Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

YouTube's Content ID copyright matching system

The party that gains answers the objection

Almost every deployment in this atlas is a machine an organisation points at its own decisions. This one is a machine an organisation points at everybody else's uploads on behalf of a third party, and the third party decides. Modelled on the documented record of YouTube's Content ID, drawn from four machine-readable editions of the operator's own Copyright Transparency Report, its help documentation, the United States Copyright Office's 2020 report on the statute, a federal criminal record, the operator's own federal complaint, and the published academic and advocacy analysis. Start with the volume, because it is the largest in this collection: 2,502,941,368 copyright claims in calendar 2025, up 14 per cent on the year before, and 99.48 per cent of every copyright action taken on the platform that year. Over 99 per cent of those claims were generated by a fingerprint comparison with nobody deciding on the claiming side. Now the part that makes this shape different from every other board here. A company admitted through an access gate delivers reference files and writes a standing instruction for each asset and each territory — block the video, take its advertising revenue, or track its figures. When the comparison finds a match, that instruction fires by itself. The platform runs the machine and hosts the objection process and says, in its own words, that it is not in a position to mediate this type of dispute because it is not a court of law. So the decision principal is an outside company, and the output is a private allocation between two parties. More than nine claims in ten take the money rather than the video. That is why the errors here are quiet: an over-broad claim usually produces no takedown, no strike and nothing anyone else can see — only a revenue stream that moves to somebody who did not make the thing. Then look at the objection process, because the inversion is the whole case. An uploader may dispute. The dispute goes to the party that made the claim, which has thirty days to answer and whose silence releases the claim automatically. If it reinstates, the uploader may appeal; the claimant then has seven days, cut from thirty in September 2022. At that rung the claimant may no longer reinstate and must either release the claim or convert it into a legal removal request — which carries a copyright strike, and three strikes in ninety days ends the account and every channel attached to it. Every rung the uploader climbs raises the chance the other party turns the matter into something that can cost them everything. The operator's own integers price that: of 45,724 failed appeals in one half-year, 13,841 produced a removal and the other 31,883 ended because the uploader cancelled the appeal or deleted the video. Roughly seven in ten people who had already lost twice walked away rather than take the next step. Hold the dispute rate carefully, because it is the number most often read backwards. About half a per cent of claims are ever disputed — 3,698,019 of 722,649,569 in the first verified half-year, 12,840,608 of 2,502,941,368 four years later — and the rate barely moved while volume more than tripled. The operator reads that as accuracy. The same report contains the reason it cannot be: pushback is highest in the tool with the widest access, over 5 per cent counter-notification on the open public webform, and lowest in the tool with the tightest, under 1 per cent here. An error signal does not behave that way. A low dispute rate is equally consistent with accuracy and with deterrence, and nothing published separates them. Two more things belong on the board before you touch it. The store is the blast radius, and the operator says so: one bad webform notice takes down a handful of videos, while one bad reference file can affect hundreds or even thousands across the site — its own example is a news channel that submitted public-domain Mars-rover footage and then claimed everyone using it, including the space agency's own channel. And the access gate cuts both ways at once, measurably. Videos requested for removal through the open webform were classified as likely false assertions of ownership at over 8 per cent, over 5 per cent and over 6 per cent across the reported periods, against 0.5 per cent or lower in the limited-access tools — while 7,626 entities with access, 4,454 of them active, generate 99.48 per cent of all copyright enforcement on the platform. The gate suppresses abuse by more than an order of magnitude and concentrates enforcement authority in seven and a half thousand hands, and both halves come from the same published table. Before you pick a target level: this board cannot be won under Service and Safety Targets or All Governance Targets, and what holds it open is a channel, not a shortfall. Every arrangement of the ten instruments on offer was read at three strengths with the budget ignored entirely, 19,088 readings on each of those two cells, and not one clears every gate. One pathway stays open in every arrangement that gets close, and it is the same one every time: an admitted claiming party lending its access to a catalogue the platform never assessed, which is the route an adjudicated fraud ran through for roughly four years. Two instruments outside this board would close it — rules for what parties pass sideways, and a named challenger inside the process — and this record bars both. Nobody in it is documented setting rules for what an admitted party may do with its access, and the operator states that it is not in a position to mediate this type of dispute because it is not a court of law. Admit either instrument and the cheapest winning arrangement costs eighteen against the twelve you are given, so what stood in the way next would be price rather than this. That is a measurement of the deployment this network is drawn from, not a puzzle waiting to be cracked. Explore and Service Targets Only can be won, and cheaply: two instruments, costing five of your twelve.

Stylized model of a documented deploymentContent moderation & editorial AI

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Rightsholder-claims matching and private allocation network: 12 components and 25 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: 4 assumed · 1 calibrated · 8 measured. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • measured

    Every quantitative figure on this board is the deployer's own published accounting of its own system, published voluntarily in the United States, and each is tagged to its reporting period. The four half-year editions to the end of 2022 were machine-readable documents with exact integers; from calendar 2023 the report is an annual, interactive, web-only publication, and the post-2022 figures here are read through verified trade analysis because the live report renders in the browser. Half-year and full-year figures are not comparable and are never chained.

