AI systems learn from data created by people, and that data contains human biases. As a result, AI outputs can be unfair or stereotyped.
How bias shows up
- Image generators defaulting to certain genders or ethnicities for jobs
- Language that assumes stereotypes
- Uneven performance across languages, accents or dialects
- Unfair outcomes in automated decisions like screening
Why it happens
Training data over-represents some groups and under-represents others, and reflects historical inequalities. Design choices and evaluation can also introduce bias.
What you can do
- Be specific in prompts about diversity when generating images of people.
- Review outputs for stereotypes before publishing.
- Don't use AI alone for decisions that affect people's opportunities, such as hiring or lending.
- Test outputs with diverse examples.
- Invite feedback from people with different perspectives.
For organisations
Audit AI systems used in decisions, document how they're used, keep humans accountable, and follow emerging regulations.