AI and human review
How Do AI Flags and Human Decisions Fit Together?
A model sorts, a person decides. Classifiers on your tooling or ours flag likely violations and route them to a queue, and a named moderator makes the decision that is logged. The pattern is the same human-in-the-loop review EICRA applies to other AI workflows, so a false positive is corrected by a person, not published by a machine.
Pre-flagging Text, image and video classifiers mark likely violations and the probable harm category, so the queue is ordered by risk rather than by arrival time.
Human decision Every removal, approval, edit or escalation is made by a moderator who reads the item in its language and context and records the policy clause applied.
User reports Items flagged by your users through a report button enter the same queue and are decided on the same policy, so reactive and proactive review do not diverge.
Escalation Anything the policy does not cover, and anything safety-critical, goes to the contact you name with the question we need answered, before any action is taken.
Model feedback Human decisions that disagree with the classifier are collected and handed back to your model owners or to EICRA’s AI model testing team as evaluation data.
Audit trail Flag, decision, clause, moderator and timestamp are stored together, so a single item’s history can be reconstructed when a user appeals or a regulator asks.