Domain Atlas / Content moderation & editorial AI
X Multilingual Hate-Speech Enforcement
Under Digital Services Act Articles 15(1)(e) and 42(2) — which require indicators of the accuracy and possible rate of error of automated content moderation, broken down by each official Member State language — X published, for the period 1 October 2024 to 31 March 2025, appeal and overturn rates for automated hateful-conduct visibility filtering across twenty named EU languages: appeal rates from 0.0 to 14.0 percent and overturn rates from 0.0 percent (Hungarian) to 72.2 percent (Swedish), with English at 40.4 percent. The independent audit of the sector's harmonised filings — a different set of reports from the April 2025 table above — found that of the eight largest EU platforms, four reported language-wise classification metrics (Facebook and Instagram across all 24 official languages, LinkedIn 22, X 7), three substituted country, and one reported none. Those are self-reported classification metrics; this atlas, not the audit, reads the April 2025 table as the only language-indexed error signal of its kind, because it indexes an outcome of the correction channel — appeals and overturns — rather than a platform's own accuracy figure. The figures are the platform's own legally mandated self-report, filed under Commission scrutiny, and no independent measurement of the deployment's per-language enforcement accuracy exists in the record.[3]
What happened
The Digital Services Act, in force since October 2022, did something no market pressure had done: it made a very large online platform's content-moderation error rate a legally mandated, language-indexed public fact. Article 15(1)(e) requires providers to publish indicators of the accuracy and the possible rate of error of their automated moderation means; Article 42(2) requires very large online platforms to break those indicators — and their moderation headcount and its linguistic expertise — down by each official Member State language. X (formerly Twitter) was designated a very large online platform in April 2023, and its reporting entity, X Internet Unlimited Company (formerly Twitter International Unlimited Company), has filed the mandated transparency reports since. The April 2025 filing, covering 1 October 2024 to 31 March 2025, is the load-bearing document: for automated hateful-conduct visibility filtering it published appeal rates across twenty named EU languages running from 0.0 percent (Bulgarian, Greek, Lithuanian) to 14.0 percent (Swedish), and overturn rates on those appeals running from 0.0 percent (Hungarian) to 72.2 percent (Swedish), with English — by far the best-staffed language — at 40.4 percent. The three languages with zero appeals carry no overturn rate at all: not a low error rate, no error measurement. Four official languages (Croatian, Irish, Maltese, Slovak) do not appear in the table.
Behind the table sits the workforce the same filings disclose. As of March 2025, X named 1,352 content moderators with primary-language professional proficiency, covering 8 of the 24 official languages — 1,197 of them (88.5 percent) in English, with seven other languages staffed by between 1 and 55 people each; a secondary list of 185 people, stated to be not distinct from the first, adds three more languages. Thirteen official languages appear in neither list. For those, the company's own words: "In situations where we need additional language support, we use translation services and/or machine translation tools, to investigate and address challenges in additional languages." Flagged content is either reviewed by a person before action or auto-actioned on the model's historical accuracy. So the deployment splits into two channels that differ in exactly one step — whether anyone who reads the language checks the machine — and the correction channel that produces the only error measurement is itself language-indexed, with several languages recording no throughput at all. What the published numbers do not show matters as much: the overturn rates do not sort by staffing (Swedish, with no named reviewer, sits highest at 72.2 percent; Hungarian, also with no named reviewer, sits at 0.0; English, with 1,197 reviewers, sits mid-table), the overturn rate is conditional on appeal with a language-dependent denominator, and the same metric family exceeds 100 percent for other policies in the same report — a period ratio, never a population error rate. Every one of these magnitudes is the platform's own legally mandated self-report, filed under Commission scrutiny; civil-society and academic reviews of the sector's filings verify the reporting ecosystem, not the enforcement's ground truth.
