Mathematician Hannah Fry entered the celebrity edition of the BBC series The Traitors and tried to win it with a statistical model rather than gut instinct. The result was instructive, if not a guaranteed route to victory. Her experiment shows how far data can narrow down suspicion in a game built on lying and paranoia, and where it stops helping.
The Traitors traces its roots to a mafia-style party game devised by the Soviet psychologist Dmitry Davydov in the late 1980s. The format later became a major hit for the BBC, and the celebrity series opener drew more viewers than England's World Cup semi-final. In the show, a group of players is split into traitors, who secretly eliminate others each night, and faithful players, who vote out whoever they suspect.
Fry, who joined the celebrity series as a contestant, built a points-based Bayesian model using voting records from earlier series. It assigned scores to each vote, with weights drawn from previous seasons. She noted a few patterns that the model picked up: traitors often cast so-called burn votes against players nobody else supports, and an unexpectedly late arrival at breakfast, later revealed as a traitor, pointed toward faithful status for the person who had come late. Her own strategy did not hold up, as she was made a traitor within 15 minutes of arriving.
When the model was run retrospectively against past rounds, the top or joint-top scorer was a traitor each time. That sounds decisive, but it was not a clean identification method. Innocent players sometimes scored highly, so the model could point to suspects without ever confirming who was lying. It also remains unclear whether a model like this could be used openly during the game, since other players would need to be persuaded by its output.