TikTok's EU and UK content-moderation operation

What this models

This example runs on the Supervised-platform moderation apparatus network: 12 components and 26 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 1 assumed · 11 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• baseline

  D48-derived new org (Phase 6, content-moderation-editorial). REGISTER FIRST, because it governs every value here and one wrong sentence would invert the case. This is a REGULATED-PLATFORM board and its structure is a GAP, not a prosecution. Two formal proceedings are open against this operator and neither names content-moderation staffing as a ground: the first, of 19 February 2024, covers protection of minors, advertising transparency, researcher data access and the risk management of addictive design and harmful content; the second, of 17 December 2024, covers election-risk recommender systems and political advertising. Four sets of preliminary findings have issued and preliminary findings are not findings of breach; the operator publicly contests at least one set; two limbs closed by binding commitments, which is not a finding of infringement either; there is no non-compliance decision and no fine. The link between the staffing substitution and the supervision is argued by unions, parliamentarians and civil society under the diligence heading, and the authority itself has written that the statute prescribes no rules on moderation resources. Any copy anywhere that says this operator is under investigation for reducing its moderators is wrong.

• baseline

  TOPOLOGY, drawn at the coarsest granularity at which every documented mechanism of this deployment is still distinguishable. Twelve nodes, all documented, none decorative. ONE model, because the record describes one pre-publication automated stage standing in front of every upload and identifies no model anywhere; the detection families it names are technologies, not systems, and none is reproduced here. THREE operator classes because the record documents three groups with three different jobs and three different fates: the desks that work what the gate routes to them, the internal complaint-handling channel the statute requires to be staffed by appropriately qualified people, and the annotation and training teams that maintain the judgement rather than exercise it and that the operator placed inside the same reduction. THREE record stores because the sources measure three separate things and the wiring between them is the case: the working decision and contest record, the compelled transparency filing, and the training label and detection-rule store. The statutory notice and order population is part of the working record here, as it is in the PAN entry for this deployment: a feed the operator does not control, with published volumes and published clocks, whose completeness the operator's own auditor could not establish. The human queue the gate fills is carried on the routing pathway into the desks rather than drawn as a separate element, because the record publishes its latencies and none of its depth and describes no prioritisation inside it. ONE external boundary because the record crosses to outside readers through the published reports and the per-decision public database, and the audit examined that crossing. And FOUR reviewers, which is derived rather than chosen — see the next entry.

• baseline

  WHY FOUR REVIEWERS, when no shipped org has more than three. Because the evidence record for this deployment describes four separately constituted examining channels with four different accesses, four different instruments and four different documented effects, and collapsing any pair would hide the mismatch that is the finding. The STATUTORY SUPERVISION CHANNEL holds the largest instrument in the record — proceedings, interim measures, binding commitments, non-compliance decisions with fines up to 6 percent of worldwide turnover — and has stated in writing that it holds no rule about the resources dedicated to content moderation. The MANDATED ANNUAL AUDIT CHANNEL holds an opinion and no remedy power, and its opinion for the year to 30 June 2025 is negative and qualified, including on the completeness of the notice and complaint populations from which the operator's own accuracy sample is drawn. The OUT-OF-COURT DISPUTE SETTLEMENT BODIES can decide and cannot bind, and the difference is measured: they disagreed with the operator in 106 cases in six months and the operator implemented 29. The LABOUR AND PARLIAMENTARY SCRUTINY CHANNEL holds no authority over the system and real authority over its labour input, and in this record it is the channel that actually produced disclosure. Two of the four have a counterpart among the PAN entry's oversight users; the supervision channel is carried in PAN as its regulator governance actor and the dispute bodies inside the enforcement store's own source line, and the Lab draws each because its access, its instrument and its documented effect all differ from the others'.

• baseline

  ABSENCES ARE DERIVED TOO, and four of them are load-bearing. There is NO enforcement node. The strikes policy that escalates to account bans is real and the volume is published — 4,906,735 account bans or suspensions in six months, 871,819 of them automatic — but the record describes it as a policy applied within the same decision record rather than as a downstream system with its own store, its own operators and its own reconciliation, and drawing one would require a transfer and a missing return leg that no source describes. The volume is carried on the write into the decision record instead. There is NO guardrail: no bounded automated screen over the gate's output before people act is documented anywhere, and what stands between the gate and a human decision is the clear-cut threshold, which is a property of the gate. There is NO retriever, although the closest candidate in this domain sits here — the repository of previously fact-checked claims that specialist misinformation moderators work against — because the record names it once, in a single clause of the published risk assessment, with no volume, no coverage figure and no description of how it is queried; it is carried in the review desks' copy and on the routing pathway into them instead. And there is NO node for any person who uploaded, reported or appealed: served people are never in these dynamics, and here the finding is that the parties who could measure what the substitution did to them are the researchers the authority preliminarily found are left with partial or unreliable data.

• baseline

  WHERE THE LAB SHAPE DIVERGES FROM THE PAN SHAPE, and nothing is asserted here that the PAN file does not already record. Three divergences. First, PAN carries five users and the Lab draws three operator classes and four reviewers: the PAN oversight user that covers mandated disclosure and independent audit together becomes the audit channel, the PAN labour-and-parliamentary oversight user becomes the scrutiny channel, and two further reviewers are drawn from material PAN carries elsewhere — the supervisor from PAN's regulator governance actor, whose label states the proceedings, the powers and the written no-resourcing-rule answer in full, and the dispute bodies from the Article 21 figures inside PAN's own enforcement store source line. Second, PAN draws no external boundary; the Lab draws one, because the record documents the crossing of the published reports and the per-decision public database to outside readers and because the independent auditor examined that crossing and qualified it. Third, PAN has no edge kind for a check: two of this board's five check pathways REDRAW PAN edges, and their widths still come from PAN on the stated mapping, while the other three are derived from the cited record directly. Where PAN folds the notice population into its enforcement store, this network now folds it the same way.

• baseline

  BASELINES, and exactly how far the PAN org carries them. The PAN entry for this deployment holds twenty-eight edges. Sixteen of this network's twenty-six pathways have a one-to-one counterpart among them, and every one of those sixteen mirrors that edge's width on one fixed four-rung mapping, stated in the derivation comment beside this network and applied with no exceptions. The other twelve PAN edges are carried on those survivors, each of which keeps its own counterpart's width: no survivor was widened or narrowed for what it absorbed. The remaining ten pathways are derived from the cited record directly. Four contrasts carry the case. The gate writes the decision record at the top rung while the moderation desks write it at the middle one, which is the substitution in two numbers. The annotation teams write the label store at the top rung and nothing on this network carries that store to any outside reader or to any of the four examiners, which is the one part of this apparatus nobody may check. The supervision channel shapes what is filed at the middle rung and reaches the size of the review layer at zero, which is the gap the whole board is about. And the scrutiny channel, which has no authority over the system at all, reaches the desks at the middle rung, wider than any check a statutory examiner holds here: the supervisor reaches the complaint desks at the low rung, the dispute bodies reach them at the low rung, and the auditor, which holds no remedy power, reaches no desk at all. EVERY PAN edge for this deployment is marked estimated, and every strength here is a modelling choice within the rungs those widths support.

