It was an image broadcast for just a few moments on a Melbourne TV news bulletin, but it’s since attracted international attention.
The digitally altered image of Victorian Animal Justice Party MP Georgie Purcell used to introduce a story on Victorian duck hunting saw the white dress she was wearing in the original photo swapped for a top exposing her midriff.
“Note the enlarged boobs and outfit to be made more revealing. Can’t imagine this happening to a male MP,” she tweeted.
After Ms Purcell called it out, media outlets including CNN and the BBC picked up the story, fuelling debate on the reach of generative artificial intelligence in our lives.
In its swift apology to Ms Purcell, broadcaster Nine News said the alteration had occurred due to an “automation by Photoshop” while resizing the original photo.
Ms Purcell says she’s not sure she buys Nine’s explanation but is happy to move on, provided a lesson is learnt by everyone to ensure it never happens again.
So how might the image have been made and what lessons should we take from it?
Expert says Nine’s explanation is plausible
Nine News has told the ABC the alteration to Ms Purcell’s midriff occurred when using Adobe Photoshop’s “generative expand” tool.
The tool allows users to make an image bigger — the program uses AI to make assumptions or guesses about how that image might be best filled out with new material.
Nine said the image it used to produce the graphic was a more tightly cropped version of the Bendigo Advertiser’s photo of Ms Purcell.
This image appears in some online image searches, and is cropped above Ms Purcell’s waist.
When that image was dragged downwards using the expand tool, the generated image exposing her midriff was created, according to Nine.
This was the image later incorporated into a graphic used in the network’s coverage of debate over the future of duck hunting in Victoria.
Adobe released a statement saying edits to the image in question “would have required human intervention and approval”.
TJ Thomson, a senior lecturer in digital media at RMIT, said it was plausible that a cropped image could produce a range of different torsos when expanded using AI.
“If you are giving Photoshop less data about the body … it has to imagine a lot more about what’s below the torso,” he said.
“So it can have a lot more creative input into what is below the torso.”
To demonstrate the tool in question, we uploaded a photo of a consenting ABC employee to Photoshop, cut off just above the chin.
When the program was presented with the image and asked to perform the “generative expand” action, it produced several variants of the torso.
The clothing in each torso differed, including a buttoned-up shirt and one with several of the top buttons undone.
In one version, the program appeared to struggle to generate the hands, presenting a jumble of fingers at the end of the man’s arms.
How a ‘vast library’ of images informs generative AI
Technology researcher Shaanan Cohney, who lectures at the University of Melbourne’s Centre for AI and Digital Ethics, said Photoshop’s generative fill tool was able to create new material for an image by drawing on a “vast library of stock images”.
“Now, Adobe claims to have done this in a way that reduces the bias in the images out there,” Dr Cohney said.
“But without more information, it’s very hard for us to know what kinds of stereotypes might be in that information.”
Dr Thomson said we could infer that many generative AI tools trawled the web to draw on images from films and pop culture, creating results that reinforced the biases present in society.
“If we look at our own landscape in terms of cinema or stock photography, often we see very biased and under-representative images and movies,” he said.
“Quite white-dominated, quite middle-aged or younger, not a lot of old folks, not a lot of people with disabilities.
“And so all those biases then are reproduced or enhanced sometimes with AI, if you’re not particularly prompting to try to get the more diverse representation.”
In a statement, an Adobe spokesperson said the company was committed to principles of accountability, responsibility and transparency when developing its AI products.
“Generative Fill and Generative Expand in Photoshop is powered by Adobe Firefly, Adobe’s family of creative generative AI models, which is trained on the highest quality of assets including Adobe Stock images, openly license content and public domain content where copyright has expired,” they said.
‘Sexualised’ images of women are training AI, researchers warn
When Ms Purcell saw her photoshopped body displayed on the TV, she said it was deeply upsetting.
“I don’t think there’s a single young person that hasn’t struggled with their body image, and seeing your own body altered on TV is very confronting,” she said.
“This has affected me in some way, and it could affect other women even more. It should never happen again.”
Walkley-award winning journalist Tracey Spicer said when she heard about Ms Purcell’s experience she was “horrified, but not surprised”.
Spicer, who recently authored a book examining the rise of artificial intelligence, said generative AI “routinely sexualises images of women and girls”.
“In fact, while designing the cover for my book, we put in prompts to design an image of a strong robot woman looking to the future with hope but concern,” Spicer said.
