports detection while maintaining privacy standards.

3. 3

   For Claude 3.7 Sonnet, encouraging safe and helpful responses to ambiguous prompts reduced unnecessary refusals by 45% while maintaining safeguards against truly harmful content.