• baseline

  DEMAND 3 / CAPACITY 1, and what the capacity value does NOT rest on. Demand 3 on the operator's own mandated and dated figures for January to June 2025: 24,534,707 Community Guidelines removals, 169,527,678 content restrictions, 2,781,470 service restrictions, 4,906,735 account bans or suspensions, about 22,700 appeals a day across 4,130,190 appeals, 308,755 illegal-content reports and 3,976 authority orders, against an audience of 169 million monthly active recipients in the European Union up from the 135.9 million declared at designation. Capacity 1 on the mandated headcount disclosure moving the other way: more than 6,000 at end-December 2023 and 4,596 at end-June 2025 including 247 not language-assigned, with five named language teams in single digits or at zero, 91 employed staff against 3,583 contracted in the following report, and whole sites closed or proposed for closure at Amsterdam, Dublin, Berlin and London across the same span. NO throughput, queue-depth, backlog or per-reviewer workload figure has ever been published for this deployment by the operator. The only per-head referents anywhere in the record are 800 to 1,000 videos reviewed a day and a typical tenure of three to four years, both from worker accounts relayed by a partisan outlet, both attributed rather than asserted throughout this bundle, and no value on this diagram is scaled by either.

• baseline

  EVIDENCE STATUS, labelled where it is used, because this record mixes four tiers that must never be flattened into one. REGULATOR-PRIMARY AND ASSERTABLE AS FACT: every proceeding date, ground and article; the two commitment closures; the written answer that the statute prescribes no rules about moderation resources; and the parliamentary record of the closure of a 300-person national moderation team. OPERATOR SELF-REPORT UNDER A LEGAL MANDATE, quotable and never launderable into an independent measurement: the headcount series and the per-language table, the enforcement and appeal volumes, the accuracy and error indicators, the three automation percentages, the composition of the London reduction and the stated reason for it. PRELIMINARY REGULATORY VIEW, carrying the word preliminary every single time: the advertisement repository, researcher data access, addictive design and minors' account settings. ALLEGATION IN AN UNADJUDICATED DISPUTE, carried with the operator's denial attached: the ballot-timing and union-busting claims, the unfair-dismissal claims, the German dismissal cases, and every union prediction that the reduction will degrade moderation quality. AND NOT ASSERTABLE AT ALL, anywhere in this bundle: that the substitution caused any measured decline in moderation quality, that the supervisor is investigating the staffing cuts, that the three automation percentages form a trend, and the operator's own claim to have the lowest error rates among all major platforms.

• baseline

  THE THREE AUTOMATION PERCENTAGES ARE NOT A TIME SERIES and this network never draws one. 72.3 percent is the automatic share of Community Guidelines removals in the mandated report for January to June 2025. 86 percent is what the operator told a parliamentary committee in October 2025, citing an April-to-June 2025 report from a different report family. 93.8 percent is the share of all violating content actioned without human review in the report for July to December 2025, the first period to include comment enforcement. They measure different things over different denominators, and the operator itself warns that the later report captures a broader range of automated enforcement actions than the previous ones. Wherever a percentage appears on this board it appears with its scope and its period attached, and no arrow is drawn between any two of them. The same discipline governs the headcount table: the January-to-June 2025 Annex D was restated on 15 April 2026, the previous half's per-language figures reach this file through a civil-society transcription rather than the original document, and the operator groups some language teams and folds some non-European-Union languages into the totals — so the direction of travel and the shape of the distribution are used, and never the differences.

• baseline

  WHAT THE RECORD DOES NOT SAY, recorded as unknowns and never as zeros on any measured quantity. No published independent measurement of a moderation-quality effect from the staffing substitution exists. The available quality signals are exactly three, and all three are weak in different ways: the operator's own upheld-and-overturned statistic, which is conditional on somebody appealing and computed by the party that made the original decision; an independent auditor who could not confirm the completeness of the notice and complaint populations that statistic is drawn from; and a preliminary regulatory view that the researchers who could measure it independently are left with partial or unreliable data. No per-language error rate is published. No appeal rate per language is published. No throughput or backlog figure is published. No measure of the training label store's quality, coverage or per-language balance is published. And the published risk assessment of 28 August 2025 contains no reference to the workforce reduction announced eighteen days earlier — which is an observation about what a published document contains, not proof that the matter was never assessed internally, because that document is the public version of a confidential one. Where this network draws a zero it is drawing a pathway the sources document as absent, refused or unreachable, never a measured quantity that came out at zero.

• assumed

  THE PEOPLE ON THE OTHER SIDE OF THIS APPARATUS ARE NOT MODELLED, and the reason is the same one that makes this case interesting. No removal, restriction, reinstatement, ban or outcome for any person is computed from anything on this diagram, and no individual user, uploader, moderator or official is named or characterised. The equity observations the evidence record carries — that the smallest named language teams stood at five, ten, ten, eleven and nineteen reviewers with two official languages at zero, and that 1,359,823 reinstatements arrived against 4,130,190 appeals in one six-month period — are recorded external observations about mandated disclosures, never error rates and never measures of moderation quality in any language, and the independent auditor found the controls behind the first of them not sufficient and appropriate. The one thing the record establishes about the served side is an absence: the parties positioned to measure what any of this did to the people using the platform are the researchers whom the supervisor preliminarily found are left with partial or unreliable data, and that is stated here as an absence rather than drawn as a pathway.

• baseline

  SEVERAL DOCUMENTED FLOWS RIDE ON A NEIGHBOUR rather than on a line of their own, because each is a finer split than the case's failure, its containment or its instruments distinguish. The gate's decisions and the desks' decisions reach the complaint channel as the appealed entries of the working record. The auditor sees the automated stage only through the records it produces. The unions and the committee see the automated stage only through the operator's reports and letters, and the audit opinion reaches them published beside the filing. The per-decision statements of reasons are part of the compelled filing, so the supervisor's read of them and their crossing to the public database travel with the filing. The desks' decisions become training examples through the record they write, and enforcement history reaches the gate through the label store. The complaint channel's reading of its own published grade and the annotation teams' reading of the labels they extend are the return legs of writes already drawn. The scrutiny channel's pressure on the annotation and training teams is the same pressure it puts on the desks, recorded as thinner. Every one of these flows is documented; none is lost, and each is stated on the pathway or the element that now carries it.

