You pause on a photo in your feed and the voice in your head asks, Is that real? I felt that same jitter the first time a deepfake flirted with my memory. That hesitation is now the place where this research finds its leverage.
I want to teach you a faster, more reliable instinct than clicking “AI detected” and hoping for the best. The Australian National University team published a study in the Proceedings of the National Academy of Sciences showing that training people to read broad facial impressions—rather than tiny glitches—can nearly double accuracy at spotting AI faces. I’ll walk you through what they did, why it works, and how you can use it in the wild.
On the timeline where thousands of faces scroll by, one face feels oddly “too smooth” — what the lab tested
The Emotions and Faces Lab at Australian National University presented 45 participants with a mix of human and AI-generated faces and asked a simple question: is this image real?
Before training, performance looked like a coin flip—people fared poorly at telling AI faces from real ones. The lab then reframed the task. Instead of hunting for visual glitches, participants rated faces on six broad traits: symmetry, proportionality, attractiveness, expressiveness, distinctiveness, and memorability. After repeated practice blocks—six sessions of 96 judgments each—average accuracy nearly doubled and a few people reached what the authors call “near-perfect accuracy.”
At a job interview where a fake candidate almost slipped through, the switch in strategy saved the day — why global impressions beat pixel-spotting
Real-world detectors—Midjourney, DALL·E, Stable Diffusion outputs, even commercial tools—often flag small errors: stray fingers, warped backgrounds, odd reflections. Those cues are fading as generators improve. The ANU team argues a different cue set works better: the math of averages that drives generative models makes AI faces subtly more typical.
AI-trained models synthesize faces from statistical averages of tens of thousands of faces. The result? Faces that are more symmetrical and attractive on average, but less distinctive, less expressive, and less memorable. AI faces often read like wax mannequins under studio lights.
That pattern is slippery to spot at a glance, but it’s something humans sense unconsciously. The study’s training sharpened that unconscious sense into a reliable decision rule: watch for too much typicality and for low memorability.
How can I tell if an image is AI-generated?
Start with six broad judgments, not tiny artifacts. Ask yourself: does this face feel unusually symmetrical or generically attractive? Is it forgettable? Does the expression really read as an emotion? If several answers point to “too perfect,” treat the image with suspicion. This approach worked in the lab across thousands of trials.
At a dating app where dozens of avatars blur into one another, you need a fast test — how the training actually works
The protocol was simple and repeatable: view a face, make a real-or-AI decision, then rate the face on the six qualities. Repetition taught participants to associate patterns of ratings with AI generation. The learning was implicit—participants weren’t told to search for symmetry; they learned to notice it by doing.
Think of the method as a metal detector for sameness, pinging when a face is suspiciously average. After training, participants made far fewer false calls than when relying on conventional cues or commercial detectors (some of which can cost about $20 (€18) per month and still produce false positives).
Can humans be trained to detect deepfakes?
Yes—at least for still images. The experiment’s gains were large, repeatable, and achieved with a short online training regimen. The catch: the paper’s results are confined to image generators. Voice and video deepfakes add temporal and audio cues that may require different training.
At your kitchen table, you can sharpen this skill—what it means for you and for platforms
If you care about disinformation, nonconsensual imagery, or simple trust on social platforms, this research gives a practical route forward. Platforms—think X/Twitter, Meta, Reddit—or newsroom verification teams could deploy short, scalable training modules to help moderators and readers improve detection rates. Some detection systems like Pangram and other commercial services have tried automated flags; humans trained in global impressions could be the guardrail those automated systems lack.
There are limits. The method won’t stop a determined fraudster who layers real photos and careful editing. It won’t automatically apply to audio or motion-based deepfakes. Still, the upside is clear: distributed human judgment, trained to notice a few broad signs, can raise the cost of deception and protect people who are targeted by AI-assisted abuse.
Are AI-detection tools reliable?
They can help, but they aren’t a panacea. AI detectors sometimes produce false positives and often operate as black boxes—Pangram and similar tools have been criticized for inconsistent outputs. Pair those tools with trained human reviewers who focus on the six traits and you get the best of both worlds: scalable automation plus informed human judgment.
This isn’t a final answer to AI slop imagery, but it’s practical, cheap, and surprisingly effective—will you teach your feed to be suspicious or will you keep trusting the eye’s first impression?