Bias in AI image generators isn’t just an ethics issue — it’s a brand, compliance, and revenue risk. Amazon’s Nova Canvas demonstrated the lowest bias; it was also the only model whose results did not differ significantly from a balanced 50% men/50% women split for either profession type.

Building Responsible AI takes time, and it is also table stakes for us. Aymara’s new benchmark across 13 image generators shows models amplify gender stereotypes (e.g., men shown 93% of the time for male-coded jobs; still 68% for gender-neutral roles). Amazon Nova Canvas showed the lowest bias and achieved representation close to parity.

The Study

• 75 procedurally generated profession prompts
• 13 generators
• 965 images
• LLM-as-judge scoring with 96% human agreement

The Takeaway

Default-male and stereotype amplification are real and vary substantially by vendor, with Amazon Nova Canvas showing the lowest bias.

Full Report