  • measured

    The dispute rate is not an error rate and nothing on this board treats it as one. The deployer's own tier comparison shows pushback rising as access broadens — over 5 per cent counter-notification on the open webform against under 1 per cent dispute on this channel — which is the opposite of what an error signal would do. A dispute rate of half a per cent is consistent with high accuracy and with high deterrence, and the published record cannot distinguish them.

  • assumed

    No figure is asserted anywhere for claims that were wrong and never disputed. Nobody has published one and none can be derived from the funnel. The contemporaneous analysis of the first edition derives a FLOOR of at least 2.2 million confirmed unjustified actions in one half-year and argues the true figure is higher; that derivation is carried as its author's, with the arithmetic shown rather than the conclusion asserted.

  • measured

    The share of disputes resolved for the uploader — over 60 per cent in the two verified half-years, over 65 per cent for calendar 2024 and 67.42 per cent for calendar 2025 — applies to the half a per cent of claims that are disputed, and is not a finding that two-thirds of claims are wrong. The deployer's own definition counts a dispute as resolved for the uploader when the claimant releases the claim OR does not answer within thirty days, so a large share of those wins are non-responses rather than determinations.

  • measured

    Two different figures on a 99 per cent scale appear in these reports and this board keeps them apart. The channel's share of ALL copyright actions on the platform is 99.43 per cent for calendar 2024 and 99.48 per cent for calendar 2025. The share of the channel's OWN claims generated by automated matching rather than partner manual claiming is over 99 per cent in every period, with manual claiming at 0.4 per cent, then under 0.5 per cent, then 0.31 per cent. Each sentence that uses one says which it means.

  • calibrated

    Rung widths for the seventeen pathways with a counterpart in the governance model this board is derived from are read off that model's own estimates on one stated four-rung mapping with no exceptions, and where one drawn pathway stands for several of that model's links it keeps its own rung and the others are narrated in its copy. Every link for this deployment is marked estimated in that model, so these are modelling widths over documented mechanism rather than measurements of the pathways themselves. The other eight pathways are derived from the cited record and say so on their own lines.

  • assumed

    No headcount, staffing level or review capacity is asserted for any desk on this board, because none is published. The deployer states that hundreds of millions of dollars have been invested in its copyright tools, and investment is not capacity. Demand and the counterfactual floor are derived from the published volume series and the published tier comparison instead.

  • measured

    The admission gate is drawn as a control with a measured effect in BOTH directions, because the deployer's own tier table supports both readings at once: abuse falls by more than an order of magnitude as access narrows, and the concentration of enforcement authority rises to roughly seven and a half thousand entities deciding 99.48 per cent of all copyright actions on the platform. The board asserts neither that the gate is unjustified nor that the concentration is harmless.

  • assumed

    The claim that access to the claiming tools is reserved for powerful rights-holders and unavailable to ordinary creators is an ALLEGATION that was never tested. It was pleaded in a federal class action and recorded in the federal copyright regulator's 2020 report as a complaint from commenters; class certification was denied in May 2023 and the case was dismissed with prejudice in June 2023, the day trial was to begin. There is no judicial finding about this system anywhere.

  • measured

    The adjudicated fraud on this board is fraud against RIGHTS-HOLDERS committed through the claiming tools, not over-claiming against uploaders. It evidences the vetting surface and the delegability of admission; it is not evidence about the comparison's accuracy, and the two are never merged.

  • measured

    The strike ledger drawn here is the COPYRIGHT strike ledger — three strikes in ninety days, written only by a claimant's conversion of a claim into a legal removal request. It is a different rule and a different count from the community-guidelines strike the same deployer applies to its own policy enforcement, which a separate case file in this catalogue models. No figure crosses between the two.

  • measured

    Uploaders and creators are not modelled. They hold the dispute and the appeal, they bear the revenue hold and the strike risk, and no outcome for any of them is computed from anything drawn here. The 31,883 abandoned appeals in one half-year are carried as the deployer's own count of a decision people made, never as an outcome this network produces.

  • assumed

    This board draws the deployment at the coarsest granularity at which every documented mechanism stays distinct. The standing match instruction is drawn inside the partner store it is set on, per asset and per territory; the annual published funnel is drawn inside the claim record it is compiled from; and the reads that only mirror a drawn write, or draw one documented act a second time, are narrated on the pathway that carries the act. Nothing documented about the deployment is dropped by drawing it this way: each such fact is stated on the element that now carries it.

