liveBook · Manning

Bias

As explained in How Large Language Models Work

Bias is a systematic skew in information, representation, or model behavior toward particular perspectives, groups, or time periods.

Bias is a systematic skew in what information represents and how a model behaves. In language and image models, it can favor particular groups, perspectives, stereotypes, or time periods. Fluent output does not make that skew harmless or obvious. section 9.1.1 section 9.3.3

Bias can enter an LLM through several connected routes:

  • Training data can reproduce stereotypes or unequal treatment, including negative evaluations of race or gender. section 9.1.1
  • Public-domain collections can overrepresent older writing, carrying outdated cultural attitudes, inaccurate science, unfamiliar language, and historical racism or sexism into training data. section 9.3.3
  • A narrow test may expose one failure without predicting harm from different prompts, users, or deployment contexts. section 9.1.1
  • Avoiding one harmful category does not automatically make a model safe, because unanticipated uses can create different harms. section 9.1.1

A business-meeting image

Prompting an image model for a business meeting may produce an image in which everyone is male and white. The example reveals a contextual stereotype, but fixing that prompt alone does not establish safety across other prompts or real-world uses. section 9.1.1

How How Large Language Models Work covers it

How deep it goes
Explains examples, data sources, and testing limits, but does not develop a formal statistical account of bias.
Examples
Uses race and gender evaluations, African American Vernacular English, business-meeting images, and older public-domain texts.
What it assumes
Assumes familiarity with LLMs and training data while defining bias through concrete outputs, groups, and historical examples.
What you can do after
Evaluate model outputs and intended uses for stereotypes, historical skew, untested contexts, and possible downstream harm.
How it connects
Connects bias with training data, stereotypes, public-domain data, model evaluation, and deployment risk.
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 28 books explain bias →

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.