Platforms Filed Nearly Four Billion Moderation Notices in Six Months. Researchers Found Most Are Templates.

What the database is

When an online platform in the European Union restricts a piece of content — removes it, demotes it, disables an account, blocks a listing — the Digital Services Act requires it to produce a statement of reasons explaining what it did and why. Article 24(5) requires that statement to be submitted to a database the European Commission maintains and publishes.

The result is a public, machine-readable record of content moderation at a scale nothing else approaches. It is searchable, downloadable, and updated continuously, and it exists specifically so that decisions made privately by companies can be scrutinised publicly by anyone who wants to.

That is an unusual regulatory design, and the unusual part is the granularity. Most transparency obligations produce a periodic report written by the regulated party, summarising its own conduct at whatever level of detail it chooses. This one produces a row per decision, filed by the platform, in a common schema, in something close to real time.

The Commission's own framing is that the database enables scrutiny of moderation decisions and monitoring of how illegal and harmful content spreads online. The theory is that aggregate visibility disciplines individual behaviour.

Whether it achieves that is a separate question from whether it is large, and the two get conflated constantly in coverage of the database. It is enormously large. What the volume actually tells you about moderation is considerably less obvious than the headline figure first suggests, and the composition is where the meaning sits.

The number, and what it actually counts

When we retrieved the database it reported 3,944,217,973 statements of reasons for the preceding 180 days, from 373 platforms, with 43 per cent recorded as fully automated decisions.

Just under four billion moderation decisions in six months works out at roughly twenty-two million a day, or around two hundred and fifty every second, filed by companies about their own enforcement. No regulator reads that volume; the premise of the design is that researchers, journalists and civil society will read it in aggregate instead.

Those figures describe a rolling window rather than a cumulative total, which matters when comparing across sources. Secondary analyses citing 3.25 billion as of May 2026, or 3.52 billion between September 2023 and May 2025, are measuring different spans and are not directly comparable to the 180-day figure.

A statement of reasons is also not the same as a piece of content. One decision may cover a single listing, and platforms differ in how they batch and count, so the number is a count of filings rather than a census of removals.

The headline is nonetheless robust in one respect. Whatever the exact denominator, and whatever the mix of platforms behind it, automated enforcement operating at population scale is now documented in a public record rather than inferred from company blog posts or leaked internal figures. That is genuinely new.

Who actually files

The most useful published analysis of the database is an academic paper, Automated Transparency: A Legal and Empirical Analysis of the Digital Services Act Transparency Database, which examined more than 131 million statements submitted during a ten-day window in November 2023.

Its headline finding is a surprise to anyone who assumes this is a social media dataset. Google Shopping accounted for 52.2 per cent of all statements in the sample, TikTok 17.1 per cent and Amazon 10.8 per cent.

The authors note explicitly that submission “is dominated by Google Shopping, TikTok and Amazon, whereas these are not the largest VLOPs” — the very large online platforms that the regulation's heaviest obligations target.

Commerce platforms delist products at industrial volume for reasons that are mostly mundane: counterfeit goods, prohibited items, listing errors, policy mismatches, expired offers. Each one generates a filing identical in form to a decision about political speech, and the schema does not distinguish between them in any way a casual reader would notice.

This is the single most important thing to understand before quoting the total. A four-billion figure that is majority product listings does not describe four billion contested speech decisions, and reporting that treats it as such is measuring the wrong thing.

The automation share

The database's own dashboard reports 43 per cent of decisions as fully automated. That is a striking figure in aggregate and a misleading one if read as an average across comparable actors, because the platforms filing the most are not a representative sample of the platforms operating in Europe.

The academic analysis found automation concentrated rather than distributed. AliExpress, Google Maps, Google Shopping and TikTok made fully automated decisions for most or all of their statements. X, by contrast, relied on human review.

So the aggregate automation rate is substantially a fact about which platforms file the most, not about how moderation is done across the sector. Change the mix of filers and the headline percentage moves without any platform changing its practice.

