liveBook · Manning

Distributional harm

As explained in Generative AI in Action

Distributional harm is harm whose effects are unequally distributed across a group or society rather than confined to one person.

Distributional harm emerges from aggregate model behavior that affects a group or society unevenly, rather than from one isolated output. It includes unequal treatment, resource allocation, opportunities, or performance across demographic groups. section 13.3.1 section 13.1

Evaluate potential distributional harm across:

  • The type of harm: representational, allocational, or capability-related. section 13.3.1
  • The affected demographic groups and whether performance or treatment differs among them. section 13.3.1
  • The application, textual, and social context, since the same error can carry different consequences across domains. section 13.3.1
  • Likelihood, severity, and frequency, then prioritize the most pressing risks for mitigation. section 13.3.1

Customer-service chatbot

A customer-service chatbot may provide outdated or incorrect information and treat users unequally by gender or race. The pattern across users constitutes a distributional concern, requiring demographic evaluation alongside privacy and accuracy review. section 13.3.1

How Generative AI in Action covers it

How deep it goes
Distinguishes harm types, contexts, recipients, and demographic groups, then prioritizes risks without detailing mitigation implementation.
Examples
Uses customer-service chatbot disparities and contrasts summarization risks in news and healthcare contexts.
What you can do after
Evaluate and prioritize group-level harms by likelihood, severity, frequency, context, and affected demographic.
How it connects
Connects distributional harm with representational harm, allocational harm, capability fairness, individual harm, privacy, and misinformation.
Cover of Generative AI in Action
liveBookmanning.com
Read Generative AI in Action on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Generative AI in Action. The section links point to the parts of the book it draws on.