Automated and AI content moderation under the DSA: what you must disclose
Filters, classifiers and AI models help small teams keep up with reports. The DSA doesn’t forbid them, but it requires transparency in five places and a human for complaints. This guide shows where and how.
Few small teams can review every report by hand straight away. Filters, classifiers and AI models help with sorting and detection. The Digital Services Act doesn’t forbid automated moderation; its very definition of “content moderation” in Art. 3(t) covers activities “whether automated or not”.
What the DSA does require is that automation is visible: to users, to notifiers and to the public. And in one place it draws a firm line.
Where the DSA mentions automation
| Article | Who | What |
|---|---|---|
| Art. 14(1) | all intermediary services | terms describe moderation tools, including algorithmic decision-making and human review |
| Art. 15(1) | all except micro and small enterprises | transparency report with information on automated means |
| Art. 16(6) | all hosting services | decision notice to notifiers states use of automated means |
| Art. 17(3)(c) | all hosting services | statement of reasons covers automated decision-making and automated detection |
| Art. 20(6) | online platforms (not micro and small enterprises) | complaint decisions not solely automated |
Art. 14(1) is the starting point: your terms describe all policies, procedures, measures and tools used for content moderation, “including algorithmic decision-making and human review”. If you use an AI pre-check, your terms say what it is used for and when a person reviews.
Notices: disclose automation in the decision notice
Under Art. 16(6), if you use automated means to process a notice or decide on it, you include information on that use in your decision notice to the notifier. This covers processing too, not only the decision: a model that sorts or pre-scores notices counts.
One sentence is enough: “Your report was pre-checked automatically. A member of our team made the decision.”
Statements of reasons: whether and how automated means were used
Art. 17(3)(c) requires every statement of reasons to include, where applicable, information on
- the use of automated means in taking the decision, and
- whether the decision concerns content detected or identified using automated means.
These are two separate questions, which gives four combinations:
| Detection | Decision | Example sentence for the statement |
|---|---|---|
| person (notice) | person | “The comment was reported and reviewed by a member of our team. No automated means were involved.” |
| automated | person | “Our filter detected the comment. A member of our team made the decision.” |
| person (notice) | automated support | “An automated pre-check assessed the comment and suggested a decision. A member of our team confirmed it.” |
| automated | automated | “The comment was detected automatically and removed without human review.” |
The DSA doesn’t say explicitly whether an AI suggestion confirmed by a person counts as use of automated means. The wording asks about their use in taking the decision, not only about fully automatic decisions, so disclosing it is the safe choice. The other required elements are covered in statement of reasons under Art. 17 DSA.
If you are an online platform submitting to the Transparency Database, you also state there, separately, whether detection and decision were automated.
Complaints: never solely automated
This is where the DSA draws the line. Under Art. 20(6), online platforms must ensure that complaint decisions “are taken under the supervision of appropriately qualified staff, and not solely on the basis of automated means”.
In practice:
- AI may summarise a complaint, show context or give an assessment.
- A person who is qualified for it, meaning they know your policies and are trained, makes the decision.
- A complaint against an automated decision must not be “confirmed” by the same system.
In Germany, for example, breaching Art. 20(6) can be fined up to EUR 300,000 (§ 33 DDG). More on the process in internal complaint handling under Art. 20 DSA.
Transparency report: purpose, accuracy, error rate
Unless you are a micro or small enterprise, you publish a transparency report every year. Art. 15(1) asks about automation in several places:
- (b) the number of notices processed using automated means;
- (c) information on own-initiative moderation, including the use of automated tools, broken down among other things by detection method;
- (e) any use of automated means for content moderation, with
- a qualitative description,
- the precise purposes,
- indicators of accuracy,
- the possible rate of error, and
- the safeguards applied.
Point (e) is the demanding one. You can only report accuracy and error rates if you measure them. A simple approach: record how often people confirm an AI suggestion and how often they change it, and how many complaints against automatically detected or decided cases succeed. The Commission’s templates in Implementing Regulation (EU) 2024/2835 set the structure. More in the DSA transparency report guide.
In practice: a pre-check, not autopilot
For small app teams, one pattern keeps most of these duties simple: AI pre-checks, a person decides.
- Detect and sort. The AI assesses reported content against your policies and moves urgent cases up the queue.
- Suggest. It proposes a decision and the matching policy.
- Decide. A person reviews the suggestion and accepts or changes it.
- Disclose. The statement of reasons and the notice to notifiers mention the pre-check; your terms describe it.
- Measure. Confirmation rates and successful complaints feed into the transparency report.
The benefit: you get the speed of the AI without a model error reaching your users directly. Complaints go to people anyway. And your statements stay honest, because they describe what actually happened.
Where a fully automatic decision makes sense, such as obvious high-volume spam, keep it narrowly scoped, describe it in your terms and disclose it in every statement of reasons. Actions against accounts are better left to people.
How to do this with Limenia
The AI pre-check is opt-in and uses a European AI provider, with processing in the EU. It assesses content against your policies, suggests a decision and raises the priority where needed. A person makes the decision (see features), and the statement of reasons mentions the pre-check in its text and in the automation field. A narrowly scoped automation is planned; actions against accounts and complaints will stay with people even then. The AI pre-check is included in Pro (5,000 checks per month) and Premium, and available for Basic as an add-on for €49 per month (excl. VAT); see pricing.
Sources
- Regulation (EU) 2022/2065 (Digital Services Act), Article 3(t), EUR-Lexeur-lex.europa.eu
- Regulation (EU) 2022/2065, Articles 14 and 15, EUR-Lexeur-lex.europa.eu
- Regulation (EU) 2022/2065, Articles 16 and 17, EUR-Lexeur-lex.europa.eu
- Regulation (EU) 2022/2065, Article 20, EUR-Lexeur-lex.europa.eu
- Commission Implementing Regulation (EU) 2024/2835 (transparency report templates), EUR-Lexeur-lex.europa.eu