The old saying is that everyone has a double somewhere in the world. For most of history that was a superstition; doppelgängers were omens. In the last decade it became a research question, and the answers are more interesting than the legend.
The look-alike study
For years the Canadian photographer François Brunelle has been photographing pairs of unrelated people who look strikingly alike, posed side by side. In 2022, a team led by Manel Esteller at the Josep Carreras Leukaemia Research Institute in Barcelona used those pairs as a natural experiment. They ran 32 pairs through three different facial-recognition algorithms; half the pairs scored as similar as identical twins do on all three. The researchers then compared the DNA of the strongest pairs and found they shared far more genetic variants than random people do, clustered in genes related to facial shape. The pairs also shared height and weight to a surprising degree, and even some lifestyle traits like smoking and education. The paper, published in Cell Reports, made the point plainly: people who look alike tend to share the genetics behind looking alike, even with no family connection.
How rare is a true double?
A 2015 study by Teghan Lucas and Maciej Henneberg at the University of Adelaide asked a stricter version of the question. Using detailed measurements of thousands of faces, they estimated that the chance of two people matching on eight key facial dimensions is less than one in a trillion, and that even with more than seven billion people on Earth, the odds that a single exact pair exists anywhere are only about one in 135. In other words, an exact double is essentially impossible; a "close enough that a stranger would do a double take" double is common. Which one people mean by doppelgänger is the whole difference.
Why look-alikes exist at all
A face is built from a limited number of variable parts, each set by genes with a limited range. Combine enough people and combinations repeat. Populations that share ancestry share more of the underlying variants, so look-alikes are more common within a population than across the world. There's also the averageness effect covered in face symmetry myths: faces near the population average are, by definition, close to many other faces, so people with typical features have more near-doubles than people with distinctive ones.
Why algorithms are good at this
Face-recognition models place each face at a point in a high-dimensional space and measure distances. That's the same operation as "does this person look like that one", performed on hundreds of dimensions instead of the two or three a human focuses on. The Barcelona team's finding that algorithms and genetics agreed is a nice validation: the algorithm's distance reflects something physically real. FaceTest's engine works on the same principle, using a general image network rather than an identity model; here's the walkthrough.
Celebrity look-alikes specifically
Being told you look like a celebrity is partly about faces and partly about exposure. You have seen that celebrity's face thousands of times, from many angles and expressions, which gives your brain a sharp reference point. The same is true for K-pop idols and actors: fans can distinguish members from a silhouette. An AI doesn't know who is famous. When a celebrity or idol test matches you, it's reporting a distance in feature space, nothing more. If it agrees with what people have told you for years, that's two independent measurements agreeing. If it doesn't, one of them is using different dimensions.
Finding yours
The photography project and the study both started with people who happened to meet their doubles. You can skip the coincidence: a face-matching test measures your distance to every face in its cast in under a second. The celebrity and K-pop tests are the closest thing to the classic "who's my famous twin", and the fun part is comparing the answer with what your friends have always said.