Deepfake faces generated via artificial intelligence (AI) have become so realistic that they routinely fool people, with some research suggesting there may be US$40 billion worth of deepfake-related fraud annually by 2027.
Not only do most people struggle to spot AI faces, but as long ago as 2023 we discovered some AI faces are "hyperreal" - they look more real than actual human faces. We also found people are overconfident they can spot AI faces, with the most confident people making the most errors.
Software-based deepfake detectors do exist, but they can't really explain the reasons for their detections - and they suffer from serious weaknesses . Some can be fooled simply by converting the image type, such as from png to jpg.
But it turns out most people can learn to spot AI faces with an hour or so of practice. In new research published in PNAS, we show there's a straightforward way to improve detection of deepfakes, by training people to pick up the tell-tale clues through experience rather than direct instruction.
The difference between human and AI faces
In our early research, we discovered a key difference between AI and human faces. AI faces are hyperaverage .
This means AI faces tend to be more symmetrical, proportional and attractive than human faces. But they're less expressive and memorable - less likely to stand out in a crowd.
Intriguingly, people can accurately and reliably judge these qualities, but frequently misinterpret the clues. For example, people often think that faces that look a bit odd are AI-generated, when in fact human faces are more likely to have distinctive, unusual features.
Although most people struggle to decide whether a face is AI or real, there is one group who are naturally good at picking up on these clues. So-called super-recognisers , who have exceptional human face perception, seem to be attuned to hyperaverageness, making them better at spotting AI faces.
This made us wonder if, for those of us who aren't super-recognisers, AI detection abilities can be trained like other forms of perceptual expertise .
Learning to spot AI
In our first study, we invited 45 participants into our lab at the Australian National University, and asked them to rate around 100 faces on six qualities that can be used to tell AI faces apart from real ones: distinctiveness, memorability, proportionality, symmetry, attractiveness and expressiveness.
We didn't tell participants how these clues might help them distinguish an AI face from a real one - they had to figure that part out for themselves.
We told participants which faces were AI and which were human, but we didn't tell them that the AI faces were more symmetrical or less expressive, for example. They had to learn these clues through experience rather than direct instruction.
Before and after training, we tested participants' ability to tell AI faces apart from human ones with new faces that were not used in the training.
Training works
In one test, participants were shown three faces - two human and one AI - and asked to select the face that was AI. On this task, average accuracy doubled from 40% before training to 80% afterwards.
Impressively, all participants improved in their AI detection abilities and several achieved close to 100% accuracy. Participants also became faster and more confident in their correct judgements.
To test the robustness of these findings, the Different Minds Lab at the University of Victoria in Canada conducted a replication of the AI detection training with Canadian participants.
The Canadian lab obtained results that were as strong as those reported in the original Australian study. This shows the training is reliable and can work for different groups of people.
The training was also just as effective when it was administered online rather than in person, which suggests it could be a cost-effective remote intervention in deepfake detection.
A promising start
But this doesn't mean we've solved the AI detection problem. Our training used faces produced with one particular generative AI model, called StyleGAN3 .
This is one of the most realistic face generators available, but the technology is advancing rapidly and there are many other models.
Our method has potential to adapt to new models by updating the training images and using multimedia, but we don't yet have evidence that this will work.
The clues we found for spotting AI faces may shift for other models. And other important questions remain: do the training benefits hold up over time? Is the training effective for people of all ages, including older adults or children?
How to improve your chances of spotting AI faces
If you want to get better at recognising AI-generated faces, looking at a lot of examples is a good start. You can see plenty at websites such as Which Face Is Real or This Person Does Not Exist .
While you're looking, bear in mind the six key factors we identified:
- how distinctive is the face?
- how memorable is it?
- how proportional is it?
- how symmetrical is it?
- how attractive is it?
- how expressive is it?
This exercise may improve your deepfake radar. But the more important takeaway is that AI deepfakes are improving very quickly - they can easily fool us, even if we think we can spot them.
The clues are no longer obvious: they are not based on specific details but on facial impressions which people form rapidly and naturally, but which can be misleading.
At the same time, there is hope. We have shown it is possible to train people to detect AI faces. By combining our human-centred approach with algorithmic detection, we may yet keep up in this cat-and-mouse game of advancing technology.
Interested in undertaking the AI face detection training? You can register here .
This research is supported by the Australian government through the Australian Research Council's Discovery Projects funding scheme (DP220101026), awarded to Amy Dawel.
Tanya George was supported by funding from the Commonwealth Government of Australia.
Eric Mah and Jim Tanaka do not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and have disclosed no relevant affiliations beyond their academic appointment.