Then the sensor dimmed. In the very next reporting period (April to June 2025), the same visibility-filtering accuracy indicators were published broken down by country rather than by language; the content-removal tables kept language columns, but their automated hateful-conduct rows are almost entirely blank or 0.00 percent. The first independent audit of the harmonised transparency reports (Trujillo, Tessa and Cresci, 2026 — an unrefereed preprint, weighed accordingly) found the pattern sector-wide: of the eight largest EU platforms, only four reported language-wise classification metrics at all, three substituted country, one reported none — and X's first harmonised filing covered 7 of the 24 official languages. The International Network Against Cyber Hate, reviewing the sector's latest reports, found the same reviewer-language asymmetry everywhere (Meta: 1 named Maltese reviewer against 3,110 Spanish) and concluded there is "no clarity on moderation outcomes per country or per language" — X's April 2025 table having been the closest thing to an exception. The predecessor report covering April to September 2024 rotted off X's own transparency site between two reads in August 2026, and this case file rests no figure on it. Enforcement around the deployment is active but has not touched this question: the Commission opened formal DSA proceedings against X in December 2023 — content moderation and the dissemination of illegal content among the grounds — which remain open with no published findings, and were extended in January 2026 to recommender systems plus a new investigation into the platform's integrated generative AI assistant. On 5 December 2025 the Commission fined X EUR 120 million, the first non-compliance decision under the DSA — for deceptive verified-checkmark design, an inadequate advertising repository, and failure to give researchers effective data access. That fine concerns transparency obligations; it is not a ruling on moderation accuracy, hate-speech enforcement, or the language tables, and no regulator has ruled on this deployment's per-language enforcement accuracy in the record read here.
The sociotechnical reading
Most deployments in this atlas fail to measure something. This one is in the atlas because a law forced a measurement into existence — and the record then documents, in the operator's own filings, what the measurement was, where it could not reach, and how it decayed. The structure worth studying is the sensor, not the classifier (a model that sorts each item into a category). The only error signal this deployment publishes is an overturn rate on appeal, which means the error measurement exists exactly where the appeal channel has throughput. The filings show that throughput is indexed by the served person's language: 0.0 to 14.0 percent by language, zero in three languages — so for those cohorts the mandated error indicator returns nothing, and an absent measurement is structurally different from a clean one. The same filings disclose why the two enforcement channels differ: 8 of 24 languages have a named fluent reviewer behind the machine, thirteen have a translation step instead, and the label pool that trains tomorrow's classifiers is generated by a reviewer population whose named capacity is 88.5 percent one language. The map draws all three facts as structure — two review channels differing in one documented step, an appeal edge whose rung is the published table, a label-pool loop no audit examines — and refuses the one inference the numbers invite: nothing in this record supports a monotone relation between staffing and measured error, and the map asserts none.
The governance arc is a disclosure loop that demonstrably ran once and then lost resolution. The regulator's authority here is over what must be measured and published, never over an individual decision — and it produced, for one reporting period, the only language-indexed accuracy table in the sector. The next period's table was broken down by country. The first harmonised filing covered 7 of 24 languages. The sector audit found three of eight platforms substituting country for language and one reporting nothing. That is the case's distinct lesson for the Field Guide: a mandated sensor is still a sensor somebody operates, and its resolution — which unit the error is indexed by — turned out to be the governable surface. A reader should also hold the enforcement posture precisely: the Commission has demonstrated both appetite and teeth against this operator (a EUR 120 million fine in December 2025 on transparency obligations — the checkmark design, the ad repository, researcher access), and the moderation limb of the December 2023 proceedings remains open with no findings; but no enforcement action in this record responds to the language-to-country regression itself. The honest boundaries travel with the case. Every per-language magnitude is vendor-tier: the platform's own mandated self-report, usable because it is filed under legal obligation and fining power, never because it has been independently verified — the independent documents in this record audit the filings, not the enforcement. The overturn rate is a period ratio conditional on appeal, not a probability. And the people whose posts are moderated are boundary-only throughout: the published per-language rates are recorded as external observations with their denominators, no protected characteristic appears anywhere in the record, and nothing about any served person is computed from the network this case pairs with.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.