What this example does not show

• STATUS, verbatim from the evidence dossier: ONGOING ON EVERY AXIS AND CONCLUDED ON NONE. Regulatory: two formal DSA proceedings open (19 Feb 2024 and 17 Dec 2024); four sets of preliminary findings issued, none confirmed; one limb closed by binding commitments (advertising transparency, 5 Dec 2025); rabbit-hole and age-assurance grounds and the whole election case still under investigation; no non-compliance decision, no fine. Deployment: the substitution is in progress and accelerating - Netherlands complete (2024), Berlin and London announced 2025 with disputes live, Dublin proposed again July 2026; automation share as the operator reports it has moved from 72.3 percent of Community Guidelines removals in H1 2025 to 93.8 percent of violating content actioned without human review in H2 2025. Labour: German dismissal cases before the Berlin Labour Court and a UK pre-action letter, neither adjudicated. Measurement: an independent DSA audit opinion that is negative and qualified, a restated moderator table, and a preliminary finding that researchers cannot get reliable data.
• NEITHER PROCEEDING IS ABOUT MODERATION STAFFING, and this scenario never says otherwise. The February 2024 grounds are protection of minors, advertising transparency, researcher data access, and the risk management of addictive design and rabbit-hole effects and age assurance; the December 2024 grounds are election-risk recommender systems and political advertising. The link between the staffing substitution and the proceedings is argued by unions, parliamentarians and civil society under the systemic-risk diligence heading; it is not a ground, and the Commission has written that the Digital Services Act does not prescribe any specific rules about the resources to be dedicated to content moderation. Any copy that says TikTok is under investigation for cutting moderators is wrong.
• PRELIMINARY FINDINGS ARE NOT FINDINGS OF BREACH. All four sets against this operator — the advertisement repository in May 2025, researcher data access in October 2025, addictive design in February 2026, and minors' account settings in July 2026 — expressly do not prejudge the outcome, and TikTok publicly rejected the addictive-design findings as a categorically false and entirely meritless depiction of its platform and said it will take whatever steps are necessary to challenge them. As of 28 August 2026 there is no final non-compliance decision and no fine against TikTok under the Digital Services Act. The two matters concluded are closures by binding commitments — the TikTok Lite Rewards programme in August 2024 and advertising transparency in December 2025 — which are not findings of infringement either.
• EVERY QUANTITATIVE FIGURE ABOUT THE DEPLOYMENT IS THE OPERATOR'S OWN, published because the law compels it. The headcount series, the per-language table, the enforcement and appeal volumes, the accuracy and error indicators and the automation percentages are all TikTok's, and the independent auditor found the internal controls over data accuracy and completeness monitoring between the source systems and the transparency report not sufficient and appropriate. This scenario attributes each figure to its report and its period and treats none of them as an audited measurement.
• ACCURACY 99.2 PERCENT MEANS OVERTURNED 0.8 PERCENT OF THE TIME. TikTok defines accuracy as the proportion of decisions upheld or maintained and error as the proportion overturned. That statistic is conditional on somebody appealing, is computed by the party that made the original decision, and is not published per language. In the same six months 3,075,758 uploader and advertiser appeals produced 1,359,823 reinstatements or lifted restrictions — a signal an order of magnitude larger on a different denominator, which TikTok itself warns does not map one to one onto the period's appeals. Neither number is used here as a population error rate.
• THE THREE AUTOMATION PERCENTAGES ARE NOT A TIME SERIES. 72.3 percent (Community Guidelines removals, the January-June 2025 mandated report), 86 percent (April-June 2025, cited by TikTok to a parliamentary committee from a different report family) and 93.8 percent (all violating content actioned without human review, July-December 2025, the first period including comment enforcement) measure different things over different denominators, and TikTok itself flags the widening scope. They appear here as three separately sourced statements and never as a trend line.
• HEADCOUNT FIGURES FROM DIFFERENT REPORT EDITIONS ARE NOT SAFELY COMPARABLE. The January-June 2025 Annex D was restated on 15 April 2026; the previous half's per-language figures reach this file through a civil-society transcription rather than the original document; TikTok groups some language teams and includes several non-EU languages in the totals. The two directly quotable anchors are more than 6,000 at end-December 2023 and 4,596 at end-June 2025, and the widely-cited 26 percent reduction since September 2023 is an analyst's arithmetic over the operator's own disclosures, not a TikTok statement and not an audited figure. This scenario uses the direction of travel and the shape of the distribution, never the differences.
• NO INDEPENDENT MEASUREMENT OF THE OUTCOME EXISTS, and this board is built around that absence rather than around an implied degradation. The available quality signals are three: the operator's own upheld-and-overturned statistic; an independent auditor who could not confirm the completeness of the notice and complaint populations that statistic is drawn from; and a Commission preliminary finding that researchers asking for public platform data are often left with partial or unreliable data. No published study measures what the staffing substitution did to moderation quality, in any language, on any policy.
• UNION-BUSTING IS AN ALLEGATION IN A PRE-ACTION LETTER, and TikTok's denial travels with it everywhere in this bundle. Two London moderators, supported by Foxglove and a Communication Workers Union branch and represented by Leigh Day, sent a letter on 19 December 2025 alleging unlawful detriment and automatic unfair dismissal after redundancy notices landed days before a scheduled union-recognition ballot. TikTok told Parliament the union-timing claims are categorically untrue, that the decisions were made globally, and that it had written to the union expressing regret about the timescales while remaining open to re-engaging after consultation. No tribunal has ruled, and the German dismissal cases before the Berlin Labour Court have no reported outcome. Every union prediction that the reduction will degrade moderation quality is a contested prediction, not a finding.
• THE PER-HEAD FIGURES ARE WORKER TESTIMONY RELAYED BY A PARTISAN OUTLET. The 800-to-1,000 videos a day and the three-to-four-year tenure are the only per-head capacity referents anywhere in this record and TikTok publishes none of its own. They are attributed wherever they appear and no value on this board is scaled by them.
• THE UNITED KINGDOM MATERIAL IS LABELLED SEPARATELY AND CARRIES NO DSA FRAMING. The London site, the CWU and UTAW ballot, the pre-action letter, Ofcom and the House of Commons Science, Innovation and Technology Committee sit under the UK Online Safety Act 2023 and the earlier Video Sharing Platform regime, not under the Digital Services Act. Ofcom was given prior notice of the London proposal and its response is not on the public record. The two regimes have different obligations, different regulators and different remedies, and nothing here lets one carry the other's framing.
• PUBLISHED RISK-ASSESSMENT AND AUDIT DOCUMENTS ARE THE PUBLIC VERSIONS. The Year 3 systemic risk assessment is headed confidential, and the absence of any reference to the workforce reduction in it is an observation about the published text rather than proof that the matter was never assessed internally. This scenario states it as what the published record does and does not contain.
• NO AI MODEL IS IDENTIFIED ANYWHERE IN THE RECORD and none is introduced here. TikTok describes vision-based, audio-based, text-based and large-language-model-based technologies and nothing more specific. A union's characterisation of the replacement as models sourced and trained by the parent company is a union statement rather than a documented system description, and it appears nowhere in this bundle. The parent company's ownership and the national-security material that surrounds it belong to a different jurisdiction and a different case and are excluded.
• SERVED PEOPLE ARE NOT MODELLED. No removal, restriction, reinstatement, ban or outcome for any person is computed from anything on this diagram, and no individual user, uploader, moderator or official is named or characterised. The equity observations the record carries — five named language teams in single digits or at zero, and 1,359,823 reinstatements against 4,130,190 appeals in six months — are recorded external observations about mandated disclosures, never error rates and never measures of moderation quality in any language.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• The European Commission has two formal Digital Services Act proceedings open against TikTok and neither names content-moderation staffing as a ground. The first, opened 19 February 2024, covers protection of minors, advertising transparency, data access for researchers, and the risk management of addictive design and harmful content, naming suspected infringements of Articles 34(1), 34(2), 35(1), 28(1), 39(1) and 40(12), and records TikTok's designation as a very large online platform on 25 April 2023 at 135.9 million EU monthly active recipients; the release states that the DSA sets no legal deadline for bringing formal proceedings to an end. The second, opened 17 December 2024, concerns election-integrity systemic risk following the annulled Romanian presidential first round of 24 November 2024 and is limited to two grounds — recommender systems including coordinated inauthentic manipulation, and policies on political advertisements and paid-for political content — under Articles 34(1), 34(2) and 35(1), with Coimisiún na Meán, the Irish Digital Services Coordinator, associated to the case and a retention order of 5 December 2024 preceding it. A third and earlier proceeding, on the TikTok Lite Rewards programme, was opened on 22 April 2024 and closed on 5 August 2024 when the Commission made binding TikTok's commitment to withdraw the programme from the EU permanently and not to launch a circumventing programme: the first DSA case closed and the first commitments accepted. As of 28 August 2026 the February 2024 proceeding has produced four sets of preliminary findings — the advertisement repository on 15 May 2025, researcher data access on 24 October 2025, addictive design on 6 February 2026, and minors' account settings on 24 July 2026 — and one closure by binding commitments, on advertising transparency, on 5 December 2025, with the rabbit-hole effect of the recommender systems and the risk of age misrepresentation still under investigation; the December 2024 proceeding has produced no preliminary findings. Preliminary findings are not findings of breach and do not prejudge the outcome, and TikTok said of the addictive-design set that 'The Commission's preliminary findings present a categorically false and entirely meritless depiction of our platform, and we will take whatever steps are necessary to challenge these findings.' There is no non-compliance decision and no fine against TikTok under the Digital Services Act.

