How do I write prompts that avoid bias?
Use neutral and inclusive terms
Replace gendered terms like 'businessman' with 'businessperson' or 'professional'. Instead of 'fireman', use 'firefighter'.
Avoid defaulting to a single race, body type, or ability unless the context requires it. For example, 'a doctor' should not automatically be depicted as a white man.
When describing people, focus on relevant attributes (e.g., 'a scientist wearing a lab coat') rather than stereotypes.
- Use 'they' as a singular pronoun when gender is unknown
- Specify 'person with a disability' rather than 'disabled person' if preferred
- Avoid terms like 'exotic' or 'normal' that imply a default
- Include diverse names and cultural references intentionally
Be explicit about diversity when needed
If you want a diverse group, say so: 'a team of people of different ages, ethnicities, and abilities collaborating'.
For image prompts, you can specify 'diverse cast' or 'mixed-race group' to counteract default biases in training data.
For text prompts, ask the AI to consider multiple perspectives or to avoid stereotypes explicitly in your instruction.
- Add 'diverse' or 'inclusive' to group descriptions
- Request 'avoid stereotypes' in your prompt
- Specify skin tone, hair texture, or clothing only when relevant
- Use 'they' or 'their' for unknown individuals
Test and revise
After generating output, review it for unintended bias. If the AI defaults to a stereotype, revise your prompt to be more specific.
Ask yourself: would this prompt make assumptions about a person's gender, race, religion, or ability? If so, rephrase.
Remember that AI models reflect biases in their training data. Your prompt can help steer away from those defaults.
- Run the same prompt multiple times to see if outputs vary
- Use negative prompts to exclude stereotypical elements
- Check for loaded words like 'aggressive' or 'submissive'
- Consider cultural context if your prompt references specific groups
Common mistakes
- Assuming that avoiding bias means ignoring race or gender entirely—sometimes specifying diversity is necessary.
- Using 'he' as a default pronoun for unknown people, which reinforces gender bias.
- Believing that AI is inherently unbiased—models often amplify stereotypes from training data.
