What "Fooling" Means Here
For a lookalike search, "fooled" doesn't mean a security breach. It means the model rates two faces as more alike than people would, or a celebrity appears in results more often than their face really justifies. Three different things cause it: similar photos, central faces, and, in research labs, deliberate attacks.
Fooled by Similar Photos: What Ollie's Data Showed
When the Ollie team compared every celebrity in the index with every other (over 17 million pairs), two groups turned up in far more close pairs than anyone else:
- Tennis players, led by Barbora Krejčíková, who turned up in more of the closest 1,000 pairs than anyone else. Tennis photos are taken with long lenses from the stands, in bright outdoor light, often mid-shot with a focused, straining expression. The photos share a lot before the faces even come into it.
- K-pop and Korean drama stars. Stage and promotional photos often share lighting, styling, hair and makeup, sometimes from the same professional teams.
In both cases, the photos have as much in common as the faces. The model can't fully separate "photographed the same way" from "looks the same". The full results are in celebrities who look alike, according to an AI.
Fooled by Training Data
There's a second reason some groups produce many close pairs. A large share of the closest pairs in Ollie's data were East Asian or South Asian celebrities. Ollie's model was trained on a research dataset in which lighter-skinned faces are over-represented, so it separates some groups less finely and gives different people higher similarity scores.
People do the same thing with faces they see less often, an effect called the own-race effect. It's a known limitation that the Ollie team discloses, and a reason to read any single close pair as what one model sees, not a fact about anyone's face. See accuracy across demographics.
Fooled by Average Faces
Faces near the centre of face space are close to many people. In 512 dimensions, some of these become "hubs": the nearest neighbour of far more faces than you'd expect. A hub celebrity can be lots of people's top match without a strong specific resemblance to any of them. See the most commonly matched celebrities.
Read more: The Most Commonly Matched Celebrities, and What It Says About Facial Geometry
Get new guides by email
Fooled on Purpose: Adversarial Attacks
Researchers have also shown that face recognition can be fooled deliberately. In 2016, Mahmood Sharif and colleagues at Carnegie Mellon University printed eyeglass frames with carefully computed colourful patterns. Wearing them, a researcher could make a face recognition system fail to recognise him, or even identify him as a specific other person, such as a celebrity.
| Type of fooling | Cause | Example |
|---|---|---|
| Photo similarity | Same lenses, light, styling | Tennis and K-pop pairs |
| Training gaps | Under-represented groups separated less finely | More close pairs among some groups |
| Hubs | Average faces in high dimensions | One celebrity matched by many people |
| Adversarial attack | Patterns designed to trick the model | Printed eyeglass frames, 2016 |
Adversarial attacks are a security research topic. They don't happen by accident in normal photos, and defending against them is an active field.
How to Tell If Your Match Is Real
- Look at the matched photo, not the celebrity's usual look. Is the resemblance in the face, or in the hair, light and pose?
- Check the gap. A top match well ahead of the rest is more specific to you than five bunched scores.
- Try another photo. Real resemblance survives a change of lighting and setting. Photo-driven matches don't.
- Look for a shared feature. If your top five share a brow, a jaw or eye spacing with you, that's real structure.
- Ask a friend. People who know your face are good at spotting whether a resemblance is real.
And use a clear, evenly lit photo to begin with. See the best photo for a celebrity match.