Regulation explainerTechnology, Media & Telecommunications

DSA content moderation: duties by service tier and a pipeline that evidences them

DSA content moderation duties depend on what your service is: every hosting service needs notice and action and statements of reasons, online platforms add complaints, trusted flaggers and database submissions, and very large platforms add risk assessments and audits. The practical answer is a moderation pipeline whose decision records already contain what each duty asks for, so reports and redress come from logs rather than from a scramble.

Reviewed 8 min read

On this page
  1. Which Digital Services Act tier your service falls into
  2. Moderation duties that stack from hosting service to very large platform
  3. Instruments that govern moderation decisions in the EU and UK
  4. A moderation pipeline whose records double as compliance evidence
  5. Turning DSA duties into engineering work
  6. A hypothetical marketplace handling counterfeit listings
  7. Where AI-assisted moderation programs fall short of the DSA
  8. Questions and answers
  9. Sources

Which Digital Services Act tier your service falls into

The Digital Services Act (Regulation (EU) 2022/2065) applies to intermediary services offered to recipients in the EU, wherever the provider is established1. Within that, obligations stack by tier. A hosting service stores information on behalf of users. An online platform is a hosting service that also disseminates that information to the public, such as a social network, video service or marketplace. A very large online platform (VLOP) or very large online search engine (VLOSE) is one designated by the Commission once it reaches 45 million average monthly active recipients in the Union1.

Micro and small enterprises are exempt from most additional online platform obligations unless designated as very large1. The exemption does not reach hosting-level duties, so even a small service storing user content needs a notice mechanism and statements of reasons.

Moderation duties that stack from hosting service to very large platform

DutyHosting serviceOnline platformVLOP or VLOSE
Notice and action (Article 16)RequiredRequired, with trusted flagger priorityRequired, with trusted flagger priority
Statement of reasons (Article 17)Required for each restrictionRequired for each restrictionRequired for each restriction
Submission to the Transparency DatabaseNot requiredEvery decision and statement of reasonsEvery decision and statement of reasons
Internal complaints and out-of-court settlementNot requiredRequired (Article 20 and Article 21)Required (Article 20 and Article 21)
Transparency reportAnnual, unless micro or smallAnnual, with extra platform dataAt least every six months
Systemic risk assessment, mitigation and auditNot requiredNot requiredRequired yearly (Article 34, Article 35, Article 37)

Summarized from Regulation (EU) 2022/2065, Articles 15 to 421. The micro and small enterprise exemption removes most online platform duties for qualifying providers that are not designated as very large.

Instruments that govern moderation decisions in the EU and UK

Cite the primary text in your own policy documents; guidance and codes change more often than the acts.

Regulation (EU) 2022/2065 on a Single Market For Digital Services (Digital Services Act)

European Union

Applies whenAn intermediary service is offered to recipients established or located in the EU, whatever the provider's place of establishment1.

  • Process notices in a timely, diligent, non-arbitrary and objective manner and disclose any use of automated means in doing so (Article 16)1.
  • Give a clear and specific statement of reasons for each restriction, stating the facts relied on, the legal or contractual ground, redress options and whether automated means were used (Article 17)1.
  • Online platforms submit decisions and statements of reasons, without personal data, to the Commission's public database (Article 24)1.

Commission Implementing Regulation (EU) 2024/2835 (templates for DSA transparency reports)

European Union

Applies whenA provider publishes the transparency reports the Digital Services Act requires; the templates apply to reporting periods from 1 July 20252.

  • Use the harmonized quantitative and qualitative templates and publish within two months of each reporting period2.
  • Report accuracy indicators for automated moderation, including precision and recall, and describe the possible rate of error2.

Online Safety Act

United Kingdom

Applies whenUnder the Online Safety Act 2023, a user-to-user or search service has links with the UK, such as a significant number of UK users or the UK as a target market4.

  • Assess the risk of illegal content and comply with the illegal content duties, in force since 17 March 2025 under Ofcom's codes5.
  • Where the service is likely to be accessed by children, complete a children's risk assessment and meet the protection duties in force from 25 July 20255.

A moderation pipeline whose records double as compliance evidence

appealthreshold tuning01Intake02Classifier scoring03Routing04Decision record05Statement of reasons06Complaint review
  1. Intake

    User notices, trusted flagger notices, authority orders and proactive detections arrive with source tags.

  2. Classifier scoring

    Models score each item against a named policy version and confidence threshold.

  3. Routing

    High-confidence items are actioned automatically where policy allows; the rest go to trained reviewers.

  4. Decision record

    One record holds the action, ground, policy version, reviewer or model, and whether automation detected or decided.

  5. Statement of reasons

    Generated from the decision record and sent to the affected user, then submitted to the database.

  6. Complaint review

    Appeals reopen the record for a human decision; reversals feed threshold tuning and report metrics.

Conceptual design of an AI-assisted moderation pipeline. It shows how records flow, not a deployed system or measured performance.

