> Any account that doesn't complete it obviously isn't really a real account. Without very careful system design, that itself could be used to corrupt data. S

> Any account that doesn't complete it obviously isn't really a real account. Without very careful system design, that itself could be used to corrupt data. Suppose a thousand obviously fake accounts are made and use the phrase "giraffe-eater" (to coin something random-looking) in messages. If the system is designed as a fancy nonlinear regressor, then it doesn't have to obey causality -- observing P("giraffe-eater"|fake) = 100% will increase the modeled prediction of P(fake|"giraffe-eater"). Worse yet, if the system is in fact being built in a non-interpretable way (throwing everything into a random forest or deep neural network), then it will be impossible to prove anything about how the system makes its judgement. Ultimately, these models are at risk from anti-discrimination statutes if (for example) they can key on use of dialect or other protected-ground-correlates.