“[The AI generator] instead created an image of a sexy gold robot with huge breasts and a tiny waist.”
Spicer also backed a warning from the Victorian Women’s Trust that the community was on “the verge of a tsunami” when it came to the weaponisation of AI against women.
Singer and songwriter Taylor Swift was the latest high-profile woman to be attacked using AI, after pornographic deepfake images were circulated online.
Spicer said as technology became freer and easier to use, its potential to be abused was growing.
“Frankly, it’s terrifying,” she said.
“Many victim-survivors of this kind of abuse say it’s like being physically, sexually assaulted.”
New York University data journalism researcher Meredith Broussard said the doctored image of Ms Purcell was an important reminder that AI “keeps failing and screwing up” in a range of ways and human intervention was critical.
“It’s a lesson in why we shouldn’t necessarily trust the defaults in artificial intelligence,” she said.
“People think that AI is going to be somehow superior to humans and that’s simply not the case.”
In Ms Purcell’s case, Dr Broussard said the person operating the program that generated the altered image should have spotted the changes and opted not to use that image.
‘Major harms’ occurring due to racial biases
In the United States, Dr Broussard said there were several examples of “major harms” being carried out against people as a result.
“Facial recognition systems tend to be better at recognising light skin than dark skin, they tend to be better at recognising men than recognising women, they tend to not take trans and non-binary folks into account at all,” she said.
“And so when facial recognition is used in policing, for example, it is disproportionately weaponised against communities of colour.”
Dr Thomson said it was also clear that AI was being used as a tool to spread misinformation.
“People are using these tools to make photorealistic generations that can deceive people or mislead people,” he said.
“We’re seeing that happening in the Hamas, Israel conflict and multiple other contexts,” he said.
In Australia, retail giants Bunnings and Kmart are being investigated over their use of facial recognition technology in stores, amid privacy concerns.
Media companies urged to offer AI training
So where does this leave the media with its use of AI?
Dr Thomson said there were a range of different approaches being taken by outlets around the world.
“The majority of folks I’ve talked to are most comfortable using these tools for things like illustrations, things that are definitely not photorealistic, where you don’t have the potential to mislead or deceive,” he said.
“If you’re doing a story on mental health or COVID, or that kind of thing, and you have some clearly illustrated image, that’s AI-generated, that’s the kind of test case or use case that photo editors feel most comfortable using.
“Some outlets are a bit more conservative, saying we’re only going to use AI-generated images when we’re reporting on AI-generated images that have gone viral, for example.”
At the ABC, there is a policy not to use generative AI images in place of real images in news reporting.
If one is used, it must be clearly labelled — as we’ve done in this story.
Spicer said media companies had an obligation to better train staff.
“Issues like this will only deepen the public’s scepticism about journalism. That’s the last thing we need, at a time when democracy is under threat,” she said.
“There are things that journalists and media organisations can be doing right now. Staff should be trained in how to use AI safely and ethically.”
Calls for greater regulation in Australia
Several experts agree greater regulation of AI is needed — particularly in Australia.
In January, the Australian government unveiled plans to target some high-risk AI technologies with audits and safeguards — but Dr Thomson said more was needed to “catch up” with the rest of the world.
“Australia has been called kind of the back of the pack in terms of AI regulation,” Dr Thomson said.
In December, the European Union passed the world’s first AI law, which paved the way for legal oversight of technology used in popular generative AI services.
In the United States, President Joe Biden made a major effort to regulate artificial intelligence under a new executive order where AI developers would need to risk-test their tools and share the results with the US government.
On Wednesday, Australian Securities and Investments Commission chair Joe Longo warned existing laws did not adequately prevent “AI-facilitated harms” before they occurred and further “transparency and oversight” may be needed.
Another idea flagged by the federal government is tagging AI-generated images with a watermark or in another way that ensures people seeing them are aware of how they’ve been made.
For its part, Adobe said that was an idea the company supported, likening it to health advice included on food packaging.
“We are working with customers and partners to enable content credentials — which are “nutrition labels” for digital content, that remain associated with content wherever it is used, published or stored,” a spokesperson said.
Dr Cohney said even with legal obligations, “ethical norms” and “AI literacy” in the community would be critical as the technology’s use expanded.
Even if we have the perfect set of laws, they still won’t catch every example of people not using them appropriately,” he said.
“So we need to think about the way we train people, whether that be in the media, or whether they be working in another industry, to use AI in an ethical fashion.”
Posted , updated
