Harm
Harm is damage, injury, or another adverse effect affecting people, organizations, assets, or other entities.
Generative AI in Action
Harm is an adverse effect or injury caused to individuals, groups, society, or other affected parties.
- How it is used
- Frames generative AI governance around identifying, classifying, measuring, and mitigating adverse effects.
- How deep it goes
- Defines harm categories and supplies a procedural evaluation workflow, stopping short of detailed mathematical measurement methods.
- Examples
- Uses customer-service chatbots, diverse test inputs, documented outputs, rubrics, red-teams, stress tests, and evaluation pipelines.
- What you can do after
- Diagnose harm by prioritizing risks, testing targeted inputs, quantifying harmful outputs, and checking mitigation regressions.
How Large Language Models Work
Harm is physical, psychological, social, legal, or economic damage caused to people or other affected entities.
- How it is used
- Uses harm to assess whether broad LLM applications are safe enough for real-world deployment.
- How deep it goes
- Explains unintended harm through bias, testing gaps, deployment context, and the option not to use an inadequately mitigated system.
- Examples
- Uses biased race and gender evaluations, a stereotyped business-meeting image, and real-world LLM deployments as examples.
- What you can do after
- Evaluate whether an intended LLM use has foreseeable harms and whether available mitigations justify deployment.
Related concepts
- impactmeasured by · Impact is the effect or consequence of an event, action, or condition, often assessed by its magnitude or significance for people, systems, assets, or outcomes.
- distributional harmincludes · Distributional harm is harm whose effects are unequally distributed across a group or society rather than confined to one person.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.