LATENT LIBRARY ETHICS & GOVERNANCE Responsible AI Fairness, Privacy, Safety and Trustin Practice LATENT LIBRARY EDITORIAL

Responsible AI

Fairness, Privacy, Safety and Trust in Practice

Latent Library Editorial

IntroductionIntroduction

Almost every organisation that uses AI has, somewhere, a statement of principles: AI should be fair, transparent, private, safe and accountable. Few people disagree with these words. The difficulty is that principles do not tell you what to do on a Tuesday afternoon when a team wants to launch a feature next week, the data is imperfect, and nobody is sure whether the model treats some customers worse than others.

This book is about that Tuesday afternoon. It translates principles into practical methods: questions to ask, measurements to take, design choices to make and processes to set up, in proportion to the risk of each use.

Discussion exercises for team workshops, additional checklists and a glossary close the book.

No technical background is required. Where a method has a technical side, the book explains what it does and what to ask of the people implementing it. This book is not legal advice: laws differ by country and change over time, so involve qualified advisers for specific decisions.

Chapter 1Why Responsibility Is Practical

AI fails differently

Traditional software usually fails visibly: a button does not work, a page crashes. AI systems more often fail quietly. They produce plausible outputs that are wrong for certain people or situations, and nobody notices until harm has accumulated. Three properties make this especially important:

  1. Scale. One flawed model can make millions of decisions before anyone reviews them.
  2. Opacity. It is often hard to see why a model produced a particular output.
  3. Unevenness. Average accuracy can look excellent while the system performs poorly for specific groups.

The business case

Responsibility is also self-interest:

Types of harm

A simple map helps structure thinking:

HarmExample
AllocationA loan, job, benefit or opportunity unfairly withheld
Quality of serviceA speech system that understands some accents much worse
RepresentationOutputs that stereotype or demean groups
PrivacyPersonal information exposed, inferred or misused
SafetyPhysical, psychological or financial harm from errors or misuse
AutonomyPeople manipulated or unable to contest decisions about them

Proportionality

Not every use deserves the same scrutiny. Suggesting a subject line for a newsletter is low risk; screening job applicants, setting insurance prices or triaging patients is high risk. Effort should scale with potential harm: how severe, how likely, how many people, how reversible, and how vulnerable those affected are.

You have reached the end of the free sample.
The full book continues with 15 more chapters, plus everything else in its 80 pages.

Get the full book