GuidesAI and agents
Moderating content with AI: reviews, comments, listings
Three outcomes
- 1PublishedWithin the rules, beyond doubt.
- 2To reviewIrony, edge case, doubt: a human decides.
- 3RejectedClear violation, reason recorded, author informed.
Rules, written down
| Category | Example |
|---|---|
| Insults, harassment | Attacking a person |
| Personal data | A third party’s phone number or address |
| Spam, advertising | Link to another shop |
| Off topic | A review that isn’t about the product |
The model receives these rules and returns, for each piece of content, a category, an outcome and a one-sentence justification.
Setting the thresholds
Strict
Lenient
Follow-up
- A sample of automatic decisions reviewed every week.
- Appeals counted: a rising rate signals a badly written rule.
- Rules versioned, so you know which version decided.
For structured output: reliable JSON from an LLM. The legal frame: AI, GDPR and the AI Act.
Frequently asked questions
- Can the AI reject on its own?
- For obvious cases (insults, spam, personal details), yes, with a recorded reason. Edge cases stay human.
- Should the author be told?
- Yes. Saying why content was rejected and how to appeal avoids frustration. In Europe, the Digital Services Act (DSA) points the same way for hosting services.
- Are negative reviews rejected?
- No. A negative but respectful review is published. Filtering it distorts the reviews, and may be illegal.