Lead: A new study published in Royal Society Open Science finds that most people cannot reliably distinguish AI-generated faces from real ones without brief training. Conducted by researchers including Katie Gray at the University of Reading, the experiments tested both typical observers and so-called super recognizers in controlled trials in the UK. Before training, participants’ ability to spot synthetic faces was close to chance; a five-minute instructional intervention substantially improved detection rates. The findings underscore growing risks from realistic synthetic imagery shared on social platforms and suggest short educational steps can reduce some vulnerability.
Key takeaways
- Study published in Royal Society Open Science by researchers from the University of Reading tested human ability to detect GAN-generated faces.
- In initial tests, typical recognizers identified 30% of fakes and super recognizers identified 41%—results near or below random-guess levels.
- A five-minute training exercise raised accuracy to 51% for typical recognizers and 64% for super recognizers when tested immediately after training.
- Participants were shown faces for 10 seconds in the first task; the trained group evaluated 10 faces in real time in the second task.
- Training focused on telltale rendering errors such as odd tooth alignment, hairline artifacts, and skin-texture inconsistencies.
- The super recognizers were recruited from a volunteer database tied to the Greenwich Face and Voice Recognition Laboratory and had ranked in the top 2% on face-memory tests.
- Authors caution that retention over time was not measured; immediate gains may not persist without reinforcement.
Background
Generative adversarial networks (GANs) and similar image models have advanced rapidly, producing humanlike faces that can be rendered at high resolution and photorealistic detail. Researchers and regulators have warned that these synthetic images enable new forms of impersonation, fraud and misinformation, from fake profiles to doctored medical advisers on social platforms. Prior work has documented technical artifacts in synthetic images, but it remained unclear how well untrained humans could spot those artifacts at scale.
At the same time, a small subset of people—often called super recognizers—perform far better than average on face-memory tasks and are sometimes deployed in security contexts. The research team used that contrast to test whether innate superior recognition skill confers meaningful advantage against modern synthetic imagery. The trial also responds to rising reports of convincing deepfakes circulating on TikTok and other networks, where realistic-looking fake accounts have misled users.
Main event
The investigators ran a pair of experiments. In the first, volunteers saw a single face on screen for ten seconds and decided whether it was a real photograph or an AI-generated image. Typical participants correctly flagged only about 30% of the synthetic faces, while super recognizers reached roughly 41% accuracy—only marginally better than chance.
The team then introduced a brief training module lasting roughly five minutes to a fresh cohort. Training explained common visual glitches—misplaced teeth, asymmetric hairlines, errant ear shapes and inconsistent skin texture—and showed examples. Immediately after training, participants completed a second task in which they viewed 10 faces and judged authenticity in real time, with a brief review of common rendering mistakes at the end.
Performance improved after training: typical recognizers rose to about 51% correct and super recognizers to about 64% correct. Trained participants also spent more time examining images before answering, suggesting the training changed inspection strategies rather than merely raising confidence. The researchers frame the regimen as a low-cost, rapid mitigation that can partially blunt the threat posed by convincing synthetic faces online.
Analysis & implications
The study demonstrates that cutting-edge synthesis can routinely trick unaided human observers, reversing an assumption that people would remain a reliable backstop against bad actors using imagery. That gap matters because visual impressions often anchor trust online—users may assume a realistic face equals a real person, which amplifies scope for impersonation, social-engineering schemes and spread of false claims.
Short, focused training improved detection markedly, but even after instruction accuracy remained far from perfect. Moving from ~30–41% to ~51–64% reduces risk but does not eliminate it, implying that training should be one component of a broader defense strategy that includes platform safeguards, provenance markers and automated detectors.
There are policy and operational implications for platforms, regulators and media literacy programs. Platforms could integrate micro-training nudges, provenance labels, or friction for newly created profile images; regulators might consider standards for required disclosure of synthetically generated imagery in certain contexts. However, technical countermeasures and detection models must keep pace with generative models that continue to improve.
Comparison & data
| Group | Pre-training accuracy | Post-training accuracy |
|---|---|---|
| Typical recognizers | 30% | 51% |
| Super recognizers | 41% | 64% |
The numeric contrast shows that training produced relative improvements of roughly 70% for typical recognizers (30% → 51%) and about 56% for super recognizers (41% → 64%) in absolute performance points. The table summarizes controlled-lab findings; real-world detection will vary with image quality, context, and the presence of additional cues such as metadata or accompanying text.
Reactions & quotes
Researchers described the training as a short, encouraging method that noticeably increased detection performance across participant groups.
Katie Gray, University of Reading (lead author, paraphrased)
Experts note the results show human viewers are no longer a reliable sole defense and recommend combining education with technical verification tools.
Independent digital forensics expert (paraphrased)
Unconfirmed
- Long-term retention: the study measured post-training performance immediately; it is unclear how long improvements persist without refresher training.
- Real-world generalizability: performance against images embedded in social feeds, video, or with accompanying text has not been established by this experiment.
- Cross-model robustness: the study used a set of GAN outputs; detection difficulty may differ for images generated by other or newer models.
- ChatGPT & Turing Test claims: separate assertions that language models have passed the Turing Test are contested and context-dependent; they are not validated by this facial-detection study.
Bottom line
The research shows modern synthetic faces can routinely fool unaided human observers, but a brief, targeted training module significantly raises detection rates. Training is a practical, low-cost mitigation that platforms and educators can deploy as part of a layered response to synthetic-media harms.
Yet training alone is not a panacea: accuracy after instruction remains imperfect and retention over time is unknown, so policymakers and platforms should combine education with technical labeling, provenance controls and automated detection to reduce the risk of impersonation and deception.
Sources
- New York Post — news article summarizing the study (media)
- Royal Society Open Science — academic journal (publisher)
- University of Reading — institutional homepage for lead authors and press context (academic)
- LiveScience — reporting on synthetic-image plausibility and public examples (media)