  empirical
  • Government European Commission (2024, February 19). Commission opens formal proceedings against TikTok under the Digital Services Act (IP/24/926) https://ec.europa.eu/commission/presscorner/api/files/document/print/en/ip_24_926/IP_24_926_EN.pdf
  • Government European Commission (2024, December 17). Commission opens formal proceedings against TikTok on election risks under the Digital Services Act (IP/24/6487) https://ec.europa.eu/commission/presscorner/api/files/document/print/en/ip_24_6487/IP_24_6487_EN.pdf
  • Government European Commission (2025-2026). Preliminary findings on TikTok's ad repository (IP/25/1223, 15 May 2025), on researcher data access (IP/25/2503, 24 October 2025), on addictive design (6 February 2026) and on minors' account settings (IP/26/1679, 24 July 2026); with the advertising-transparency commitments decision of 5 December 2025 and the TikTok Lite Rewards closure of 5 August 2024 (IP/24/4161). PRELIMINARY FINDINGS ARE NOT FINDINGS OF BREACH https://ec.europa.eu/commission/presscorner/api/files/document/print/en/ip_25_2503/IP_25_2503_EN.pdf

• The European Commission has stated in writing that the Digital Services Act gives it no rule about content-moderation resourcing. Asked in European Parliament written question E-002454/2024, submitted 6 November 2024 by Kim Van Sparrentak, whether a platform can comply with Articles 16, 20, and 34 to 35 after firing an entire national moderation team of 300 people in the Netherlands, the Commission answered, in a reply last updated 15 January 2025: 'The DSA does not prescribe any specific rules about the resources to be dedicated to content moderation.' It added that platforms must enforce their moderation rules 'in a diligent, objective and proportionate manner', that 'qualified staff' must ensure fair and unbiased decision-making in internal complaint handling, that 'it is important that designated companies put in place adequate content moderation processes and dedicate enough resources for diligent content moderation', and that it 'is closely monitoring TikTok's compliance with the DSA and will follow up with formal enforcement steps if appropriate'. No formal enforcement step naming moderation staffing has followed as of 28 August 2026. The question itself put the staffing change in Digital Services Act terms — Article 16 notice and action, Article 20 internal complaint handling, which requires reasoned decisions taken under the supervision of appropriately qualified staff and not solely on the basis of automated means, and Articles 34 and 35 on systemic risk taking into account specific regional or linguistic aspects — and cited 5.7 million Dutch monthly users. The boundary this establishes is the structure of the case: the regulator holds a diligence lever and not a headcount lever.

  empirical
  • Government European Parliament written question E-002454/2024 (Van Sparrentak) and the Commission's answer of 15 January 2025: 'The DSA does not prescribe any specific rules about the resources to be dedicated to content moderation' https://www.europarl.europa.eu/doceo/document/E-10-2024-002454-ASW_EN.html

• TikTok's EU content-moderation workforce fell across 2024 to 2026 while its EU audience grew, on its own legally mandated disclosures. Its second DSA transparency report recorded 'More than 6k moderators are dedicated to the moderation of content in the European Union as of the end of December 2023'. Its fifth report, for January to June 2025, gives 4,596 people dedicated to EU content moderation at the end of June 2025, of whom 247 are not language-specific, with an Annex D per-language table — as corrected on 15 April 2026 — of English 1,552, German 567, French 525, Spanish 437, Italian 331, Polish 144, Portuguese 143, Romanian 103, Dutch 100, Swedish 63, Hungarian 37, Bulgarian 34, Czech 31, Greek 29, Finnish 28, Slovenian 26, Slovak 25, Danish 19, Latvian 11, Croatian 10, Estonian 10, Lithuanian 5, Irish 0, and Maltese 0; TikTok notes that Czech, Slovak, and Slovenian are one team, that Croatian moderators also cover Serbian, and that the totals also include Arabic, Catalan, Hindi, Pashto, Persian, Turkish, Ukrainian, Norwegian, Russian, and Icelandic capacity. EUobserver, reading the sixth report for July to December 2025, records 91 employed staff against 3,583 contracted human moderators. Over the same span TikTok reported 169 million EU monthly active recipients in the first half of 2025 against the 135.9 million declared at designation in April 2023; Social Media Today computes the September-2023-to-June-2025 change as an audience up about 25 percent and a moderation workforce down about 26 percent, which is an analyst's arithmetic over TikTok's disclosures rather than a TikTok statement or an audited figure. The site-level events: the entire 300-person Netherlands moderation team closed in September 2024, as stated in European Parliament question E-002454/2024; fewer than 500 Malaysian roles were cut in October 2024; about 300 Dublin trust-and-safety roles were notified in March 2025 and about 300 more proposed on 1 July 2026 out of more than 2,000 Irish staff; about 150 Berlin trust-and-safety and TikTok Live roles were announced on 10 August 2025; and approximately 430 London roles were put at potential risk with notices on 22 August 2025. TikTok stressed to Parliament that no changes had yet taken effect, and site-level figures vary between sources — Berlin is reported as both about 150 and 160 of roughly 400 staff, and London as over 400, approximately 430, and 439 — so each figure is pinned to its source and its date and none are summed.

  empirical
  • Vendor TikTok Technology Limited (2025, 29 August). DSA Transparency Report, January to June 2025 (fifth report), Annex D values restated 15 April 2026 https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA_H1_2025/TikTok-DSATransparencyReport-January-June-2025.pdf
  • Trade press EUobserver (2026). TikTok used automation in nearly 100% of violating-content moderation in Europe (the independent reading of the sixth report carrying the 91 employed against 3,583 contracted split) https://euobserver.com/205083/tiktok-used-ai-in-nearly-100-of-violating-content-moderation-in-europe/
  • Trade press Social Media Today (2025). TikTok Continues To Grow in EU, Reduces Moderation Staff (the 26 per cent reduction is this analyst's arithmetic over the operator's own disclosures, not an operator statement and not an audited figure) https://www.socialmediatoday.com/news/tiktok-europe-user-numbers-moderation-staff-enforcement/759058/
  • Vendor Letter from TikTok to the Chair of the House of Commons Science, Innovation and Technology Committee, 7 November 2025, published by the Committee (the thirds breakdown, the labelling rationale and the vertical re-partition) https://committees.parliament.uk/publications/50179/documents/270768/default/
  • Advocacy European Federation of Journalists and UNI Global Union (2025). Germany: TikTok workers on strike to secure collective agreement; with the Business & Human Rights Resource Centre entry carrying the company response https://europeanjournalists.org/blog/2025/09/23/germany-tiktok-workers-on-strike-to-secure-collective-agreement-over-ai-taking-their-jobs/
  • Trade press RTE and the Irish Examiner (2026, July 1). Around 300 jobs under threat at TikTok's Irish operation, and the Communications Workers' Union response on user safety https://www.rte.ie/news/business/2026/0701/1581229-tiktok-to-cut-300-irish-jobs/

