Module · AI moderation
Moderation that handles 85% before the team touches it.
Hybrid AI + human moderation pipeline. The model auto-handles 85% of incoming contributions, your team arbitrates the 15% that's ambiguous. DSA-grade logs by default. Trained on 1M+ comments from European newsrooms.
- 85% of toxic content filtered automatically Der Spiegel deployment
- 6 langs native moderation (FR · DE · IT · ES · PT · EN) Multilingual model
- DSA transparency reports auto-generated, exportable on demand Built-in
Why this module
Manual moderation scales to nothing.
A regional daily can hit 187,000 comments a year. Reviewing each by hand is one full-time moderator per shift, and the team still misses the worst content during off-hours. The AI-first model flips the economics: the model handles the clean cases and the clearly toxic ones, the team only sees the 15% that needs judgment. That's the difference between drowning and operating.
Moderation queue
Today · 1 412 comments processed
- FLAG
"Cette ministre est une incompétente, elle devrait démissionner avant qu'on ne la pousse dehors."
personal-attack · unverified-claim · borderline-rhetoric
0.62 - REJECT
"CLICK HERE to earn 5000€/month from home, no skills required! Free signup at example.com"
spam · commercial-promotion · external-link-bait
0.94 - APPROVE
"Le rapport INSEE de mars contredit la tendance évoquée. La part des salaires en pourcentage du PIB a en réalité baissé sur 10 ans."
sourced · substantive · on-topic
0.04
What you get
Built for newsroom workflows.
- 01
Three-stage pipeline
Every contribution goes through (1) a publisher-defined blacklist + suspicious-words list, (2) algorithmic moderation that auto-approves clean cases and auto-rejects clear violations, (3) a human queue for everything ambiguous. Each stage is auditable.
- 02
Social media moderation
The same pipeline can moderate the comments on your Instagram, YouTube and Facebook channels in real time, filtering illegal content, scams and unreadable messages. Around 95% is automated, every accepted or rejected item is visible in the admin, and you can override any decision. See <a href="/social-media-moderation">social media moderation</a>.
- 03
Logora moderation team as a service
Optional managed service : our German-speaking team reviews the human queue 1 to 4 times a day, with the timing aligned to your traffic peaks. You stay in control of the rules, we operate the queue.
- 04
Keyboard shortcuts + multi-select
Moderators accept (Ctrl+A), skip (Ctrl+S), reject with one of six reasons (Ctrl+1 to 6). Multi-select handles batches of related contributions in one action. Decisions persist : re-opening a contribution shows the prior call.
- 05
Six DSA-compliant rejection reasons
Incivility · Inappropriate language · Personal attack / hate · Incomprehensibility · Off-topic / advertising · Repetition. Each user sees the reason in their profile, with redress information per DSA Article 17.
- 06
Pre-moderation toggle for sensitive debates
Pre-moderation : enable on individual debates (politics, elections, public-safety topics) and every contribution lands in the human queue before publication. Trade-off between latency and editorial control, decided per debate.
- 07
User bans + transparent reader profile
Configurable bans : 1 day, 1 week, 1 month, permanent. Bans are surfaced in the user's profile with the reason. Every rejected contribution stays visible to its author with the rejection category, in line with DSA transparency.
Try the moderation engine
One engine. Every editorial charter.
Pick a media context — sport, politics, people, local news — then type any reader comment. The same Logora engine running on production newsrooms applies the matching editorial charter and returns a verdict, a reason, a toxicity score, and the categories it flagged. Disagree? Tell it why. It learns for the rest of your session.
- APPROVE — respectful, on-topic, no slurs, contributes substance. Auto-publishes on production.
- FLAG — borderline language, unverified claims, ambiguous tone. Routes to human review queue.
- REJECT — hate speech, threats, spam, harassment. Auto-blocked with a DSA Article 14 statement of reasons sent to the user.
—
What should the verdict have been?
Want this calibrated for your newsroom?
Book a 30-min call →0 adjustment(s) active for this charter ·
Same engine, in production at
- Der Spiegel Germany
- Milenio Mexico
- and other European publishers FR · DE · IT · ES · PT · EN
Same prompt calibration, same threshold defaults as a production deployment. On your tenant we add publisher-specific rules, custom thresholds per article category, and your team takes final authority on FLAG verdicts via the moderation admin.
In production
Used by Der Spiegel.
Der Spiegel moved from manual moderation overload to a hybrid pipeline that filters 85% of toxicity before human review. The moderation team has time again, and the reader experience is measurably civil.
85%
of toxic content auto-filtered, freeing the team to focus on borderline cases
Read the full caseWhat changes
Before, and after.
Before · Manual moderation only
- Team drowns at peak hours
- Toxic content visible during off-shifts
- No structured audit trail
- DSA transparency report rebuilt by hand
- No multilingual coverage without doubling the team
After · Logora hybrid moderation
- 85% auto-handled, team sees the 15% that matter
- Real-time scoring at submission, before publication
- Every decision journalised with reason text
- Transparency report exportable in minutes
- Six native languages on the same pipeline
Common questions
What publishers actually ask.
-
Do we have to use Logora's comment widget to use the moderation module?
No. We can moderate any comment stream you send us (existing widget, Facebook, Disqus, custom). The full stack works best together, but the moderation API is standalone.
-
What happens to a comment scored ambiguously by the model?
It lands in the human queue with the toxicity score, the model's reasoning, the article context, and the user history. The moderator decides. Time-to-decision drops below 30 seconds with the keyboard shortcuts.
-
How does the blacklist and the suspicious-words list work?
The blacklist contains words that auto-reject any contribution containing them, with the standard reason "Inappropriate language" or "Personal attack" depending on the term. The suspicious-words list contains terms whose meaning depends on context ("victim", for example) and routes the contribution to the human queue instead of rejecting outright. Both lists are publisher-editable.
-
Can we override the model in real time?
Yes. Editors and head-moderators can change thresholds, force-approve, force-reject, ban specific users (1 day, 1 week, 1 month, permanent), and retrain on specific batches. The model is the default; your team has final authority.
-
How does this interact with the DSA?
Directly. Every automated decision generates an Article 17 statement of reasons. The annual Article 24 transparency report is assembled from the same log. Compliance is built-in, not bolted on. See our full DSA mapping for the per-article rundown.
-
Can we enable pre-moderation on specific debates?
Yes. Pre-moderation is a per-debate toggle : when active, every contribution lands in the human queue before publication, regardless of the algorithmic score. Recommended for politically sensitive debates (elections, immigration coverage, hot societal topics).
-
Can the Logora team operate the moderation for us, or do we have to staff it?
Both options exist. You can run the moderation in-house with your editorial team (the admin is built for that), or you can delegate the human queue to our Logora moderation team. Our German-speaking moderators review the queue 1 to 4 times a day, timing aligned with your traffic. You stay in control of the rules and the rejection labels, we operate the workflow.
A moderation pipeline your team can stop dreading.
A 60-min call. We show you the moderation admin live with real data, walk through the DSA logging, and come back with a pilot scope on one of your sections.