Study finds generative AI tools produce racist and sexist stereotypes of Australians
Researchers tested five popular image-generating AI tools with 55 prompts about Australia and Australians in May 2024 and found roughly 700 images riddled with racist and sexist stereotypes. The outputs depicted idealized Australian families as overwhelmingly white, suburban and heteronormative, while reproducing caricatures anchored in the country's settler-colonial past.
What happened
In May 2024, researchers entered 55 text prompts into five popular image-generating AI tools: Adobe Firefly, Dream Studio, DALL-E 3, Meta AI and Midjourney, collecting approximately 700 images to examine how generative AI depicts Australia and Australians.
The images produced were riddled with racist and sexist caricatures anchored in the country's settler-colonial past, relying on tired tropes such as red dirt, Uluru, the outback, untamed wildlife and bronzed Australians on beaches.
Images of Australian families were overwhelmingly white by default, suburban and heteronormative, while prompts about 'An Australian mother' typically resulted in white, blonde women in domestic settings holding babies, with one tool producing only images of exclusively Asian women.
Reported impact
The outputs reproduce sexist and racist caricatures that represent an imagined monocultural past rather than contemporary Australian diversity.
The findings challenge the marketing narrative that generative AI is intelligent and creative, demonstrating instead how these tools embed existing biases in their training data.
Index analysis
The research identified that generative AI tools produce images representing Australians through a narrow, historically biased lens that fails to capture the country's demographic and cultural diversity.
Different tools showed varying biases: most produced white-dominant imagery for family prompts, while one tool (Adobe Firefly) produced exclusively Asian women for 'Australian mother' prompts, indicating inconsistent and stereotype-driven outputs across platforms.
Governance controls
Some AI tools refused certain requests, particularly those containing the words 'child' or 'children', indicating that some providers mark children as a risk category, though this represents only partial risk management of bias.
The study revealed that researchers used default settings on all tools without modifications, meaning the biased outputs reflect the systems as publicly deployed without customization.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.