• TikTok's own account of the London reduction is that two thirds of it is not front-line moderation but the functions that maintain moderation. Writing to Dame Chi Onwurah MP, Chair of the House of Commons Science, Innovation and Technology Committee, on 7 November 2025, TikTok stated: 'In the UK, there are approximately 430 roles at potential risk under this proposal', and that 'a third of these roles are in teams involved in the labelling of data for AI model training. Progress in the development of these models has significantly decreased the need for this kind of manual labelling. Another significant proportion of those potentially affected are in ancillary roles, for example training teams, whose duties include activities such as training moderators on our Community Guidelines... Around a third of those impacted are front line moderation teams.' Its earlier letter of 20 October 2025 named the labelling team as the AI Data Service and Operations team and said that 'the majority of those potentially affected are not in front line moderation roles'. The same 7 November letter describes a structural re-partition rather than only a reduction: TikTok is 'moving from a region-based structure of generalised moderators to one based on different types of products or types of risk, known as verticals' — harassment, misinformation, fraud — consolidated into fewer sites and supplemented by third-party specialists offering 'greater ability to rapidly expand' and 'greater levels of language-specific, follow-the-sun coverage', which it says is 'not a like-for-like replacement'. TikTok's published Year 3 risk assessment describes the same functions from the other side: safety-topic experts and local-market experts write and update the keyword lists and detection rules the classifiers execute, and moderators apply a Moderation Policy Framework. The language-indexed staffing table the DSA requires TikTok to publish therefore describes a partition the operator states it is leaving.

  empirical
  • Vendor Letter from TikTok to the Chair of the House of Commons Science, Innovation and Technology Committee, 7 November 2025, published by the Committee (the thirds breakdown, the labelling rationale and the vertical re-partition) https://committees.parliament.uk/publications/50179/documents/270768/default/
  • Vendor Letter from TikTok to the Chair of the House of Commons Science, Innovation and Technology Committee, 20 October 2025, published by the Committee (the 86 per cent figure, the over-400 London figure and the categorically-untrue denial) https://committees.parliament.uk/publications/49886/documents/267685/default
  • Vendor TikTok Technology Limited (2025, 28 August). DSA Risk Assessment Report 2025 (Year 3 systemic risk assessment); carried also for the absence of any reference to the workforce reduction in the published text https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA/DSARiskAssessmentReport-TikTokTechnologyLimited-2025.pdf
  • Vendor TikTok Technology Limited (2025, 29 August). DSA Transparency Report, January to June 2025 (fifth report), Annex D values restated 15 April 2026 https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA_H1_2025/TikTok-DSATransparencyReport-January-June-2025.pdf

• TikTok's own automation figures rose across the substitution period under changing definitions, and they are three separately sourced statements rather than a trend. Its DSA transparency report for January to June 2025 gives 17,729,896 automatic removals of 24,534,707 total Community Guidelines removals, or 72.3 percent. Its letter to the House of Commons Science, Innovation and Technology Committee of 20 October 2025 states that '86% of the content we remove is now removed by automation', attributed to an April-to-June 2025 transparency report from a different report family. Its sixth DSA transparency report, for July to December 2025, states that 'Automated systems actioned 93.8% of all violating content without human review', with '97.6% of automated enforcement decisions being confirmed as correct', against approximately 112 million pieces of violating content — and that period was the first to include comment enforcement volumes, so the denominator is not the previous periods'. TikTok itself cautions that the later report 'captures a broader range of automated enforcement actions, including automated LIVE enforcement, when compared with our previous reports'. Social Media Today reads the same disclosures as automated detection for mental and behavioural health concerns rising from 49 percent in 2023 to 90 percent, and for youth safety from 38 percent to 77 percent. The three headline percentages measure different things over different denominators and are never presented here as a time series.

  empirical
  • Vendor TikTok Technology Limited (2025, 29 August). DSA Transparency Report, January to June 2025 (fifth report), Annex D values restated 15 April 2026 https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA_H1_2025/TikTok-DSATransparencyReport-January-June-2025.pdf
  • Vendor Letter from TikTok to the Chair of the House of Commons Science, Innovation and Technology Committee, 20 October 2025, published by the Committee (the 86 per cent figure, the over-400 London figure and the categorically-untrue denial) https://committees.parliament.uk/publications/49886/documents/267685/default
  • Vendor TikTok (2026, 27 February). Digital Services Act: our sixth transparency report on content moderation in Europe (July to December 2025) https://newsroom.tiktok.com/digital-services-act-our-sixth-transparency-report-on-content-moderation-in-europe?lang=en-150
  • Trade press EUobserver (2026). TikTok used automation in nearly 100% of violating-content moderation in Europe (the independent reading of the sixth report carrying the 91 employed against 3,583 contracted split) https://euobserver.com/205083/tiktok-used-ai-in-nearly-100-of-violating-content-moderation-in-europe/
  • Trade press Social Media Today (2025). TikTok Continues To Grow in EU, Reduces Moderation Staff (the 26 per cent reduction is this analyst's arithmetic over the operator's own disclosures, not an operator statement and not an audited figure) https://www.socialmediatoday.com/news/tiktok-europe-user-numbers-moderation-staff-enforcement/759058/

• The independent audit the Digital Services Act itself mandates returned a NEGATIVE, qualified opinion on TikTok Technology Limited for the year 1 July 2024 to 30 June 2025. KPMG Advisory N.V.'s Article 37 assurance report, dated 29 August 2025, covers 90 specified requirements and reaches four negative conclusions, all inside the moderation record. On Article 16(6) it identified notices where TikTok did not perform moderation actions and could not evidence the monitoring controls over the interface between notice intake and the moderation systems, so it 'could not obtain sufficient assurance to support the completeness of the total population of notices'. On Article 20(4) complaint records could not be retrieved 'due to limitations in documentation retention', so KPMG 'was unable to confirm that all complaints were handled in a timely, non-discriminatory, diligent, and non-arbitrary manner'. On Article 24(5) duplicate statements of reasons were transmitted to the Commission's DSA Transparency Database, producing 'more records... than the actual number of decisions taken', until a remediation in May 2025, and post-remediation sampling still found statements of reasons that were never transmitted. And the advertisement repository was found defective under Article 39(3). Six requirements were DISCLAIMED because they sit under the Commission's open proceedings — Articles 28(1), 34(1), 34(2), 35(1), 39(1) and 40(12) — with material observations on two. On Article 42(2), the article that mandates the per-language human-resources figures and the accuracy indicators themselves, the conclusion is 'Positive with comments', the comment being that 'internal controls concerning data accuracy and completeness monitoring, between the various source systems and Transparency Report are not sufficient and appropriate'. The auditor holds no remedy power of any kind.

  empirical
  • Vendor KPMG Advisory N.V. (2025). DSA Assurance Report: independent practitioner's assurance report concerning Regulation (EU) 2022/2065, TikTok Technology Limited, 1 July 2024 to 30 June 2025 (NEGATIVE, qualified opinion). Statutorily mandated, operator-commissioned and operator-published, hence vendor tier despite independent authorship https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA/DSAAssuranceReport-TikTokTechnologyLimited-2025.pdf
  • Vendor TikTok Technology Limited (2025, 29 August). DSA Transparency Report, January to June 2025 (fifth report), Annex D values restated 15 April 2026 https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA_H1_2025/TikTok-DSATransparencyReport-January-June-2025.pdf

