Human in the loop: how AI and moderators work best together

Different strengths

	AI	Human moderators
Speed	milliseconds, 24/7	minutes or hours
Consistency	same answer every time	depends on the person and the day
Nuance	improving, but limited	irony, history between users, intent
Scale	thousands of messages per minute	a few hundred per shift
Cost of a mistake	can repeat it everywhere	usually isolated

Neither is enough alone. The goal is to give each one the work it does best.

A practical split

1. Automatic, reversible actions for clear cases with high scores: hide the message, slow down the user, ask them to rephrase.
2. Review queue for uncertain scores: a moderator sees the message, its scores and the conversation, and decides.
3. People only for serious, hard-to-reverse actions: bans, reports to authorities, decisions about minors.
4. Appeals handled by a person who did not take the first decision.

How to set the score bands is explained in choosing a threshold.

Give moderators the context

A message alone can mislead a person as much as a model. Show the previous messages, the user's history and why the message was flagged (which category, which rule, which score). Context helps people too.

Close the loop

Each human decision is a label: “this was a false alarm”, “this one was missed”. Collected over time, these decisions are the most valuable data to tune thresholds and to improve the next model version.

Be transparent with users

Tell your community that messages are checked automatically, what happens when a message is flagged and how to appeal. Rules people understand are followed more, and mistakes are forgiven more easily when they can be fixed.

Protect the moderators

Moderators see the worst content. Automatic filtering of the clearest cases, blurred previews and limits on exposure protect their wellbeing: another reason to let AI handle the first pass.