That matters for anyone drawing conclusions about machine decision-making from the number. The database is excellent evidence that automated enforcement happens at scale, and weak evidence about how any individual platform behaves relative to another.

It also creates an odd incentive structure worth naming. A platform that files diligently and automates heavily looks worse in the aggregate statistics than one that files less or classifies its automation more conservatively, which is the opposite of what a transparency regime should reward.

What researchers found when they read them

The most damaging finding in the academic work is not about volume or automation at all. It is about whether the statements say anything, which is the only question that determines whether the exercise achieves what it was designed to achieve.

The authors report that most platforms submit “generic templates that may not provide very meaningful information” rather than explanations specific to the decision taken. The field is populated; the field is not informative.

That is the familiar failure mode of any disclosure regime that specifies a form rather than a standard of explanation. Compliance is measured by whether a statement was filed, so filing a statement becomes the objective.

It is the same pattern we documented in a different context in our report on paid-content labelling: a label that technically exists but conveys nothing satisfies the letter of a rule while defeating its purpose.

None of which makes the database worthless. It is the only dataset of its kind, it is public, and the criticism itself was only possible because the data is open enough to be analysed and found wanting.

Why this matters outside Europe

For publishers and companies distributing content, the database is a practical research asset rather than a regulatory burden. It records, per platform, what categories of material get restricted and how often, which is ordinarily invisible.

It is also a precedent, and precedents in this area travel. A rolling, row-level, machine-readable disclosure obligation is a far more demanding transparency standard than the periodic PDF that most jurisdictions currently require, and regulators outside the European Union are watching how well it actually performs before copying it.

The performance lesson so far is that mandating the filing is easy and mandating the meaning is hard. Four billion rows of template text is compliance; it is not necessarily transparency.

For anyone whose content has been restricted on a major platform, the database is the first place to look for the stated ground — with the caveat that the stated ground may be boilerplate.

And for anyone quoting the number in an argument, the honest framing includes the denominator, the rolling window it covers, and the fact that a majority of the best-studied sample was commerce rather than speech. A figure this large is persuasive precisely because nobody checks its composition.

Methodology and limitations

The headline figures — 3,944,217,973 statements of reasons over the preceding 180 days, 373 reporting platforms and 43 per cent fully automated decisions — were read directly from the European Commission's DSA Transparency Database on 26 September 2026 and are quoted as displayed. The database reports a rolling window, so these figures will differ from any cumulative total and from figures retrieved on another date. Platform-level shares, the automation breakdown by platform and the criticism of template submissions are from Kaushal, van de Kerkhof, Goanta, Spanakis and Iamnitchi, Automated Transparency: A Legal and Empirical Analysis of the Digital Services Act Transparency Database, dated 3 April 2024, which analysed a ten-day sample of 131 million statements in November 2023. That sample is early and its distribution may not reflect current filing. We attempted to retrieve the database's own statistics endpoint for a current platform breakdown and it returned HTTP 404. We have not independently analysed the underlying data.

Conclusion

The database is the largest public record of content moderation that has ever existed, and its headline number is genuinely arresting: nearly four billion filings in half a year, 43 per cent of them made without a human. But the composition matters more than the total. A majority of submissions in the best-studied sample came from commerce platforms delisting products, automation is concentrated among a handful of heavy filers rather than typical across the sector, and the academic verdict on the explanations themselves is that most are templates. Volume was the easy part of transparency. Meaning is the part still outstanding.

Frequently Asked Questions

Sources

1. official European Commission, DSA Transparency Database, retrieved 26 September 2026. Source of the 180-day totals, the platform count and the automation share, quoted as displayed.
2. academic Rishabh Kaushal, Jacob van de Kerkhof, Catalina Goanta, Gerasimos Spanakis and Adriana Iamnitchi, “Automated Transparency: A Legal and Empirical Analysis of the Digital Services Act Transparency Database”, 3 April 2024.
3. official Regulation (EU) 2022/2065 (Digital Services Act), Article 24(5), for the obligation on providers of online platforms to submit statements of reasons to the Commission's database.