Human-in-the-loop moderation

Human-in-the-loop moderation means software makes the clear calls and sends the uncertain ones to a person, whose decision is final and recorded. In practice it is a review queue: items enter when a flag's confidence falls in a middle band, when a user reports them or when someone appeals; a moderator chooses an action; and your system carries it out.

What goes into the queue

The band starts at 0.5: the API raises a flag at its threshold, 0.5 by default, and findings below it never come back, so your own lines sit above it. The docs give 0.5 to 0.7 as an example band for human review (flags).

The loop, step by step

1. Your server sends the item with your own ids in metadata, at once or with "async": true.
2. The result arrives in the answer, or as a moderation.completed or moderation.flagged webhook (async requests).
3. Your rules decide: publish, act above your blocking line, or hold for review.
4. A moderator opens the item in the dashboard's Logs, or in your own tool, and takes one of your actions.
5. Detectivision sends a moderation.action webhook to each of your endpoints that takes that event, and your system hides, restores or bans.
6. The request's state moves from pending to actioned, and GET /api/v1/moderation/{id} lists the actions taken.

Detectivision never carries out an action itself (actions).

What a reviewer needs

• The content as users saw it. Detectivision does not keep images (image moderation), so load them from your own storage.
• The result. The flags raised, their confidence, and the rule that held the item.
• Context. The thread, listing or profile around the item and earlier decisions about the same user, found through the ids in metadata.
• The rule. Your written policy for that flag, with examples on both sides of the line.
• A note field. Why they decided, for appeals and for training new reviewers.

Actions and the audit trail

You define up to 20 actions in the dashboard, each with a key your system receives, such as hide, restore or ban. Moderators take them in Logs, up to 600 a minute, or your own tool takes them through the API with a key that has write permission, up to 60 a minute; above that the API answers 429 with Retry-After. Taking actions and sending their webhooks is free.

curl https://api.detectivision.ai/api/v1/moderation/1042/actions \
  -H "Api-Key: $DETECTIVISION_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"action": "hide", "note": "insult aimed at another user"}'

{
  "type": "moderation.action",
  "timestamp": "2026-10-16T12:00:00.000Z",
  "data": {
    "action": { "key": "hide", "label": "Hide comment" },
    "note": "insult aimed at another user",
    "by": { "type": "api", "name": "review tool" },
    "item": {
      "id": 1042,
      "type": "text_moderation",
      "flagged": true,
      "flags": [{ "flag": "insult", "confidence": 0.62 }],
      "metadata": { "userId": "u_123", "commentId": "c_456" },
      "createdAt": "2026-10-16T11:58:10.000Z"
    }
  }
}

Example request and webhook payload (format from the API reference; values illustrative). by.type is user for the dashboard and api for an API key, named by the key's label. Logs shows each action's delivery state for 30 days. Keep your own record of who decided what, when, why and under which rule: the yearly figures below need it for longer.

Where the law requires a person

Who is who. Turkey's Presidency of Cyber Security (Siber Güvenlik Başkanlığı) issued a regulation, published in the Official Gazette (Resmî Gazete) on 10 October 2026 and in force from 1 November 2026, for social network providers reachable from Turkey. A social network provider (sosyal ağ sağlayıcı) lets users create, view or share content such as text, images, sound and location for social interaction (Article 4). The EU Digital Services Act (DSA) covers intermediary services offered to users in the EU; an online platform is a hosting service that makes users' content public at their request (Article 3).

The full Turkish rules are on Turkey's under-15 social media rules.

Looking after reviewers

• Blur by default. Show flagged images blurred, with a click to reveal. Detectivision returns no blurred or censored copy, so blur in your own review screen, for example with the CSS filter: blur().
• Share the hardest queues. Rotate people through violence and self-harm items, and let them pass an item to a colleague.
• Keep the rules close. When reviewers disagree on an item, add it to the written rule as an example.

Measuring the queue

Track these without borrowing targets from elsewhere:

• items entering per day, by source and by flag;
• the backlog, and the age of the oldest item;
• time from entering the queue to an action, per flag;
• the share of band items reviewers confirm: near all, and your blocking line may sit too high; few, and the band catches too much harmless content (false positives);
• the share of decisions reversed on appeal.

For where the queue sits in a full system, see content moderation system design.

What Detectivision does not do

• It provides no reviewers. Logs, your own actions and webhooks are tools; the decisions are your team's.
• It never acts on your platform. Your system hides, restores and bans.
• It returns no censored copies or word positions. Flags judge the whole text or image.
• It does not keep images. Your review screen needs your own copy.
• It does not check video, audio or live streams yet. They are coming soon.

See how actions work