Concept guide

Responsible AI

AI now helps review loan applications, screen job applicants and support doctors. Is it enough for it to be accurate? You also have to ask whether it treats some groups unfairly, whether anyone can explain why it decided what it did, and who is responsible when it gets something wrong. This guide walks through the principles of responsible AI and the practical ways teams actually live up to them.

01

The principles of responsible AI

Governments, international bodies and AI companies have all published their own AI principles. The names and the number of principles vary a little, but what they say is nearly the same. In the scene below, pick a principle and see what goes wrong when it breaks.

Principle
Fairness
Don't treat people in similar situations differently
When it breaks
Applicants with the same experience get lower scores because of their gender

It takes all six principles together to hold up trustworthy AI

Fairness: Don't treat people in similar situations differently. When it breaks: Applicants with the same experience get lower scores because of their gender.

PrincipleWhat it means
FairnessDon't treat people in similar situations differently because of a group like gender, age or region
Reliability & safetyWork consistently as expected under different conditions (robustness), and cause no harm
Privacy & securityProtect data, and use it only within what people agreed to
InclusivenessAnyone can use it, whatever their disability, age, language or region
TransparencyTell people AI is being used and how it reaches decisions (includes explainability)
AccountabilitySomeone is responsible for the outcome, and there is a way to appeal and fix mistakes

Writing principles down doesn't make anyone follow them. The rest of this guide covers what you actually measure and build for each one.

If a question says… → pick this

  • “Treats people in similar situations differently by group”A fairness problem
  • “Works consistently under many different conditions”Reliability & safety (robustness)
  • “Usable by anyone, whatever their disability or language”Inclusiveness
  • “Tells users AI is involved and how it decides”Transparency
  • “A named owner and a process for the outcome”Accountability

The principle that causes trouble most often is fairness. So why would an AI treat one group worse than another?

02

Where bias comes from

AI doesn't set out to discriminate. Most bias comes from the data. A model learns from the examples it's shown, so if the data leans one way, the results lean the same way. The scene below shows a speech recognition model: how much training data it had for each accent, and how often it got each accent wrong.

Standard accent
900 hrs · 5% errors
Regional accent
100 hrs · 22% errors

Speech recognition model — training data and error rate by accent (illustrative)

Regional-accent audio is only 10% of the training data, so the model misunderstands those speakers about one time in four (22%). The model isn't discriminating on purpose. The data is lopsided.

Type of biasHow it happensExample
Sampling (representation) biasSome groups have too little dataToo few recordings of regional accents, so the model misunderstands them
Historical biasUnfair decisions from the past are baked into the dataA group that was rarely hired in the past gets low scores from the model too
Measurement / label biasThe way answers were labeled leans one wayOne group's behavior was flagged as "risky" more often
Proxy variablesOther features reveal the sensitive information you removedGender was dropped, but names and hobbies still give it away
Here's what trips people up: removing sensitive attributes like gender, age or region from the inputs doesn't make the bias go away. Other features can carry the same information. That's why you have to split the results by group and actually measure them.

If a question says… → pick this

  • “Performance is low because one group has too little training data”Sampling (representation) bias → collect more data
  • “The model copies past discriminatory decisions”Historical bias
  • “Removing sensitive attributes eliminates bias”Wrong — proxy variables remain
  • “Check the share of each group in the dataset”Bias detection (data analysis)

So how do you check, with actual numbers, whether a model is fair?

Pro only

That’s the free part of this guide

The other 5 chapters are open to Pro subscribers. Each one continues with 3D visuals and “if a question says…” tips, and the last chapter lets you check yourself with a short quiz.

  1. 03Measuring fairness
  2. 04Explainability
  3. 05Transparency and documentation
  4. 06Governance and regulation
  5. 07Reading situations as principles
See ProAlready on Pro? Log in

Advertisement

PasslyExam

You have read the concept. Now test it on exam questions

PasslyExam offers original practice questions, explanations and full mock exams built around each certification’s exam guide. The explanations cover why the right option is right and why the look-alike options are wrong. Sign up to practice for free every day.

Start for freeLog in

Exams that cover this concept