What this example does not show

  • LITIGATION AND REGULATORY POSTURE, verbatim from the evidence dossier and load-bearing. Ongoing and expanding. Content ID is the single highest-volume automated adjudication system in the catalogue by decision count (2.5 billion claims in calendar 2025). No regulator has ordered a change to it; the only US litigation that directly attacked its access structure ended without any finding. The deployer continues to publish an annual funnel.
  • THE DISPUTE RATE IS NOT AN ERROR RATE, and this scenario never uses it as one. The operator's own tier comparison shows pushback rising as access broadens — over 5 per cent counter-notification on the open webform against under 1 per cent dispute on this channel — which is the opposite of what an error signal would do. A rate of half a per cent is consistent with high accuracy and with high deterrence at once, and the published record cannot distinguish them. No figure is asserted anywhere for claims that were wrong and never disputed, because nobody has published one and none can be derived.
  • THE ACCESS-ASYMMETRY CRITIQUE IS AN UNTESTED ALLEGATION. It was pleaded in a federal class action in the Northern District of California and recorded in the United States Copyright Office's 2020 report as a complaint from commenters, one of whom later sued. Class certification was denied on 22 May 2023 on the ground that classwide copyright ownership would entail individualised proof precluding certification, and on 12 June 2023, the day trial was to begin, the parties stipulated to dismissal WITH PREJUDICE of all claims raised or that could have been raised. There was no trial and no verdict. The opposite complaint is on the same federal record from rights-holders, who told the Office that the system misses a significant share of unauthorised uploads.
  • THE CRIMINAL RECORD IS FRAUD AGAINST RIGHTS-HOLDERS, NOT OVER-CLAIMING AGAINST UPLOADERS. Two principals of a rights company falsely claimed ownership of over fifty thousand recordings and monetised them through the claiming tools via a third-party rights administrator; both were convicted by plea after a thirty-count federal indictment in November 2021 and one was sentenced in June 2023 to seventy months. It evidences the vetting surface and the delegability of admission. It is not evidence about the comparison's accuracy, and the two are kept apart everywhere in this bundle. The individual defendants are not named and the intermediary is identified in the charging documents by initials only, so it is not named or guessed.
  • THE 2019 ACTION IN THE DISTRICT OF NEBRASKA IS A PLEADING, and it concerns a different channel. The operator itself sued under the statute's remedy for malicious takedown notices, alleging dozens of bogus notices and a threat to trigger a third strike unless creators paid. It settled in October 2019 without adjudication, the allegations are attributed to the operator as the pleading party, and they concern the public webform and strike system rather than this matching channel — though they terminate in the same strike ledger the appeal ladder here feeds.
  • THE REPORT CHANGED SHAPE AFTER 2022, and that matters for anyone quoting a figure from it. The first four editions were biannual downloadable documents with labelled exhibits and exact integers; from calendar 2023 the report became an annual, interactive, web-only publication, and the post-2022 figures used here are read through verified trade analysis because the live report renders in the browser. Comparisons across that boundary compare a half-year to a full year and a documented methodology to a rendered dashboard, and no figure in this scenario chains across it. Two trade readings of the same 2024 edition give 'over 65 per cent' and '70 per cent' for the uploader win rate; this scenario uses the operator's own published prose for 2024 and the exact 67.42 per cent for 2025, and records the discrepancy rather than averaging it.
  • SEVERAL 2025-EDITION FIGURES REST ON A SINGLE TRADE READING with ambiguous framing — notably a count of claims said to have reached a court stage and a stated appeal success rate. Neither is used anywhere in this bundle, and neither should be added without independent confirmation.
  • THIS IS NOT THE SAME DEPLOYMENT AS THE POLICY-ENFORCEMENT CASE ON THE SAME PLATFORM, and no figure crosses between them. That case is community-guidelines enforcement: the platform's own classifiers deciding whether a video breaks the platform's own rules, with the platform as decision-maker, a removal-appeal-reinstatement ladder, and a finding about what happened when human reviewers went home. This one is copyright matching: an identity comparison against a store, with an outside rights-holder as decision-maker, a statutory backdrop, a claim-dispute-appeal-removal-counter-notification ladder, and money rather than removal as the dominant outcome. The two strike ledgers are different ledgers under different rules. Neither case's findings substitute for the other's.
  • THE EUROPEAN DIMENSION IS CONTEXT, NOT A UNITED STATES REGULATORY FACT. The operator is a United States company and the litigation and criminal record cited are United States federal matters. The deployment is global and its published figures are worldwide with no country breakdown, a gap the contemporaneous analysis flags. Article 17 of the Copyright in the Digital Single Market Directive and the Digital Services Act supply reporting and redress obligations in Europe and are, on the analysis of the commentators cited here, part of why this report is published at all; they are named as context and nothing on this board is derived from them.
  • Uploaders, creators and rights-holding artists are not modelled. They hold the dispute and the appeal, they bear the revenue hold and the strike risk, and no outcome for any of them is computed from anything drawn here. The 31,883 abandoned appeals in one half-year, the individual artist losses recorded in the criminal case, and the ten-hour white-noise recording that drew five claims are all carried as measurements of the environment this network sits in — never as outputs of it, and no individual is named or characterised.

