How moderation works

Automatic moderation (AI + filters)

Automatic moderation is the first line of defense. It runs in real time on all generated content.

How it works:

1. Keyword filters. A list of forbidden words/expressions (common insults, slurs, hateful terms) triggers automatic masking. The message is replaced with "[Message hidden by moderation]" and the author is notified.

2. Contextual AI detection. We use an AI model trained on thousands of examples to detect inappropriate content beyond simple keywords. For example, "You're really [disguised insult]" is detected even if the insult is creative.

3. Spam detection. If you send 10 identical messages in 1 minute, or if you send suspicious external links, the system automatically mutes you for 15 minutes.

4. Manipulation detection. If multiple accounts coordinate their messages (same content, same timing), the system detects the pattern and applies measures (temporary mute, escalation to human moderation).

Limits: it can produce false positives (a legitimate message hidden). In that case, the user can appeal.

Reactive human moderation (on reports)

When a user reports content, reactive human moderation takes over.

How it works:

1. The report lands in the moderation queue.
2. A human moderator picks up the ticket. They review:
   • The reported message or behavior.
   • The context (the entire conversation).
   • The reported user's history (repeat offender or first time).
   • Any multiple reports (if several people reported, that's a strong signal).
3. Decision made within 24h on average, sometimes up to 7 days for complex cases.
4. The reported user receives the decision (warning, mute, ban, etc.).
5. The user who reported receives a follow-up message: "Your report has been handled. Sanction applied: [type]" or "Your report was reviewed, but the reported content did not violate our rules."

Team: ~25 full-time-equivalent human moderators, based in Yaoundé, Douala, Dakar, and Paris. 24/7 coverage through rotation.

Volume: moderation handles about 5,000 reports per day on average, with peaks of up to 15,000 during big events (major matches, elections).

Proactive human moderation (pattern monitoring)

On top of the reactive layer, proactive moderation watches for suspicious patterns without waiting for reports.

Cases monitored:

1. New accounts with suspicious behavior. An account created yesterday that posts 100 messages in 1 hour, or that follows 500 people right away, gets analyzed.

2. Collective fraud patterns. Multiple accounts that "match" against each other in the arena to launder money, or that vote for each other in challenges.

3. Shifts in overall toxicity. If a topic (a country's politics, for example) suddenly becomes very toxic, we can strengthen moderation on that topic.

4. Emerging trends. New types of spam, new forms of phishing, new problematic content detected on other platforms - proactive moderation updates the filters.

Team: ~5 moderators dedicated to proactive work, plus a technical team of anti-fraud engineers.

Transparency and published reports

Ktkarena publishes a quarterly moderation report in the "Transparency" section of the website. This report contains:

• Total number of reports received during the quarter.
• Breakdown by type (insults, spam, fraud, etc.).
• Decisions made (sanctions applied, reports dismissed).
• Average processing times.
• Number of appeals and success rate.
• Number of permanent bans.

It's our commitment to transparency. We don't hide moderation challenges - we lay them out and improve.