HomeEverythingEducation
Equities & FundsCrypto & Digital AssetsAI & TechnologyBusiness & CorporateUS Politics & PolicyGeopolitics & Global RiskMacro, Rates & FXCommodities & EnergyEuropean Politics & MarketsAsia-PacificReal Estate & Property
Story archiveAll categories
← All Stories

AI deepfakes are getting harder to spot, but training can help

Created at 11 Jul · 12:12 AM1 source↑ Market-relevant
IN SHORT

Researchers have found that people can be trained to identify AI-generated deepfake images with increased accuracy, moving from around 40% to 80% after just an hour of instruction. The training focuses on subtle perceptual qualities rather than obvious flaws.

✉Newsletter

PiQ Daily

Pick your topics. Get only what matters, on your cadence.

Key Numbers

40%pre-training accuracy in spotting AI fakes
80%post-training accuracy in spotting AI fakes
1 hourtraining duration for accuracy improvement
£25mamount transferred in a deepfake boss scam

Who's Involved

Clare Sutherland
Psychologist from the University of Aberdeen leading research
Amy Dawel
Director of the Australian National University Emotions and Faces Lab
StyleGAN3
AI image tool used to create deepfake faces
Katie Jones
Fictitious deepfake profile used in alleged Russian intelligence operation

↳ Why This Matters

The increasing sophistication of AI-generated deepfakes poses significant risks for fraud and disinformation, making it crucial for individuals to develop the ability to identify them to protect themselves and maintain trust in digital media.

Key facts

  • AI-generated images are becoming increasingly realistic and difficult to distinguish from real photographs.
  • A new training approach helps people identify AI deepfakes by focusing on subtle perceptual qualities.
  • Participants in studies increased their accuracy in spotting AI fakes from approximately 40% to 80% after training.
  • The training highlights characteristics such as symmetry, proportionality, attractiveness, distinctiveness, expressiveness, and memorability in faces.
  • The ability to detect deepfakes is important for preventing fraud and political espionage.
  • AI image generation also has beneficial applications, such as visualizing the potential appearance of missing children over time.

Artificial intelligence has advanced to a point where AI-generated deepfake images are increasingly difficult to distinguish from real photographs. Researchers are exploring whether humans can be trained to identify these sophisticated fakes.

A study led by psychologists from the University of Aberdeen and the Australian National University has found that training can significantly improve people's ability to spot AI-generated faces. Participants in the study typically increased their accuracy from around 40% to 80% after receiving training, which focused on subtle perceptual qualities rather than obvious flaws like extra fingers.

The training draws attention to six key areas: symmetry, proportionality, attractiveness, distinctiveness, expressiveness, and memorability. Researchers noted that AI faces often appear more attractive, generic, and less emotionally expressive than human faces. They also highlighted that AI models, often trained on data predominantly featuring young white individuals, may struggle to accurately recreate faces of other demographics.

Professor Amy Dawel, who is part of the research team, explained that the AI is learning from its mistakes, making traditional methods of spotting fakes less effective. The training aims to develop a 'gut feeling' for AI imposters by attuning individuals to their characteristics.

The ability to discern AI fakes is crucial due to the potential for fraud and political espionage. The report cited a scam where an employee transferred £25 million after a video call with a deepfake of their boss, and a past investigation into a fictitious LinkedIn profile allegedly created by Russian intelligence.

Despite these risks, researchers also acknowledge positive uses for AI image generation, such as visualizing the aging of long-missing children. The findings suggest that while AI is rapidly improving, humans can still be trained to identify its creations, though continuous practice is necessary.

Frequently asked questions

Training can significantly improve accuracy, with participants typically increasing their success rate from about 40% to 80% after an hour of instruction.

The training focuses on symmetry, proportionality, attractiveness, distinctiveness, expressiveness, and memorability, noting that AI faces often appear too perfect or generic.

The ability to identify deepfakes is vital for preventing fraud, such as financial scams, and mitigating risks associated with political espionage and disinformation.

Yes, AI image generation can be useful for creative purposes and applications like visualizing how a missing child might look at different ages.

What Happens Next

01Further research may explore more advanced training techniques for spotting AI deepfakes.
02The study's findings could inform the development of new tools and educational programs for deepfake detection.

Get the newsletter.

Pick the topics you actually care about. We'll email when there's news worth your time, on the cadence you choose. Cancel any time from your account.

Cadence

How It Developed

AI has become adept at creating realistic images, making it difficult to distinguish real from fake.
Researchers have developed training methods to help people identify AI-generated deepfakes.
Training focuses on subtle perceptual qualities like symmetry, proportionality, attractiveness, distinctiveness, expressiveness, and memorability.
Participants' accuracy in spotting AI fakes increased from about 40% to 80% after training.
The ability to spot deepfakes is crucial for combating fraud and political espionage.
AI also has potential positive uses, such as visualizing missing children's aging faces.

Sources

T1
See if you can spot an AI deepfake with our testBBC News

Related Stories

Chinese parents use AI to guide university degree choices
11 Jul · 1:06 PM
Ukraine's Secret Drone Factories Illustrate Shift in Military Spending
11 Jul · 12:11 PM
Australia Debates Copyright Law Changes Amid AI Investment Push
11 Jul · 8:11 PM
Chinese team develops wearable dopamine patch for tracking depression and Parkinson's
11 Jul · 12:06 PM
AI energy demand spurs push for cleaner power alternatives
11 Jul · 1:27 PM