AI Terminology Course
AI Terminology
/
Intermediate

Bias in AI

Definition

Systematic and unfair prejudice generated by an AI model, usually stemming from skewed, unrepresentative, or historically prejudiced training data.

Explain Like I'm New

If you train an AI by having it read the internet, the AI will learn and mathematically amplify all the racism, sexism, and toxic stereotypes that currently exist on the internet.

Real World Example

Amazon built an AI to screen resumes. Because the training data consisted of 10 years of past Amazon hires (who were predominantly male), the AI learned that 'Being Male' was a good feature. It began actively penalizing resumes that contained the word 'Women's' (like 'Women's Chess Club Captain'). Amazon had to scrap the project.

Common Use Cases

  • •Algorithmic fairness
  • •AI ethics

Interview Questions

basic

  • Is AI inherently unbiased because it is just a mathematical machine?

intermediate

  • What is 'Representation Bias' in computer vision datasets?

Flash Cards

Question

Inherently unbiased?

Click to reveal answer
Answer

NO. AI is a mirror. It mathematically reflects the biases present in the human-created data it was fed.

Question

Representation Bias?

Click to reveal answer
Answer

When a dataset heavily over-represents one group. If a facial recognition AI is trained on 1 million faces of white men, and 10 faces of minority women, it will have a massive failure rate when trying to identify minorities in the real world.