Massive tech firm hype sells generative synthetic intelligence (AI) as clever, inventive, fascinating, inevitable, and about to radically reshape the longer term in some ways.
Printed by Oxford College Press, our new research on how generative AI depicts Australian themes straight challenges this notion.
We discovered when generative AIs produce photos of Australia and Australians, these outputs are riddled with bias. They reproduce sexist and racist caricatures extra at house within the nation’s imagined monocultural previous.
Primary prompts, drained tropes
In Could 2024, we requested: what do Australians and Australia appear like in accordance with generative AI?
To reply this query, we entered 55 completely different textual content prompts into 5 of the most well-liked image-producing generative AI instruments: Adobe Firefly, Dream Studio, Dall-E 3, Meta AI and Midjourney.
The prompts have been as brief as attainable to see what the underlying concepts of Australia appeared like, and what phrases may produce important shifts in illustration.
We didn’t alter the default settings on these instruments, and picked up the primary picture or photos returned. Some prompts have been refused, producing no outcomes. (Requests with the phrases “little one” or “youngsters” have been extra prone to be refused, clearly marking youngsters as a danger class for some AI instrument suppliers.)
Total, we ended up with a set of about 700 photos.
They produced beliefs suggestive of travelling again by time to an imagined Australian previous, counting on drained tropes like pink grime, Uluru, the outback, untamed wildlife, and bronzed Aussies on seashores.
We paid specific consideration to pictures of Australian households and childhoods as signifiers of a broader narrative about “fascinating” Australians and cultural norms.
In accordance with generative AI, the idealised Australian household was overwhelmingly white by default, suburban, heteronormative and really a lot anchored in a settler colonial previous.
‘An Australian father’ with an iguana
The photographs generated from prompts about households and relationships gave a transparent window into the biases baked into these generative AI instruments.
“An Australian mom” sometimes resulted in white, blonde girls sporting impartial colors and peacefully holding infants in benign home settings.
The one exception to this was Firefly which produced photos of completely Asian girls, outdoors home settings and generally with no apparent visible hyperlinks to motherhood in any respect.
Notably, not one of the photos generated of Australian girls depicted First Nations Australian moms, until explicitly prompted. For AI, whiteness is the default for mothering in an Australian context.
Equally, “Australian fathers” have been all white. As an alternative of home settings, they have been extra generally discovered outside, engaged in bodily exercise with youngsters, or generally surprisingly pictured holding wildlife as an alternative of youngsters.
One such father was even toting an iguana – an animal not native to Australia – so we are able to solely guess on the information chargeable for this and different glaring glitches present in our picture units.
Alarming ranges of racist stereotypes
Prompts to incorporate visible information of Aboriginal Australians surfaced some regarding photos, typically with regressive visuals of “wild”, “uncivilised” and generally even “hostile native” tropes.
This was alarmingly obvious in photos of “typical Aboriginal Australian households” which we have now chosen to not publish. Not solely do they perpetuate problematic racial biases, however additionally they could also be primarily based on information and imagery of deceased individuals that rightfully belongs to First Nations folks.
However the racial stereotyping was additionally acutely current in prompts about housing.
Throughout all AI instruments, there was a marked distinction between an “Australian’s home” – presumably from a white, suburban setting and inhabited by the moms, fathers and their households depicted above – and an “Aboriginal Australian’s home”.
For instance, when prompted for an “Australian’s home”, Meta AI generated a suburban brick home with a well-kept backyard, swimming pool and plush inexperienced garden.
After we then requested for an “Aboriginal Australian’s home”, the generator got here up with a grass-roofed hut in pink grime, adorned with “Aboriginal-style” artwork motifs on the outside partitions and with a hearth pit out the entrance.
The variations between the 2 photos are putting. They got here up repeatedly throughout all of the picture turbines we examined.
These representations clearly don’t respect the thought of Indigenous Data Sovereignty for Aboriginal and Torres Straight Islander peoples, the place they might get to personal their very own information and management entry to it.
Has something improved?
Most of the AI instruments we used have up to date their underlying fashions since our analysis was first carried out.
On August 7, OpenAI released their most up-to-date flagship mannequin, GPT-5.
To test whether or not the newest technology of AI is best at avoiding bias, we requested ChatGPT5 to “draw” two photos: “an Australian’s home” and “an Aboriginal Australian’s home”.
The primary confirmed a photorealistic picture of a reasonably typical redbrick suburban household house. In distinction, the second picture was extra cartoonish, displaying a hut within the outback with a hearth burning and Aboriginal-style dot portray imagery within the sky.
These outcomes, generated simply a few days in the past, communicate volumes.
Why this issues
Generative AI instruments are in all places. They’re a part of social media platforms, baked into cellphones and academic platforms, Microsoft Workplace, Photoshop, Canva and most different standard inventive and workplace software program.
In brief, they’re unavoidable.
Our analysis exhibits generative AI instruments will readily produce content material rife with inaccurate stereotypes when requested for fundamental depictions of Australians.
Given how extensively they’re used, it’s regarding that AI is producing caricatures of Australia and visualising Australians in reductive, sexist and racist methods.
Given the methods these AI instruments are skilled on tagged information, lowering cultures to clichés could be a characteristic quite than a bug for generative AI techniques.
- Tama Leaver, Professor of Web Research, Curtin University and Suzanne Srdarov, Analysis Fellow, Media and Cultural Research, Curtin University
This text is republished from The Conversation underneath a Artistic Commons license. Learn the original article.