Turning DSA duties into engineering work

  1. Version every policy and threshold

    Give each community guideline and classifier threshold an identifier and an effective date, so every decision record can cite the exact clause in force at the time.

    Output
    Policy and threshold registry
    Owner
    Trust and safety policy lead
  2. Tag the source and the automation level

    Record whether the item came from a user notice, a trusted flagger, an authority order or proactive detection, and whether a model detected it, decided it, or both. Article 17 and the transparency templates ask for that distinction.

    Output
    Decision schema with automation flags
    Owner
    Moderation engineering
  3. Set automation boundaries per harm type

    Allow automatic action only where precision on a recently labelled, held-out set meets the bar your policy team sets for that harm. Satire and hate speech need more human judgment than known counterfeit listings.

    Output
    Automation boundary table
    Owner
    Policy lead with data science
  4. Generate statements of reasons from the record

    Build the statement from structured fields: restriction and its scope, facts relied on, ground, automation used and redress routes. Free-text reviewer notes should add detail, not replace the fields.

    Output
    Statement templates per action type
    Owner
    Product and legal
  5. Submit to the database and strip personal data

    Online platforms send each decision to the DSA Transparency Database3. Pseudonymize or drop personal data before submission and keep the mapping internal for complaints.

    Output
    Database submission job
    Owner
    Moderation engineering
  6. Staff the complaint queue with qualified reviewers

    Complaints must be decided under the supervision of qualified staff, not solely by automated means1. Track reversal rates by policy and model version, since a high reversal rate is evidence a threshold is wrong.

    Output
    Appeals workflow and reversal dashboard
    Owner
    Trust and safety operations
  7. Build the transparency report as a query

    Map each field in the harmonized templates to a column in your decision store, including precision and recall estimates for each automated tool, and run a draft report well before the deadline.

    Output
    Report field map
    Owner
    Compliance analytics

A hypothetical marketplace handling counterfeit listings

Where AI-assisted moderation programs fall short of the DSA

Automation not recorded per decision

Early signalThe team can say a model is used but not which decisions it made.

MitigationMake detection and decision automation flags mandatory fields in the decision schema.

Accuracy measured once, at launch

Early signalPrecision and recall figures in reports come from the original validation set.

MitigationRe-label a fresh sample each reporting period and report against that.

Appeals handled by the same model

Early signalComplaint outcomes match the original decision almost every time.

MitigationRoute complaints to qualified reviewers who can see, but are not bound by, the model score.

UK and EU handled as one regime

Early signalOne risk assessment document is reused for both without mapping to Ofcom's codes.

MitigationKeep a shared evidence base but produce separate assessments for each act.

Questions and answers

Does every content removal under the DSA need a statement of reasons?

Hosting providers must give one for restrictions imposed because content is illegal or incompatible with their terms, including visibility limits, demonetization, suspensions and account closures. Article 17 applies only where the provider knows the user's electronic contact details, does not apply to deceptive high-volume commercial content, and does not cover action taken under an authority order issued under Article 91.

How should a platform report the accuracy of automated moderation?

The harmonized templates in Implementing Regulation (EU) 2024/2835 ask for accuracy indicators for automated tools, including precision and recall, together with a qualitative description of the tools and their possible error rate2. The defensible way to produce them is to label a fresh random sample of automated decisions each period and calculate the indicators per tool and harm type.

Does the DSA apply to companies outside the EU?

Yes, if they offer services to recipients in the EU. The regulation applies irrespective of where the provider is established, and a provider without an EU establishment must appoint a legal representative in a Member State1. Mere technical accessibility of a website from the EU is not enough on its own; a substantial connection, such as a significant number of EU users or targeting EU countries, is.

Can AI make moderation decisions without any human involvement?

For initial decisions, the DSA permits automated means but requires the provider to disclose their use in notice handling and in each statement of reasons. Complaints are different: the internal complaint system must decide under the supervision of qualified staff and not solely by automated means1. Most teams therefore automate high-precision, low-ambiguity cases and keep humans on contested and context-heavy ones.

Sources

  1. Regulation (EU) 2022/2065 on a Single Market For Digital Services (Digital Services Act) — EUR-Lex · checked 10 October 2026
  2. Commission Implementing Regulation (EU) 2024/2835 laying down templates concerning the transparency reporting obligations under Regulation (EU) 2022/2065 — EUR-Lex · checked 10 October 2026
  3. DSA Transparency Database — European Commission · checked 10 October 2026
  4. Online Safety Act 2023 — legislation.gov.uk · checked 10 October 2026
  5. Important dates for Online Safety compliance — Ofcom · checked 10 October 2026

More in Technology, Media & Telecommunications

Back to Technology, Media & Telecommunications

Next step

Send us your moderation decision schema for a DSA evidence review

Share the fields your moderation tools record today and your service tier. We will map them against the statement of reasons, database and reporting requirements and list the gaps to close.

Request a schema review