Sources and evidence

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

  • YouTube's Content ID is a fingerprint-matching copyright claiming system whose deciding party is an outside rights-holder rather than the platform. Rights-holders admitted through an eligibility gate deliver reference files; YouTube derives fingerprints and compares every upload against the reference store; on a match the partner's pre-set match policy fires automatically — block, monetize, or track — and the policy can differ country by country on the same video, with no case-by-case human decision on the claiming side. YouTube states that it 'is not in a position to mediate this type of dispute as we are not a court of law', and that when a matter reaches a legal removal request 'the ownership issue has exited the Content ID claim and dispute system built by YouTube, and enters the legal removal and remediation process defined by the DMCA and similar applicable laws'. The volume, on the deployer's own published reporting: 2,502,941,368 Content ID claims in calendar 2025, up 14 per cent on approximately 2.2 billion in calendar 2024, against 722,649,569 in the first half of 2021. TWO DISTINCT 99-PER-CENT FIGURES appear in these reports and mean different things. Content ID's share of ALL copyright actions taken on the platform was 99.43 per cent in 2024 and 99.48 per cent in 2025. The share of Content ID's OWN claims generated by automated matching rather than by a partner's manual claiming feature was 'over 99 per cent' in every reported period, with manual claiming at 0.4 per cent in the first half of 2021, 'fewer than 0.5 per cent' in the second half of 2022, and 0.31 per cent — about 6.9 million claims — in 2024. YouTube states the system cannot assess fair use: 'it's impossible for matching technology to take into account complex legal considerations like fair use or fair dealing.' Every quantitative figure here is the deployer's own accounting of its own system, published voluntarily in the United States, and none has been independently verified.

    empirical
    • Vendor YouTube / Google (2021-2023). Copyright Transparency Report, H1 2021, H2 2021, H1 2022 and H2 2022 (the four machine-readable PDF editions; the biannual series ends with H2 2022) https://storage.googleapis.com/transparencyreport/report-downloads/pdf-report-22_2022-7-1_2022-12-31_en_v1.pdf
    • Trade press TorrentFreak (2025, 2026). YouTube Processed 2.2 Billion Content ID Copyright Claims in 2024, and YouTube Processed 2.5 Billion Content ID Copyright Claims in 2025 (verified trade readings of the annual web-only editions, which render client-side) https://torrentfreak.com/youtube-processed-2-5-billion-content-id-copyright-claims-in-2025/
    • Vendor YouTube Help (2026). How Content ID works; Qualifying for Content ID; Dispute a Content ID claim; Copyright strike basics; Monetization during Content ID disputes https://support.google.com/youtube/answer/2797370
    • Government U.S. Copyright Office (2020, May). Section 512 of Title 17: A Report of the Register of Copyrights https://www.copyright.gov/policy/section512/section-512-full-report.pdf

  • About half a per cent of Content ID claims are ever disputed, the rate is stable across five years and a tripling of volume, and it is not an error rate. Verified from YouTube's four machine-readable Copyright Transparency Report editions: 3,698,019 disputes on 722,649,569 claims in the first half of 2021 (0.512 per cent) with 38,864 copyright removals originating from disputes; 3,810,395 on 759,540,199 in the second half of 2021 (0.502 per cent) with 43,198 removals; 3,690,786 on 757,993,607 in the first half of 2022 (0.487 per cent) with 24,931 removals; and 826,242,639 claims in the second half of 2022. Four years later the 2025 edition reports 12,840,608 disputes on 2,502,941,368 claims, 0.51 per cent. The share of disputes resolved in the uploader's favour was 'over 60 per cent' in the first half of 2021 and the second half of 2022, 'over 65 per cent' in the 2024 edition, and 67.42 per cent in the 2025 edition — and YouTube's own definition counts a dispute as resolved for the uploader when the claimant 'either voluntarily released the claim or did not respond within the 30-day window', so a large share of those outcomes are claimant non-responses rather than determinations. Two independent trade readings of the same 2024 edition give 'over 65 per cent' and '70 per cent'; the discrepancy is recorded rather than averaged. YouTube's own tier comparison is why the dispute rate cannot be read as accuracy: counter-notifications ran at over 5 per cent of removals through the open public webform in the first half of 2021 and over 4 per cent in the second half of 2022, against fewer than 2 per cent in the limited-access tools and under 1 per cent against Content ID claims — pushback rising as access broadens, the inverse of what an error signal would do. YouTube also reports repeatedly that manual claims are more than twice as likely to be disputed as automated ones: under 0.6 per cent against over 1 per cent in the first half of 2021, under 0.5 against over 0.9 in the second half of 2022, and 0.54 against 1.13 in 2024. Paul Keller of the Communia Association, writing for infojustice in December 2021, derived a FLOOR from the first edition's published numbers — 729.3 million copyright actions in six months, 3.7 million disputes, roughly 60 per cent resolved for the uploader, therefore at least 2.2 million confirmed unjustified actions in half a year — and argued the true figure is necessarily higher because most affected uploaders never complain, concluding that 'over-enforcement (both unjustified blocking and unjustified demonetisation) is a very real issue that affects the rights of a substantial number of uploaders on a regular basis'. That is his derivation from the deployer's own numbers, not a finding by anyone. YouTube publishes a caveat that cuts the other way too: dispute and counter-notification counts are trailing events that keep accruing after a period closes, so it snapshots them three months after period end and any rate read from a freshly closed period is an undercount by construction. No figure exists anywhere for claims that were wrong and were never disputed.

