AI-supported comment moderation for a media company

01Situation

A media company runs a large number of digital news sites with comment and community sections. Moderation relied on keyword filters with static word lists. Since February 2024, the Digital Services Act has required platforms with user-generated content to moderate transparently, give reasons and offer complaint routes.

02Challenge

Keyword filters don't understand context: quotes, irony or factual debates about weapons, drugs or extremism get blocked, while paraphrased hate speech, coded language, leetspeak or approaches to minors slip through. Images in comments aren't checked at all. At the same time, every decision has to be explained, documented and open to appeal – across many sites with their own house rules.

03Solution

Pre-publication moderation that analyses text and images through language understanding rather than word lists, documents each decision with a reason and routes borderline cases to moderators.

04Implementation

• —Connecting comment functions via API so every post is checked before it appears.
• —Text analysis across categories such as hate speech, insults, threats, spam, grooming and personal data – each with a confidence score.
• —Image analysis for nudity, violence, extremist symbols, manipulated images and scam graphics.
• —Configurable thresholds and rule sets per site, derived from each site's house rules.
• —Automatically generated statements of reasons, internal complaint handling and audit logs in line with the DSA.
• —A moderator interface for borderline cases, queues and reporting; phased rollout across sites.

05Impact

• —Legitimate, even controversial posts are wrongly blocked less often.
• —Disguised hate speech, evasion attempts and harmful images are caught before they're visible.
• —DSA duties such as reasons, complaints and documentation are built into the workflow rather than added on.
• —Moderators focus on the cases that need human judgement.