• TikTok restated the moderator counts in Annex D of its January-to-June 2025 DSA transparency report on 15 April 2026, nearly eight months after publication, noting that the values 'have been updated with the correct values'. The mandated record of how many people moderate content in each EU official language was wrong when it was first published, and the independent auditor's comment on the article that mandates that record is that the internal controls over data accuracy and completeness monitoring between the source systems and the transparency report are not sufficient and appropriate. The consequence for anyone reading the series is that headcount figures from different report editions are not safely comparable: the July-to-December 2024 per-language figures reach this file through a civil-society transcription by INACH rather than the original document — English 1,524, German 532, Spanish 531, French 509, Italian 290, Portuguese 160, Polish 146, Dutch 99, Romanian 99, Swedish 72, Czech 53, Hungarian 51, Greek 50, Bulgarian 38, Slovenian 37, Slovak 33, Finnish 31, Croatian 29, Latvian 22, Lithuanian 19, Estonian 17, and Danish 15, totalling 4,357 across 22 named languages — one of the two editions was restated, and TikTok groups several language teams and folds several non-EU languages into its totals. The aggregate is nearly flat between the two while the smallest columns fall hard and the largest rise; the direction and the shape of that distribution are usable and the differences are not. INACH's own judgement of the sector's reports is that TikTok 'has almost no data available specifically on hate speech as a separate category' and that none of the reports reviewed provides a breakdown of hate-speech enforcement by country or by protected characteristic.

  empirical
  • Vendor TikTok Technology Limited (2025, 29 August). DSA Transparency Report, January to June 2025 (fifth report), Annex D values restated 15 April 2026 https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA_H1_2025/TikTok-DSATransparencyReport-January-June-2025.pdf
  • Advocacy International Network Against Cyber Hate (2025). Overview of the latest transparency reports under the DSA (executive summary brief) https://www.inach.net/wp-content/uploads/Transparency-Reports-2024_2025-DSA-2-1.pdf
  • Vendor KPMG Advisory N.V. (2025). DSA Assurance Report: independent practitioner's assurance report concerning Regulation (EU) 2022/2065, TikTok Technology Limited, 1 July 2024 to 30 June 2025 (NEGATIVE, qualified opinion). Statutorily mandated, operator-commissioned and operator-published, hence vendor tier despite independent authorship https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA/DSAAssuranceReport-TikTokTechnologyLimited-2025.pdf

• TikTok's contest machinery under the Digital Services Act runs on two levels with very different force, and both are measured in its own January-to-June 2025 report. Internally, 3,075,758 appeals came from uploaders and advertisers and 1,054,432 from users who had reported content — about 22,700 a day against 4,596 moderators — producing 1,359,823 pieces of content reinstated or unrestricted and 61,095 items removed after a reporter's appeal, at a median decision time under two hours on both tracks; TikTok warns the reinstatement count does not map one to one onto the period's appeals. Externally, under Article 21, out-of-court dispute settlement bodies received 1,121 complaints and closed 498 in period: the body agreed with TikTok in 113 cases, disagreed in 106, and 189 closed without a formal decision. Of the 106 decisions that went against TikTok, TikTok implemented the decision in 29 — 27.4 percent — at a median handling time of about 26 days against the internal channel's under two hours. That gap is the difference between a body that can decide and a body that can bind. Article 20 requires the internal channel to produce reasoned decisions under the supervision of appropriately qualified staff and not solely on the basis of automated means, and the Commission repeated that requirement in its written answer on moderation resourcing. The independent auditor's conclusion on Article 20(4) is negative: complaint records could not be retrieved 'due to limitations in documentation retention', so it 'was unable to confirm that all complaints were handled in a timely, non-discriminatory, diligent, and non-arbitrary manner'.

  empirical
  • Vendor TikTok Technology Limited (2025, 29 August). DSA Transparency Report, January to June 2025 (fifth report), Annex D values restated 15 April 2026 https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA_H1_2025/TikTok-DSATransparencyReport-January-June-2025.pdf
  • Vendor KPMG Advisory N.V. (2025). DSA Assurance Report: independent practitioner's assurance report concerning Regulation (EU) 2022/2065, TikTok Technology Limited, 1 July 2024 to 30 June 2025 (NEGATIVE, qualified opinion). Statutorily mandated, operator-commissioned and operator-published, hence vendor tier despite independent authorship https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA/DSAAssuranceReport-TikTokTechnologyLimited-2025.pdf
  • Government European Parliament written question E-002454/2024 (Van Sparrentak) and the Commission's answer of 15 January 2025: 'The DSA does not prescribe any specific rules about the resources to be dedicated to content moderation' https://www.europarl.europa.eu/doceo/document/E-10-2024-002454-ASW_EN.html

• No published independent measurement of a moderation-quality effect from TikTok's staffing substitution exists, and the three available quality signals are each weak in a different way. The first is TikTok's own indicator, which is an overturn-rate complement rather than a population error rate: it defines accuracy as 'the proportion of content where the original enforcement decision was upheld or maintained' and error as 'the proportion... overturned', reporting automated accuracy of 99.2 percent and error of 0.8 percent for January to June 2025 against 99.12 percent the previous half, with per-Member-State error running from 0.3 percent in Slovakia and Bulgaria to 1.6 percent in Austria and Germany and France at 1.4 percent, and 97.6 percent of automated enforcement decisions 'confirmed as correct' for July to December 2025. That statistic is conditional on somebody appealing and is computed by the party that made the original decision, and in the same six months 3,075,758 uploader appeals produced 1,359,823 reinstatements or lifted restrictions on a different denominator that cannot be turned into an error rate. The second is the independent Article 37 auditor, which could not establish the completeness of the notice population or of the complaint population from which that sample is drawn. The third is negative: on 24 October 2025 the European Commission preliminarily found that TikTok and Meta may have put in place burdensome procedures and tools for researchers to request access to public data, 'often leav[ing] them with partial or unreliable data' — so the outside parties who could measure a quality effect independently are the ones the regulator says cannot get reliable data. Nothing in this file attributes any change in moderation quality to the staffing substitution, because no such measurement has been made. TikTok's claim to Parliament that its platform has 'the lowest error rates and highest accuracy rates among all major platforms' is an operator claim across disclosures that civil-society and academic reviewers describe as mutually incomparable, and the ranking is not repeated here.