    empirical
    • Vendor YouTube / Google (2021-2023). Copyright Transparency Report, H1 2021, H2 2021, H1 2022 and H2 2022 (the four machine-readable PDF editions; the biannual series ends with H2 2022) https://storage.googleapis.com/transparencyreport/report-downloads/pdf-report-22_2022-7-1_2022-12-31_en_v1.pdf
    • Trade press TorrentFreak (2025, 2026). YouTube Processed 2.2 Billion Content ID Copyright Claims in 2024, and YouTube Processed 2.5 Billion Content ID Copyright Claims in 2025 (verified trade readings of the annual web-only editions, which render client-side) https://torrentfreak.com/youtube-processed-2-5-billion-content-id-copyright-claims-in-2025/
    • Academic Keller, P. (2021, December 10). YouTube Copyright Transparency Report: Overblocking is real. infojustice (American University Washington College of Law); with the same author's 2024 reading of the first annual edition on the Kluwer Copyright Blog https://infojustice.org/archives/43833
    • Vendor Google Transparency Report. YouTube Copyright Transparency Report (current interactive edition; live but client-rendered, which is why post-2022 figures here are carried through verified trade analysis) https://transparencyreport.google.com/youtube-copyright/everyone-has-access

  • Money rather than removal is the dominant outcome of a Content ID claim, which is why its error surface is nearly invisible. Over 90 per cent of Content ID claims are monetized rather than blocked, on YouTube's own reporting for the second half of 2022 and for calendar 2024: the claimed video stays up and the advertising revenue goes to the claimant, so an over-broad claim usually produces no takedown, no strike, and no visible trace, only a diverted revenue stream that the uploader must notice and contest to reverse. The revenue clock is published and it turns on speed rather than correctness. Revenue on a claimed video is held while the claimant reviews a dispute, but held from the CLAIM date only if the uploader disputes within five days of the claim; if the uploader disputes later, the hold runs only from the dispute date; and if the uploader takes no action within those five days, the revenue accrued in that window is paid to the CLAIMANT regardless of how the dispute is later resolved. Revenue data is also suppressed in the uploader's analytics while a claim is active. Cumulative Content ID payouts to rights-holders reached 5.5 billion United States dollars from advertising as of December 2020, 9 billion as of December 2022, and over 12 billion as of December 2024, of which approximately 3 billion in 2024 alone. The Electronic Frontier Foundation's 2020 study argues that this pricing produces pre-emptive self-censorship rather than contested claims: because Content ID cannot assess fair use and each rung of the ladder risks deplatforming or lost income, creators cut clips to a few seconds, re-edit videos as the matcher changes, and surrender revenue on uses copyright law would permit, being in that study's words 'so afraid of being deplatformed or losing that income' that the loop goes unused. That is the study's analysis, attributed to it.

    empirical
    • Vendor YouTube / Google (2021-2023). Copyright Transparency Report, H1 2021, H2 2021, H1 2022 and H2 2022 (the four machine-readable PDF editions; the biannual series ends with H2 2022) https://storage.googleapis.com/transparencyreport/report-downloads/pdf-report-22_2022-7-1_2022-12-31_en_v1.pdf
    • Vendor YouTube Help (2026). How Content ID works; Qualifying for Content ID; Dispute a Content ID claim; Copyright strike basics; Monetization during Content ID disputes https://support.google.com/youtube/answer/2797370
    • Trade press TorrentFreak (2025, 2026). YouTube Processed 2.2 Billion Content ID Copyright Claims in 2024, and YouTube Processed 2.5 Billion Content ID Copyright Claims in 2025 (verified trade readings of the annual web-only editions, which render client-side) https://torrentfreak.com/youtube-processed-2-5-billion-content-id-copyright-claims-in-2025/
    • Advocacy Trendacosta, K. (2020, December). Unfiltered: How YouTube's Content ID Discourages Fair Use and Dictates What We See Online. Electronic Frontier Foundation https://www.eff.org/wp/unfiltered-how-youtubes-content-id-discourages-fair-use-and-dictates-what-we-see-online

