📖 Content moderation glossary
What is model arbitration?
In short
Model arbitration is a technique in which several AI models analyse the same content and their verdicts are compared, so that a moderation decision on a delicate case does not rest on a single model's judgement.
Every AI model has blind spots. Asking a second or third model to review a borderline comment, and comparing their answers, catches many of those errors. When the models agree, confidence in the decision is higher; when they disagree, it is a strong signal that the case is ambiguous.
Example: a comment uses a word that is a slur in one context and a reclaimed term in another. The first model flags it as hate speech, the second reads it as neutral. Rather than picking one at random, the system treats the disagreement as doubt and sends the comment to a human moderator.
Arbitration is usually reserved for delicate cases, since running several models on every comment would be slow and costly. OpenModeration uses its own moderation AI and cross-checks delicate cases with other AI models; new models are plugged in as they come out, and cases that remain uncertain are handled by the OpenModeration team.