A toxicity score is a number, often between 0 and 1, that a model assigns to a piece of text to estimate how likely it is to be perceived as toxic. Google Jigsaw's Perspective API, for years one of the most widely used examples, returned scores like this for comments, though Google has announced it will be retired after 2026.‍

‍

Why it matters

Scores let a platform sort a huge stream of comments. Very high scores might be hidden automatically, middle scores sent to a reviewer, and low scores left alone. They're a triage tool, not a verdict. A score says something about how the text reads, not whether it breaks a specific policy, and models can be wrong in patterned ways. Reviews of these tools have documented cases where they over-flag certain dialects and mentions of identity groups, and where mild profanity in harmless sentences pushes scores up.

‍

‍Setting thresholds

Teams typically choose thresholds by running a sample of comments through the model and comparing the scores with human labels, then picking cut-offs that balance wrongly hidden comments against missed abuse. The middle band, where the model is unsure, is usually where human review adds the most.

‍

‍A common misreading

A score of 0.8 doesn't mean a comment is 80% toxic in severity. For Perspective, the score reflected the likelihood that a reader would perceive the comment as toxic. A mild insult and an extreme one can both score high. Severity, targets, and context still have to come from a human or a different model.