  • The Content ID objection ladder is documented, asymmetrically priced at every rung, and its deterrent effect is measurable in the deployer's own integers. An uploader may dispute a claim; the claimant has 30 days to respond and non-response releases the claim automatically. If the claimant reinstates, the uploader may appeal; the claimant then has 7 days, cut from 30 in September 2022 when an 'Escalate to Appeal' route was introduced. At that point the claimant may no longer reinstate and must either release the claim or file a legal removal request. A Content ID claim by itself carries no copyright strike; a legal removal request does. Three strikes in 90 days terminates the account and all associated channels, and strikes expire after 90 days if the uploader completes YouTube's Copyright School while holding fewer than three. After a counter-notification the claimant has 10 business days to show it has initiated court action or the content is reinstated — that window is statutory rather than YouTube's. THE MEASURED DETERRENCE: of 45,724 failed appeals in the second half of 2022, 13,841 (just over 30 per cent) resulted in a copyright removal, and the remaining 31,883 ended because the uploader cancelled the appeal or deleted the video rather than accept the strike risk. Roughly seven in ten uploaders who had already lost twice abandoned the matter rather than proceed. THE FUNNEL'S TAIL: in the same half-year YouTube accepted fewer than 25 per cent of the counter-notifications submitted to it and fewer than 1 per cent of counter-notifications resulted in a lawsuit, against 826,242,639 claims — six orders of magnitude of attrition from claim to court. Perel and Elkin-Koren documented in 2016 that appeal eligibility historically depended on the account being in 'good standing', so a prior strike could remove the ability to appeal the next claim; current documentation places Content ID appeal behind advanced-feature verification. This is the COPYRIGHT strike ledger, a different rule and a different count from the community-guidelines strike the same platform applies to its own policy enforcement.

    empirical
    • Vendor YouTube / Google (2021-2023). Copyright Transparency Report, H1 2021, H2 2021, H1 2022 and H2 2022 (the four machine-readable PDF editions; the biannual series ends with H2 2022) https://storage.googleapis.com/transparencyreport/report-downloads/pdf-report-22_2022-7-1_2022-12-31_en_v1.pdf
    • Vendor YouTube Help (2026). How Content ID works; Qualifying for Content ID; Dispute a Content ID claim; Copyright strike basics; Monetization during Content ID disputes https://support.google.com/youtube/answer/2797370
    • Academic Perel, M., & Elkin-Koren, N. (2016). Accountability in Algorithmic Copyright Enforcement. 19 Stanford Technology Law Review 473 https://law.stanford.edu/wp-content/uploads/2016/10/Accountability-in-Algorithmic-Copyright-Enforcement.pdf

  • Access to Content ID is rationed, and YouTube's own tier comparison shows the ration suppressing abuse and concentrating enforcement authority at the same time. In calendar 2025, 7,626 entities held Content ID access and 4,454 actively used it; in calendar 2024 the figures were 7,703 and 4,564; earlier editions give 'over 9,000 partners' as of December 2022 and the U.S. Copyright Office recorded over 9,000 rights-holders as of 2020. Against that, 295,531 claimants used the public copyright webform in 2025 and 173,338 used the Copyright Match Tool, with over 4 million channels holding Copyright Match Tool access as of December 2025 — up from over 2 million in July 2021 and over 2.5 million in December 2022. So roughly seven and a half thousand entities generate 99.48 per cent of all copyright actions on the platform. The stated criterion is exclusive rights to 'a substantial body of original material that is frequently uploaded by the YouTube creator community', plus demonstrated need and capacity, with categories excluded by rule: mashups, compilations, and remixes; video game footage and software visuals; unlicensed media; licensed content without exclusive rights; and recordings of performances, concerts, events, and speeches. A refused applicant may respond once with additional information. THE GATE'S MEASURED EFFECT, from the deployer's own reporting: videos requested for removal through the open public webform that YouTube's review team deemed 'a likely false assertion of copyright ownership' ran at over 8 per cent in the first half of 2021, over 5 per cent in the second half of 2022 and over 6 per cent in 2025, against 0.2 per cent or lower in the limited-access tools in the first half of 2021 and 0.5 per cent or lower in the second half of 2022; the 2025 edition describes the webform abuse rate as more than ten times that of all other copyright removal tools. YouTube states it terminates 'tens of thousands of accounts each year that attempt to abuse our copyright tools' and that claimants who repeatedly make erroneous Content ID claims can have Content ID access disabled and their partnership terminated — and publishes no count of partners actually de-accessed for erroneous claiming, the one funnel number absent from every edition. No headcount or review capacity is published for any of the copyright teams; the figure YouTube gives is 'hundreds of millions of dollars' invested in the Copyright Management Suite, which is investment rather than capacity.