  empirical
  • Vendor TikTok Technology Limited (2025, 29 August). DSA Transparency Report, January to June 2025 (fifth report), Annex D values restated 15 April 2026 https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA_H1_2025/TikTok-DSATransparencyReport-January-June-2025.pdf
  • Vendor TikTok (2026, 27 February). Digital Services Act: our sixth transparency report on content moderation in Europe (July to December 2025) https://newsroom.tiktok.com/digital-services-act-our-sixth-transparency-report-on-content-moderation-in-europe?lang=en-150
  • Vendor KPMG Advisory N.V. (2025). DSA Assurance Report: independent practitioner's assurance report concerning Regulation (EU) 2022/2065, TikTok Technology Limited, 1 July 2024 to 30 June 2025 (NEGATIVE, qualified opinion). Statutorily mandated, operator-commissioned and operator-published, hence vendor tier despite independent authorship https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA/DSAAssuranceReport-TikTokTechnologyLimited-2025.pdf
  • Government European Commission (2025-2026). Preliminary findings on TikTok's ad repository (IP/25/1223, 15 May 2025), on researcher data access (IP/25/2503, 24 October 2025), on addictive design (6 February 2026) and on minors' account settings (IP/26/1679, 24 July 2026); with the advertising-transparency commitments decision of 5 December 2025 and the TikTok Lite Rewards closure of 5 August 2024 (IP/24/4161). PRELIMINARY FINDINGS ARE NOT FINDINGS OF BREACH https://ec.europa.eu/commission/presscorner/api/files/document/print/en/ip_25_2503/IP_25_2503_EN.pdf
  • Government UK House of Commons Science, Innovation and Technology Committee (2025). Chair's letter of 28 October 2025 on the staffing of TikTok's Trust and Safety teams, and the Committee news release of 13 November 2025, 'TikTok fails to share evidence behind increased AI use in content moderation' https://committees.parliament.uk/committee/135/science-innovation-and-technology-committee/news/210394/tiktok-fails-to-share-evidence-behind-increased-ai-use-in-content-moderation/
  • Advocacy International Network Against Cyber Hate (2025). Overview of the latest transparency reports under the DSA (executive summary brief) https://www.inach.net/wp-content/uploads/Transparency-Reports-2024_2025-DSA-2-1.pdf

• The House of Commons Science, Innovation and Technology Committee asked TikTok on 28 October 2025 six questions about the UK trust-and-safety reductions, including the total job losses, the core responsibilities of the roles and how they support moderation, whether a risk assessment of the job losses for UK user safety had been conducted and what its outcome was, whether third-party moderation teams would replace the responsibilities, Ofcom's response, and how UK users would be safeguarded by staff in other countries. The Chair's letter quoted TikTok's own earlier written evidence back to it — 'tens of thousands' of safety professionals working 24/7 and 'In 2024, we invested over $2 billion in our Trust and Safety efforts' — and its oral evidence of 25 February 2025 distinguishing what automation handles well, 'pornographic material, blood and that kind of thing', from content referred to 'human moderators who have to use their nuance, skills and training to be able to rule on other elements that can include hateful behaviour and misinformation'. TikTok replied on 7 November 2025 describing an internal analysis expecting improvements in 'speed of moderation' and 'efficacy of moderation (i.e. how often the moderation decision is the correct one)', and supplied no data. On 13 November the Committee published the reply, stating that 'TikTok did not share its data or risk assessment that justified this in its reply to the Committee Chair', with Chair Dame Chi Onwurah commenting that 'TikTok's response represents a commitment to reducing staffing levels in favour of increasing the use of AI to moderate content on its platform. But TikTok have come up empty to show that this transition to AI won't lead to more harms for its users', and that 'TikTok refers to evidence showing that their proposed staffing cuts and changes will improve content moderation and fact-checking - but at no point do they present any credible data on this to us.' The Committee's instrument is scrutiny and publication; it has no authority over the moderation design. TikTok confirmed in the same correspondence that it had given Ofcom prior notice of the London proposals and that it has been regulated by Ofcom since 2021 under the Video Sharing Platform regime and now under the Online Safety Act; Ofcom's response is not on the public record.

  empirical
  • Government UK House of Commons Science, Innovation and Technology Committee (2025). Chair's letter of 28 October 2025 on the staffing of TikTok's Trust and Safety teams, and the Committee news release of 13 November 2025, 'TikTok fails to share evidence behind increased AI use in content moderation' https://committees.parliament.uk/committee/135/science-innovation-and-technology-committee/news/210394/tiktok-fails-to-share-evidence-behind-increased-ai-use-in-content-moderation/
  • Vendor Letter from TikTok to the Chair of the House of Commons Science, Innovation and Technology Committee, 7 November 2025, published by the Committee (the thirds breakdown, the labelling rationale and the vertical re-partition) https://committees.parliament.uk/publications/50179/documents/270768/default/
  • Vendor Letter from TikTok to the Chair of the House of Commons Science, Innovation and Technology Committee, 20 October 2025, published by the Committee (the 86 per cent figure, the over-400 London figure and the categorically-untrue denial) https://committees.parliament.uk/publications/49886/documents/267685/default

• Two labour disputes over the reductions are live and neither has been adjudicated. In Berlin, ver.di held five one-day strikes in July 2025, beginning 23 July, and a four-day strike from 23 September 2025, over the announcement of 10 August 2025 that the Trust and Safety and TikTok Live teams of about 150 people would close; its demands were a collective agreement with severance worth three years' salary and a twelve-month notice extension, under the slogan 'We trained your machines, pay us what we deserve!'. Jacobin, reporting worker accounts, gives 160 of roughly 400 Berlin staff affected, moderators reviewing 800 to 1,000 videos a day, and a typical tenure of three to four years before psychological strain ends it — testimony relayed by a partisan outlet and the only per-head capacity referents anywhere in this record, since TikTok publishes none. TikTok's response, carried in the Business & Human Rights Resource Centre tracker, is that the changes would 'streamline workflows and improve efficiency' with 'full commitment to protecting safety and integrity'. Dismissal cases went to the Berlin Labour Court and no outcome is on the public record. In London, the redundancy notices of 22 August 2025 preceded a scheduled union-recognition ballot with UTAW, a branch of the Communication Workers Union, and on 19 December 2025 two moderators supported by Foxglove and UTAW and represented by Leigh Day sent a pre-action letter alleging unlawful detriment and automatic unfair dismissal, citing internal TikTok documents of May 2025 referencing the 'complexity and volume of certain categories of moderation that require human judgment for safety and compliance, rather than automated tools'. TikTok told Parliament the union-timing claims are 'categorically untrue', that the decisions were made globally, and that it had written to the CWU expressing 'regret about these timescales' while remaining open to re-engaging after consultation. These are allegations and denials, not findings; no tribunal has ruled. In Ireland the Communications Workers' Union objected to the July 2026 Dublin proposal on user-safety grounds, arguing that 'only quality jobs can provide the level of rapid policy response, moderation oversight, and overall safety that are required' while TikTok is under Commission and national investigation; that is a union contention about a future effect, not a measurement.

  empirical
  • Advocacy European Federation of Journalists and UNI Global Union (2025). Germany: TikTok workers on strike to secure collective agreement; with the Business & Human Rights Resource Centre entry carrying the company response https://europeanjournalists.org/blog/2025/09/23/germany-tiktok-workers-on-strike-to-secure-collective-agreement-over-ai-taking-their-jobs/
  • Trade press Jacobin (2025, August). Berlin's Striking TikTok Workers Stand Up to a Tech Giant (worker accounts relayed by a partisan outlet; the only per-head throughput and tenure referents in the record, and the operator publishes none) https://jacobin.com/2025/08/berlin-tiktok-tech-ai-strike
  • Advocacy Foxglove (2025, December 19). Press release: TikTok faces first legal action over unlawful union-busting in London (a pre-action letter and an allegation; the operator denies it and no tribunal has ruled) https://www.foxglove.org.uk/2025/12/19/tiktok-first-legal-action-union-busting-london/
  • Vendor Letter from TikTok to the Chair of the House of Commons Science, Innovation and Technology Committee, 20 October 2025, published by the Committee (the 86 per cent figure, the over-400 London figure and the categorically-untrue denial) https://committees.parliament.uk/publications/49886/documents/267685/default
  • Trade press RTE and the Irish Examiner (2026, July 1). Around 300 jobs under threat at TikTok's Irish operation, and the Communications Workers' Union response on user safety https://www.rte.ie/news/business/2026/0701/1581229-tiktok-to-cut-300-irish-jobs/

