Tumblr uses a machine learning system to scan user accounts and posts for spam behavior — things like mass-following, repetitive posting, or coordinated bot activity. The system scores accounts on how likely they are to be spam and flags suspicious ones for a human reviewer to investigate, rather than automatically banning them.

Tumblr's engineering team documented the spam detection pipeline in an August 2019 post on its engineering blog. The system uses semi-supervised machine learning — meaning it learns from both labeled examples of spam and unlabeled behavioral data — and achieves 98% accuracy in identifying spammers. It runs on Spark ML for model training and uses Redis to store spam probability scores for rapid lookup. Tumblr has acknowledged that AI-generated spam content is making the problem progressively harder, as AI tools allow bad actors to produce convincing fake posts at scale.