    empirical
    • Trade press TorrentFreak (2025, 2026). YouTube Processed 2.2 Billion Content ID Copyright Claims in 2024, and YouTube Processed 2.5 Billion Content ID Copyright Claims in 2025 (verified trade readings of the annual web-only editions, which render client-side) https://torrentfreak.com/youtube-processed-2-5-billion-content-id-copyright-claims-in-2025/
    • Vendor YouTube / Google (2021-2023). Copyright Transparency Report, H1 2021, H2 2021, H1 2022 and H2 2022 (the four machine-readable PDF editions; the biannual series ends with H2 2022) https://storage.googleapis.com/transparencyreport/report-downloads/pdf-report-22_2022-7-1_2022-12-31_en_v1.pdf
    • Vendor YouTube Help (2026). How Content ID works; Qualifying for Content ID; Dispute a Content ID claim; Copyright strike basics; Monetization during Content ID disputes https://support.google.com/youtube/answer/2797370
    • Government U.S. Copyright Office (2020, May). Section 512 of Title 17: A Report of the Register of Copyrights https://www.copyright.gov/policy/section512/section-512-full-report.pdf
    • Academic Keller, P. (2021, December 10). YouTube Copyright Transparency Report: Overblocking is real. infojustice (American University Washington College of Law); with the same author's 2024 reading of the first annual edition on the Kluwer Copyright Blog https://infojustice.org/archives/43833

  • The reference store, not the comparison, is where Content ID's errors scale, and YouTube publishes both the mechanism and a worked example. Its own words: 'Just one bad copyright webform notice can result in a handful of videos being temporarily removed from YouTube. In Content ID the impact is multiplied due to its automated nature; one bad reference file can impact hundreds or even thousands of videos across the site.' The example the report gives is its own — a news channel uploaded public-domain NASA Mars-rover footage as a reference file and made claims against every other channel using the same footage, including NASA's own channel. There is a dedicated correction loop over the store, upstream of and independent from the per-claim dispute loop: a dedicated team plus automated systems detect bad or low-quality reference files; the partner may exclude the offending segment, remove the whole reference file, or ask for re-review; and if the partner does not respond the reference file is marked invalid and removed and ALL claims associated with it are released at once. YouTube names the recurring causes: partners delivering non-exclusive content, public-domain material, licensed-but-not-owned clips, or reference files capturing indistinct sound effects and nature sounds. Content ID also holds a queue of PENDING claims where reference files carry flawed or conflicting ownership data and the system is uncertain whether a claim should be made at all, with the conflict resolved between partners rather than against the uploader — an explicit abstain-and-hold path inside an otherwise fully automated channel. A DOCUMENTED FALSE-CLAIM CASE ON NON-COPYRIGHTABLE AUDIO: a ten-hour white-noise recording uploaded in 2015 had drawn five separate Content ID claims by January 2018, at least two of them matching other white-noise recordings held by a single company; all five claimants chose to MONETIZE rather than block, so the error's only visible effect was a diversion of advertising revenue, and the claims were released after press attention rather than through the dispute process. Nothing about a released claim propagates back into the reference file unless the integrity team independently flags it, so a reference file that should not have been admitted keeps generating claims against every future matching upload.

    empirical
    • Vendor YouTube / Google (2021-2023). Copyright Transparency Report, H1 2021, H2 2021, H1 2022 and H2 2022 (the four machine-readable PDF editions; the biannual series ends with H2 2022) https://storage.googleapis.com/transparencyreport/report-downloads/pdf-report-22_2022-7-1_2022-12-31_en_v1.pdf
    • Trade press TorrentFreak (2025, 2026). YouTube Processed 2.2 Billion Content ID Copyright Claims in 2024, and YouTube Processed 2.5 Billion Content ID Copyright Claims in 2025 (verified trade readings of the annual web-only editions, which render client-side) https://torrentfreak.com/youtube-processed-2-5-billion-content-id-copyright-claims-in-2025/
    • Advocacy Electronic Frontier Foundation (2018). Ten Hours of Static Gets Five Copyright Notices (Takedown Hall of Shame) https://www.eff.org/takedowns/ten-hours-static-gets-five-copyright-notices

  • Claimant-side fraud through Content ID is an adjudicated criminal fact, and it exposes the delegability of the access gate rather than any inaccuracy in the matching. Two principals of MediaMuv L.L.C. were indicted on thirty counts in the District of Arizona on 16 November 2021 for conspiracy, wire fraud, money laundering, and aggravated identity theft arising from false Content ID ownership claims; both pleaded guilty, one in April 2022 and the other in February 2023, and one was sentenced in June 2023 to 70 months in prison. On the charging record they falsely claimed ownership of over 50,000 Latin music recordings and monetized them through Content ID via a third-party rights administrator, obtaining $20,776,517.31 by the indictment's count and approximately $23.4 million by the plea. The indictment describes the method: staff found unmonetized music on the platform, downloaded and re-uploaded it, and asserted ownership through the content management system, presenting the administrator with a contract stating they were the 'writer, author, publisher, copyright holder and creator' of the catalogue, backed by forged letters from artists — assertions the administrator accepted without ownership verification. Individual artist losses recorded on the charging record run to $132,702, $128,339 and $102,626, and the scheme ran roughly four years before it was stopped. THE STRUCTURAL POINT: an approved Content ID partner can present claims for a catalogue the platform never assessed, so the eligibility gate is delegable and the vetting failure in this case was at the intermediary as much as at the platform. THE BOUNDARY: this is fraud against RIGHTS-HOLDERS committed through the claiming tools, not over-claiming against uploaders, and it is not evidence about the comparison's accuracy. The individual defendants are not named here and the third-party rights administrator is identified in the charging documents by initials only and is not named or guessed.