• TikTok's published Year 3 systemic risk assessment, dated 28 August 2025 — eighteen days after the Berlin closure announcement and six days after the London redundancy notices — records no change to its Fundamental Rights inherent risk score, which stands at Medium-High and Likely and is described as 'consistent with TikTok's score in Year 2', and contains no reference to the trust-and-safety workforce reduction, restructuring, or redundancies. A keyword sweep of the published report for redundancy, restructuring, headcount, layoff, workforce, union, and labour returns nothing on the reduction announced in the same month. This is an observation about what the published text contains and not proof that the matter was never assessed internally: the document is headed 'Confidential' and what is public is its published version. The assessment does name over-moderation and under-moderation by 'content moderation systems and human moderators' as a standing moderation risk, and it records the internal approval chain through specialist Risk Assessment Review Groups, the Online Safety Oversight Committee, and the Board of Directors of TikTok Technology Ireland.

  empirical
  • Vendor TikTok Technology Limited (2025, 28 August). DSA Risk Assessment Report 2025 (Year 3 systemic risk assessment); carried also for the absence of any reference to the workforce reduction in the published text https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA/DSARiskAssessmentReport-TikTokTechnologyLimited-2025.pdf

• TikTok's published Year 3 systemic risk assessment describes the moderation pipeline in its own words. 'All video, photo and text-based content uploaded to the Platform are subject to a real time, technology-based automated review. While a video is undergoing this review, it is visible only to the uploading user/creator.' Detection uses 'vision-based, audio-based, text-based and LLM-based' technologies together with keyword lists and natural-language processing; no model, vendor, or product is identified anywhere in the record. Automated removal is 'applied when violations are the most clear-cut', and otherwise the item is routed to a human queue; high-view content may be routed for additional human review. Specialist misinformation moderators work against a repository of previously fact-checked claims from IFCN-accredited partners, and separate lanes handle illegal content, advertising, and marketplace listings. A strikes policy escalates to account bans, and every decision is appealable. The assessment also records the internal governance chain: each risk module goes to a specialist Risk Assessment Review Group of senior internal stakeholders with Compliance input, then to the Online Safety Oversight Committee, a cross-functional leaders' steering group, then to the Board of Directors of TikTok Technology Ireland for review and approval, across twelve risk modules under four categories. It names as a standing moderation risk 'The risk that TikTok's content moderation systems and human moderators may: (i) over-moderate... or (ii) under-moderate'.

  empirical
  • Vendor TikTok Technology Limited (2025, 28 August). DSA Risk Assessment Report 2025 (Year 3 systemic risk assessment); carried also for the absence of any reference to the workforce reduction in the published text https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA/DSARiskAssessmentReport-TikTokTechnologyLimited-2025.pdf
  • Vendor TikTok Technology Limited (2025, 29 August). DSA Transparency Report, January to June 2025 (fifth report), Annex D values restated 15 April 2026 https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA_H1_2025/TikTok-DSATransparencyReport-January-June-2025.pdf

• The enforcement, notice, and appeal volumes for TikTok in the European Union, January to June 2025, all from its own DSA transparency report: 24,534,707 Community Guidelines removals of which 17,729,896 were automatic, 2,470,592 advertising removals, and 829,861 TikTok Shop removals; 169,527,678 content restrictions, roughly seven times the removal layer; 2,781,470 service restrictions; and 4,906,735 account bans or suspensions of which 871,819 were automatic. The highest-volume removal policies were Regulated Goods and Commercial Activities at 9,485,450 (7,729,860 automatic), Sensitive and Mature Themes at 8,552,437 (6,305,376), Youth Safety and Well-Being at 5,944,993 (4,612,550), Mental and Behavioral Health at 5,581,036 (5,026,613, or 90.1 percent automatic) and Safety and Civility at 4,352,589 (2,855,504, or 65.6 percent, the lowest automation share of the major policies). On the notice side: 308,755 illegal-content reports from EU users covering 151,354 unique items, of which 26,512 were actioned as unlawful and 15,365 as policy breaches — 27.7 percent of unique reported items actioned at all — at a median decision time under 17 hours on policy grounds and under 21 hours on legal grounds; 3,976 removal orders from Member State authorities with a median action time under 3 hours, and 782 such requests in the following period led by France and Romania; and 82 trusted-flagger reports under Article 22, against 22,429 user reports of illegal hate speech and only five trusted-flagger hate-speech reports in the previous period on a civil-society transcription. In the July to December 2025 period EUobserver records approximately 112 million removals with 99.3 percent removed before any user report and 714,000 user-reported incidents. Every figure here is TikTok's own, published because the Digital Services Act compels it, and the independent auditor found the controls over the data behind the transparency report not sufficient and appropriate.

  empirical
  • Vendor TikTok Technology Limited (2025, 29 August). DSA Transparency Report, January to June 2025 (fifth report), Annex D values restated 15 April 2026 https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA_H1_2025/TikTok-DSATransparencyReport-January-June-2025.pdf
  • Trade press EUobserver (2026). TikTok used automation in nearly 100% of violating-content moderation in Europe (the independent reading of the sixth report carrying the 91 employed against 3,583 contracted split) https://euobserver.com/205083/tiktok-used-ai-in-nearly-100-of-violating-content-moderation-in-europe/
  • Advocacy International Network Against Cyber Hate (2025). Overview of the latest transparency reports under the DSA (executive summary brief) https://www.inach.net/wp-content/uploads/Transparency-Reports-2024_2025-DSA-2-1.pdf
  • Vendor KPMG Advisory N.V. (2025). DSA Assurance Report: independent practitioner's assurance report concerning Regulation (EU) 2022/2065, TikTok Technology Limited, 1 July 2024 to 30 June 2025 (NEGATIVE, qualified opinion). Statutorily mandated, operator-commissioned and operator-published, hence vendor tier despite independent authorship https://sf16-va.tiktokcdn.com/obj/eden-va2/zayvwlY_fjulyhwzuhy%5B/ljhwZthlaukjlkulzlp/DSA/DSAAssuranceReport-TikTokTechnologyLimited-2025.pdf

Where this connects

Institutional pressures in this domain

• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
• Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
• Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).

All of them in context on the Content moderation & editorial AI domain page.

Levers available here and the patterns behind them

• Gate record entries — Human-in-the-loop write gating
• Review the riskiest first — Risk-tiered oversight
• Escalate checks — State-feedback vigilance
• Mark AI-written records — Provenance labeling
• Check copied records — Reconcile copied records
• Review on schedule — Oversight cadence & retrospectives
• Understand the system — Understand the system
• Assign a challenger — Structured dissent
• Upgrade model — Improve the model
• Gate vendor updates — Vendor quality gate

Documented case histories

• TikTok EU and UK trust-and-safety staffing substitution
• The errors that became visible when the reviewers went home
• The most built-out correction structure and the reach it doesn't have
• The byline nobody was behind
• A staff byline the AI wrote and the review it implied
• StopNCII & Take It Down
• X Multilingual Hate-Speech Enforcement
• X Community Notes (crowd annotation)
• GIFCT hash-sharing database
• Google CSAM detection and total account closure
• Meta cross-check: the enforcement-exemption tier
• The CyberTipline: triage under a rule against looking
• Sama Nairobi: the review workforce as the governed subsystem
• The score is published and the service cannot act on it
• YouTube Content ID