Can you spot an AI deepfake? Researchers say they can train people to tell what is real and what is not

Can you tell the difference between an AI-generated face and that of a real person?

Major advances in artificial intelligence are powering hyper-realistic deepfake images, video and audio, making it challenging to distinguish what is authentic and what is artificial.

These false depictions, as we have seen at schools and in a first-of-its-kind case involving high-profile victims, present major hazards.

“Deepfake fraud is a real and growing problem,” UNSW Sydney IDeA Lab director Dr James Dunn told 7NEWS.com.au.

“AI-generated faces can make fake profiles, romance scams and misinformation campaigns appear more credible, and Interpol reports that synthetic images are already being used in social-engineering schemes worldwide.”

About one in four Australians say they have come across a deepfake scam in the past years, according to a study from the Commonwealth Bank.

Amid the ongoing battle against deepfake fraud, a team of researchers led by the Australian National University (ANU) Emotions and Faces Lab say they have successfully trained people to identify what is real and what is fake when it comes to a person’s visage.

“Training on visual artifacts, like looking for a sixth finger or odd earrings, has had limited success, partly because the AI is getting too good, and fraudsters may avoid using pictures with obvious flaws anyway,” lead researcher Associate Professor Amy Dawel said.

People in the study were asked to consider six perceptual qualities rather than hunt for a “single giveaway” sign.

Dunn explained each of them below:

Symmetry: “AI-generated faces are often a little too symmetrical. Real faces usually have small natural asymmetries — a slightly uneven smile, one eyelid sitting lower than the other, or other subtle quirks. If a face appears almost perfectly balanced, it’s worth taking a closer look.”

Proportionality: “This relates to how facial features are sized and positioned. While AI can sometimes produce unusual proportions, it more commonly creates faces with very balanced, conventionally proportioned features. Natural human variation is often greater than what AI tends to generate.”

Attractiveness: “AI-generated faces frequently appear unusually attractive or polished. This is a subjective characteristic, but many AI faces have a broadly appealing, idealised look that feels almost ‘too perfect’.”

Distinctiveness: “Consider what would make the face memorable or recognisable in a crowd. AI-generated faces often gravitate towards average-looking features, making them seem more generic and less individually distinctive.”

Expressiveness: “AI faces often display relatively muted or restrained emotions. They may smile or frown, but the expressions can lack the subtle emotional cues and complexity that are common in genuine photographs.”

Memorability: “Closely related to distinctiveness, AI-generated faces are often surprisingly difficult to remember. Because they tend towards average features, they don’t leave as strong a mental impression as many real faces.”

Because AI-generated faces have become so convincing, Dunn said it is “often the overall impression that matters rather than one obvious flaw” when trying to pick them.

“One important point we emphasised is that these characteristics are intentionally somewhat overlapping and subjective,” he said.

“There is rarely a single ‘tell’ that identifies an AI-generated image. Instead, the goal is to become familiar with the patterns AI tends to produce and develop an overall sense of when something feels slightly off.”

Researchers say all participants in the study improved, and high performers could spot a fake almost every time.

“It was amazing to see the dramatic improvement in people’s ability to detect AI faces,” Dawel said.

“We’ve shown our training is effective for some of the most convincing fakes available.

“We are also working on how to optimise the training – making it shorter and ensuring the benefits last over time.’’

The research was replicated by a team at the University of Victoria in Canada, which showed “the findings weren’t a fluke”.

The study comes amid a major shift in the landscape for artificial intelligence in Australia.

On Wednesday, Prime Minister Anthony Albanese announced he would oversee an Office of AI within his department, as the government prepares to introduce new legal standards for AI and data centres.

The standards will set rules for large data centres, including an obligation to underwrite their own power supply, pay their full share of connection costs so energy bills are not impacted, reduce power when needed, and be water efficient. The government has also made big promises to artists and creators, with the shift to ensure their copyright is not infringed upon by AI companies. 

“This world-leading framework is about Australia choosing to shape the future rather than letting the future of AI shape us,” Albanese said.

“This framework is about protecting our national interests and ensuring certainty for growth, jobs and investment.”

RMIT University distinguished professor Lisa Given, co-founder of the Centre for Human–AI Information Environments, said Australians have a wide range of concerns around AI, from energy and water-guzzling data centres near residential neighbourhoods to potential job losses and copyright infringements.

“Addressing such varied challenges, particularly when AI technologies are evolving at such a fast pace, warrants the coordinated approach this office will take,” Given said.

“Given the known risks posed by generative AI technologies like deepfake images that mislead customers or hallucinated content that misinforms chatbot users, Australians need government intervention to protect them from harm.

“The creation of this office marks a significant shift in the government’s overall approach to governing AI, towards being more hands-on and proactive.”

A big Meta miss

Change strategist and trends broadcaster Michael McQueen told Sunrise this week that Australians had embraced AI at a remarkable pace, with many naming their chatbots and even assigning them genders.

“It’s so familiar now in a way that’s unthinkable two years ago,” he said.

But some boundaries remain, as Meta found out when privacy concerns forced it to ditch an AI tool it had just launched this month.

Muse Image allowed users to generate images using public Instagram accounts, with the tech giant saying its “intent was to provide a useful creative tool and to give people control ⁠over whether their public content could be referenced in this way”.

“We’ve heard the feedback that ⁠this feature missed the mark, so it’s no longer available,” Meta said.

SAG-AFTRA, the ⁠union representing actors and other media professionals, welcomed the about-face.

“With the dangers of non-consensual digital ​replicas well ⁠known to all, a feature that encouraged that behaviour is unwise,” a union spokesperson said.

The ANU Emotions and Faces Lab would like to hear from people interested in undertaking the AI face detection training or participating in other AI face studies.