Most people cannot reliably tell AI-generated faces from real faces. In controlled tests, accuracy for average observers was only slightly better than chance, and even super-recognisers showed only a modest edge. As state-of-the-art face-generation models have shed the obvious artefacts that once gave them away, visual judgment alone is no longer a dependable way to spot synthetic images.
What did the UNSW and ANU study find?
Researchers from UNSW Sydney and the Australian National University tested 125 people, including 36 super-recognisers, using a curated set of faces with obvious glitches removed. Participants judged whether each face was real or AI-generated. According to the summary from UNSW, people with typical face-recognition skills scored only slightly above chance, and super-recognisers did better, but by a slim margin overall (UNSW newsroom). The peer-reviewed paper appears in the British Journal of Psychology (DOI: 10.1111/bjop.70063).
Across participants, confidence in being able to spot synthetic faces was high, but performance lagged, highlighting a consistent overconfidence effect (UNSW).
You can try a short version of the test yourself here: UNSW Face Test.
Why are AI-generated faces so hard to spot?
Early synthetic faces were betrayed by artefacts such as mismatched earrings, distorted teeth, or hair blending into backgrounds. Modern face-generation models have largely eliminated those tells, producing high quality, coherent images. The best outputs now look statistically very typical of human faces and avoid glaring mistakes, which narrows the gap between real and fake at a glance.
The most advanced AI faces tend to be unusually average and highly symmetrical, features that normally signal attractiveness and familiarity, which can make them feel convincingly real (UNSW).
In other words, what used to be reliable cues no longer apply. As systems improve, the set of visual signs that humans can exploit is shrinking, and cues in backgrounds, lighting consistency, or facial-background transitions, while sometimes helpful, are not consistently reliable.
Are super-recognisers better at detecting AI-generated faces?
Yes, but only a little. In the study, super-recognisers outperformed typical participants, yet their advantage was modest compared with their usual superiority on tasks that involve recognising real faces. Some non-expert participants even outscored super-recognisers, indicating substantial overlap between groups (UNSW).
Spotting AI faces does not appear to be a skill that most people can easily learn with simple tips, and even people who are excellent with real faces only gain a limited edge in this task (UNSW).
Super-recognisers in the experiment were more sensitive to a face’s “centrality” and symmetry, using the fact that many synthetic faces look statistically average as a subtle cue. But this cue is probabilistic, not definitive, and it can fail as models diversify outputs.
How should you verify whether a face image is AI-generated?
Because human inspection is unreliable, treat images as claims to be verified. Combine technical and contextual checks:
- Use provenance when available. Look for C2PA content provenance or Content Credentials that show where and how a photo was captured and edited. Absence of credentials is not proof of fakery, but valid credentials add trust.
- Run reverse image searches. Check if the image appears elsewhere or predates a claimed event using tools like Google image search or TinEye.
- Check context and sources. Verify who posted the image first, whether reputable outlets corroborate it, and whether names, locations, and dates align with independent reports.
- Use platform or tool signals. Some services are piloting invisible watermarks, for example SynthID, and platform labels. These are imperfect, but they help when present.
- Apply basic forensic checks cautiously. Inconsistencies in reflections, lighting, text, jewelry, or background geometry can be clues, but modern generators often get these right, and compression can hide tells.
Algorithmic deepfake detection tools exist, but their accuracy can degrade as generative models evolve, and they can be circumvented. Think of them as one input to an image verification workflow, not as final arbiters.
What are the limitations and what comes next?
The UNSW and ANU results apply to curated still images without obvious artefacts. Performance may vary with higher resolution, motion video, or multiple images of the same person, though current trends suggest detection will keep getting harder as generators improve.
On the defensive side, builders and publishers are expanding content provenance so capture devices and editing tools can cryptographically sign media history, and platforms are testing watermarking schemes. Standards like C2PA aim to make provenance portable across ecosystems, and research groups continue to evaluate forensic detectors and robustness.
For individuals and organizations, the practical takeaway is to update assumptions. If an image matters for trust, policy, or safety, do not rely on a quick look. Use provenance, search, corroboration, and, when available, professional verification services.