    empirical
    • Government U.S. Department of Justice, District of Arizona (2021, 2023). MediaMuv Principals Charged with Stealing Millions in Music Royalties, and MediaMuv L.L.C. Principal Guilty of Stealing Millions in Music Royalties Sentenced to 70 Months https://www.justice.gov/usao-az/pr/mediamuv-llc-principal-guilty-stealing-millions-music-royalties-sentenced-70-months

  • No court and no regulator has found Content ID unlawful, ordered it changed, or sanctioned it anywhere as of 28 August 2026, and the one United States case that attacked its access structure produced no finding of any kind. Schneider et al. v. YouTube, LLC (N.D. Cal. 3:20-cv-04423) alleged that Content ID was reserved for powerful copyright owners and unavailable to ordinary creators. Class certification was DENIED on 22 May 2023 on the ground that classwide copyright ownership 'will entail individualized proof that precludes certification', with the court adding that 'the takedown of content in response to a DMCA notice is miles away from substantive proof of copyright ownership or infringement'. On 12 June 2023 — the day trial was scheduled to begin, and after YouTube's 25 May withdrawal of its safe-harbour defence — the parties stipulated to dismissal WITH PREJUDICE of all claims raised or that could have been raised. There was no trial and no verdict. The eligibility complaint is nonetheless on the federal record: the U.S. Copyright Office's 2020 Section 512 Report describes Content ID as a voluntary filtering system beyond section 512's requirements, quotes the participation criterion, and reproduces commenters objecting that it 'unfairly excludes smaller copyright owners', that 'every artist should be entitled to this service', and — from the party who would later sue — 'basically, that means the little guy need not apply. That's wrong.' The same report records the OPPOSITE complaint from rights-holders that Content ID misses a significant share of unauthorized uploads, one commenter reporting a contractor identifying 1,488,035 infringing copies since December 2012 that Content ID had not caught, and user-advocacy comments that the system is 'prone to false positives and cannot properly take fair use considerations into account'. Both error directions sit on the same federal record from opposing parties. Separately, in YouTube, LLC v. Christopher L. Brady (D. Neb. 8:19-cv-00353, filed 19 August 2019) YouTube itself brought an action under 17 U.S.C. 512(f), alleging the defendant sent dozens of false takedown notices and threatened to trigger a third strike, terminating channels, unless creators paid him; the case settled in October 2019 without adjudication. Those are allegations attributed to YouTube as the pleading party, and they concern the public webform and strike channel rather than Content ID, though they terminate in the same strike ledger the Content ID appeal ladder feeds. THE STRUCTURAL CHARACTERISATION, attributed: Maayan Perel and Niva Elkin-Koren wrote in the Stanford Technology Law Review in 2016 that Content ID welds ex ante algorithmic blocking onto DMCA-style ex post removal and 'has turned algorithmic copyright enforcement into a private-financial model' protecting owners 'beyond the basic removal process provided by the DMCA', proposing transparency, due process, and public oversight as the accountability frame.

    empirical
    • Academic Goldman, E. (2023). Class Certification Denied Over Rightsowners' Demands for Content ID Access, Schneider v. YouTube (Technology & Marketing Law Blog); with TorrentFreak (2023). No Trial Today or Ever: YouTube Content ID Lawsuit Dismissed at 11th Hour https://blog.ericgoldman.org/archives/2023/05/class-certification-denied-over-rightsowners-demands-for-content-id-access-schneider-v-youtube.htm
    • Government U.S. Copyright Office (2020, May). Section 512 of Title 17: A Report of the Register of Copyrights https://www.copyright.gov/policy/section512/section-512-full-report.pdf
    • Government Complaint for Violation of the Digital Millennium Copyright Act, 17 U.S.C. 512(f), YouTube, LLC v. Christopher L. Brady, No. 8:19-cv-00353 (D. Neb., filed 19 August 2019), settled October 2019 without adjudication (a pleading, attributed to the operator as the pleading party; the webform and strike channel, not Content ID) https://torrentfreak.com/images/Youtube-v-Christopher-Brady-DMCA-abuse-complaint-191908.pdf
    • Academic Perel, M., & Elkin-Koren, N. (2016). Accountability in Algorithmic Copyright Enforcement. 19 Stanford Technology Law Review 473 https://law.stanford.edu/wp-content/uploads/2016/10/Accountability-in-Algorithmic-Copyright-Enforcement.pdf
    • Vendor YouTube / Google (2021-2023). Copyright Transparency Report, H1 2021, H2 2021, H1 2022 and H2 2022 (the four machine-readable PDF editions; the biannual series ends with H2 2022) https://storage.googleapis.com/transparencyreport/report-downloads/pdf-report-22_2022-7-1_2022-12-31_en_v1.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

Documented case histories