For many of us, it might feel like the internet is entering a trust recession.
Just last month, the Princess of Wales's long-awaited social media reappearance was marred by revelations of digital alterations to the image, fuelling already-frenzied internet speculation around the royal figure.
The saga highlighted for many the increasingly difficult task of spotting media that has been digitally altered — and in some cases, fabricated altogether using artificial intelligence (AI) generative technology.
So how can audiences learn to navigate an internet landscape increasingly populated by AI-generated images?
Experts say while some tips can help, a bigger shift in the way we view material online is inevitable, due to rapid advancements in AI.
How do you identify an AI-generated face online?
We headed into Melbourne's CBD and asked people if they could tell us which of the following two faces were generated.
Only one person correctly guessed the face on the right.
There are few subtle clues to give it away — one participant pointed to asymmetrical earrings as a giveaway.
Digital media lecturer Brendan Murphy from Central Queensland University says despite recent improvements in the technology, a trained eye can usually detect visual cues that give away AI-generated faces.
"There will still usually be details that don't quite line up," Mr Murphy says.
"Even if you look at things like teeth you might find that they're slightly asymmetrical — one too many, one too few.
"[AI] doesn't understand what a face is, it doesn't understand what a face does."
In 2019, ABC identified the mistakes frequently made by AI algorithms, as visual cues to help "spot" AI:
- Water splotches: Shiny, blurry blobs appearing across the image
- Background problems: Backgrounds appearing chaotic, corners of rooms not lining up, not making sense contextually
- Eyeglasses: Often asymmetric, crooked or jagged
- Other asymmetries: Look out for different earrings in the left and right ear, and different forms of collar or fabric on the left and right side
- Hair: Disconnected from head
- Teeth: Often odd-numbered or asymmetrical
However, in the time since then, the margin of error in these AI-algorithms has slimmed, making it near-impossible to internet users to still rely on visual imperfections alone to detect AI-generated images.
Researchers at The Australian National University recently found audiences relied on more intuitive cues to distinguish AI from real faces.
But they found this still led many of them astray when asked to distinguish between photographs of real human faces and images of faces generated using AI.
The study found more often than not, people incorrectly identified AI-generated images as real people.
They were also more likely to incorrectly identify actual humans as AI-generated.
See how you go with these six images, some of which ANU sourced from Flickr-Faces-HQ Dataset for its study:
AI or Real?
This is a real person.
90 per cent of participants in the study incorrectly identified this image as generated by AI.
AI or Real?
This is AI.
92 per cent of participants in the study incorrectly identified this image as a real person.
AI or Real?
This is AI.
93 per cent of participants incorrectly identified this image as a real person.
AI or Real?
This is a real person.
86 per cent of participants in the study incorrectly identified this image as generated by AI.
AI or Real?
This is a real person.
84 per cent of participants in the study incorrectly identified this image as generated by AI.
AI or Real?
This is AI.
90 per cent of participants incorrectly identified this image as a real person.
Ben Steward, a PhD candidate and co-author of the ANU research said participants were often using "attractiveness" as a distinguishing factor in their decision-making.
"We found that AI faces were rated as more attractive, and that these faces were also more likely to be correctly identified as AI," he says.
"This suggests that people are using cues present in the faces, such as attractiveness, to help distinguish between AI and human faces, even if they do not know it."
The study proposes that the program's output of faces that are more attractive is a result of the algorithm producing faces which have more "averaged" features.
"We know that faces with averaged features are considered more attractive, but also more symmetrical, and less memorable – two more features which we found could be used to distinguish between AI and human faces," Mr Steward says.
Will detecting AI-generated faces be impossible in the future?
While cues like proportionality, familiarity and memorability were identified as traits associated with human faces by participants in the study, they applied those qualities in the "wrong direction", the researchers found.
Basically, participants over-identified these traits in AI faces more often than human faces.
Similarly, human faces were found to be more "memorable" — a trait which participants associated with AI faces.
The ANU researchers described this phenomenon as "AI hyperrealism".
"AI hyperrealism is the phenomenon where people think faces generated by AI algorithms are real people more often than they think images of actual humans are real," Mr Steward says.
Mr Murphy says a future in which hyperrealism is the industry standard for AI-production is just around the corner.
"I don't think it's far in the future that there won't be any ways for a human to tell by eye," he says.
"As one thing gets ticked off, the systems move on to the next thing to perfect.
"It's a really interesting time for considering what a photo is, especially as things circulate on social media."
What does this mean for the future of the internet?
Earlier this year, Facebook parent company Meta launched an investigation into their platforms after pornographic deepfake images of Taylor Swift flooded the internet.
In February, CNN reported a finance clerk in Hong Kong fell victim to an elaborate scam, paying $39 million (AUD) to fraudsters who used generative deepfake technology to pose as higher-ups and colleagues of the worker over a Zoom call.
The string of high-profile incidents has prompted more and more questions about how to account for this technology once it becomes weaponised.
Carl Bergstrom is one of the University of Washington researchers behind Which Face Is Real?, an online quiz designed to demonstrate the "ease with which digital identities can be faked".
For Professor Bergstrom, mitigating these risks comes down to awareness.
"The real danger is having technologies that exist that you aren't aware exist," Professor Bergstrom says.
"We've always assumed that if you see a picture of a person's face that it has to be a real person, and it was clear to me that we were very near a time and place where this was no longer true."
Jean Burgess, a professor of digital media at the Queensland University of Technology, says in the midst of an "arms race between good and bad actors", media literacy has become a necessity.
"It's coming at a time where there's already a pretty big crisis in trust in the media, and a big crisis in media literacy," she says.
Professor Burgess says navigating and questioning representations of truth in the media requires an "active participation" in the technology.
"The more we use these tools, the better we will be at spotting the visual cues," she says.
How will 'AI literacy' protect us?
For Professor Bergstrom, however, the key going forward will not be "learning the subtleties of AI", but rather, knowing what can be faked and finding ways to "triangulate what is true".
"AI literacy has become something different than AI detection," he says.
"It is knowing that just because you see a face, it doesn't mean there's a person there.
"Even [current AI models] are good enough to fool people, so as it gets better and better it sort of doesn't matter if a person can identify it."
In March, tech giant OpenAI debuted their new voice-cloning program Voice Engine, but delayed a public rollout in order to "make a more informed decision about whether and how to deploy this technology".
The delay has called into question the increasingly accessible technology used to generate deepfakes and other forms of AI-synthesised text, audio and video, in an otherwise with many countries struggling to keep up with the pace of the potential cybersecurity threats posed by these tools.
Last month, the European Union took lead, passing down a landmark AI regulation rules — making it the first of its kind globally.
The laws seek to restrict the use of "high risk" AI products, forcing companies to label fake images.
Professor Bergstrom says regulation of this sort is not the first time internet audiences have had to contend with a new frontier.
"In the past we've been quite good at triangulating around these sorts of things — people were quite worried about photoshop when it came about," he says.
"But just as we did with photoshop, rather than being concerned and panicked about [AI], which I don't think is necessary, I think we just need to be prepared for it."
