Read this part
Ethics & limitations
StarMatch ranks visual similarity between your face and a fixed gallery. That is all it does. It is not an identification system, and the numbers it shows should not be read as confidence that you are anyone.
The threshold does not hold
Face descriptors are conventionally treated as matching the same person below a distance of 0.6. Across the 2,133,145 pairs of genuinely different people in this gallery, the closest pair sits just 0.310 apart — well inside that threshold.
In other words: a rule that is supposed to mean “this is the same person” fires on strangers within a set of only 2066 faces. Scale that to a national database and false matches stop being a curiosity and start being someone’s afternoon in a police station.
For scale: two photographs of Keanu Reeves taken five years apart measure 0.371 apart in this same descriptor space — you can reproduce that on the doppelgänger page. The closest pair of different people in this gallery is 0.310. The strangers are closer together than the two photos of one man. There is no threshold that separates those cases, because they are the wrong way round.
That gap widened as the gallery grew. At 366 faces the closest stranger pair was 0.377; at 2066 it is 0.310. More faces means more chances for two of them to collide, so a system like this gets less reliable as its database gets bigger — which is the opposite of how people assume it works.
- Same person, 5 years apart
- 0.371
- Closest stranger pair
- 0.310
- 5th percentile
- 0.671
- Median stranger pair
- 0.831
- Furthest pair
- 1.218
The gallery is not representative
Faces were drawn from Wikidata, ranked by how many Wikipedia language editions cover the person, and filtered to those with a freely-licensed portrait on Wikimedia Commons. Every one of those steps inherits Wikipedia’s well-documented skew toward men, toward the West, and toward the recent past.
The breakdown below is by the occupation each person was queried under, not what they are best known for: anyone with a single acting credit lands in “actor” because that query runs first. It is a fair picture of how the gallery was assembled, not a biography. No correction is applied for gender, ethnicity or era — the bias is stated rather than quietly smoothed over.
- footballer
- 295
- actor
- 291
- politician
- 264
- comedian
- 213
- singer
- 184
- model
- 184
- film director
- 130
- tennis player
- 112
- television presenter
- 102
- musician
- 91
- rapper
- 79
- basketball player
- 70
- youtuber
- 23
- boxer
- 10
- streamer
- 9
- screenwriter
- 9
Accuracy is not uniform across faces
Published evaluations of face recognition — most prominently NIST’s FRVT work — repeatedly find error rates vary by skin tone, sex and age, often by more than an order of magnitude. The descriptor used here is a 2017-era ResNet trained on a web-scraped dataset and inherits those disparities.
If StarMatch performs worse on your face, that is a property of the model, not of your face.
What is deliberately not built
- No image is uploaded, so none can be retained or subpoenaed.
- No enrolment: you cannot add a private individual to the gallery.
- No search by name to reverse-look-up a face.
- No analytics on results, and no record that a match happened.
The gallery contains only public figures who already have a freely-licensed portrait published about them. That boundary is the point.
If you want it gone
Everyone in the gallery is a public figure with a Commons portrait, but inclusion is still automated and unasked-for. Open an issue on the repository naming the entry and it will be removed from the index on the next build — no justification required.
Built as a portfolio project. It is a demonstration of how face embeddings behave and of what a privacy-preserving architecture looks like — not a product, and not a tool for identifying anybody.