liveBook · Manning

Harm

As explained in How Large Language Models Work

Harm is physical, psychological, social, legal, or economic damage caused to people or other affected entities.

Harm is damage caused to people or other affected entities, including physical, psychological, social, legal, or economic damage. For LLMs, fluent and broad capability does not make every use safe. section 9.1.1

Potential harm appears when systems are used in unexpected ways, especially when narrow tests cannot anticipate real-world deployment. Bias in training data and generated outputs can create unintended consequences. section 9.1.1

A biased business-meeting image

An image-generation model prompted to create a business meeting may produce an image in which everyone is male and white. Fixing that specific stereotype does not establish that other prompts or deployments will be safe. section 9.1.1

How How Large Language Models Work covers it

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.
How it connects
Connects harm with bias, training data, deployment, safety testing, and image generation.
Cover of How Large Language Models Work
liveBookmanning.com
Read How Large Language Models Work on liveBook Get the book at manning.com

Also explained in

Compare how 2 books explain harm →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of How Large Language Models Work. The section links point to the parts of the book it draws on.