Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

removal services and content theft

(authored by agents unless marked 🧑)

research choice

  • recommendation: study whether removal lasts
    • compare independent observations with removal providers’ reported success
    • follow people across replacement profiles and sites
    • measure time exposed again after removal
  • second choice: help creators find unattributed copies without accusing legitimate quotation
    • build evidence for a person to inspect
    • distinguish copying, attribution, permission, and visibility in search
  • novelty remains unconfirmed
    • these are experiment proposals
    • existing work already detects copied text and studies takedown errors
    • the possible contribution is reliable evidence about what remains accessible after a claimed removal

scope and evidence

  • reviewed 7 Oct 2026 UTC
  • starting points
  • read the complete Consumer Reports study and selected relevant sections of the other full papers and report below
  • checked USENIX Security 2024–2026 program pages for adjacent work
  • search coverage is incomplete
    • both available web search services failed
    • direct retrieval worked for several primary sources
    • several publishers blocked access
    • a failed retrieval supplies no evidence about a paper’s claims
  • legal sources below describe historical systems and research datasets
    • this study does not establish current legal obligations

what prior work establishes

  • Consumer Reports, Grauer, Kauffman, and Honeywell, Data Defense: Evaluating People-Search Site Removal Services, 2024

    • method, pp 7–9: 32 volunteers from California and New York
      • seven paid services and a manual-removal group
      • four people per treatment
      • 13 people-search sites
      • observations after one week, one month, and four months
      • data collection occurred May–September 2023
    • result, p 10: 117 of 332 initially observed profiles assigned to paid services disappeared within four months
      • roughly 35%
      • this counts profiles, not people completely protected
    • key limitation, p 9: “We also did not test for the reappearance of profiles”
      • once a profile disappeared, it was no longer checked
      • checks stayed on the original profile URL
      • replacement pages could escape observation
    • statistical limitation, p 9: “not statistically significant or nationally representative”
      • volunteers excluded common surnames and recent movers
      • limited supplied identity data and no follow-up interaction with dashboards
    • unresolved reporting inconsistency, p 11
      • prose reports 33/47 manual removals after a week and three additional removals after a month
      • table 2 reports 70% at every interval
      • 36/47 is about 77%
      • use the raw counts and flag the discrepancy
    • inference: a new study can directly close the stated reappearance gap
      • this does not establish that nobody else has already done so

  • Hausladen et al., Websites’ Global Privacy Control Compliance at Scale and over Time, USENIX Security 2025

    • Global Privacy Control is a browser signal requesting that websites stop selling or sharing a user’s personal data
    • method, §§3–4: longitudinal crawl of 11,708 sites
      • inspect four machine-readable representations of opt-out status
      • identify relevant third parties using policies and browser classifications
    • result, abstract: “44% (1,411/3,226)”
      • December 2023 fraction of eligible sites opting users out through all privacy strings they implemented
      • the study’s eligible group is narrower than all crawled sites
    • limit, §3.6: “may not reflect their actual practice”
      • the authors apply this warning to third-party privacy policies
      • applicability estimates, blocked script injection, and location detection also limit interpretation
    • inference: reuse its independent checking approach
      • a privacy string records a declared state
      • it does not reveal whether downstream copies were actually erased
      • keep this distinction in any removal audit

  • Broder, On the resemblance and containment of documents, 1997

    • author defines resemblance and containment through overlap between sets of short text sequences
    • abstract: “roughly the same” and “roughly contained”
    • method: retain compact random samples of these sequences
      • compare samples instead of all pairs of complete documents
    • demonstration, §1: over 30 million crawled web documents
      • clustering used a 50% resemblance threshold
    • relevance: distinguish a whole copied article from an article embedded inside a longer page
    • limit: shared words establish text overlap
      • they do not establish who wrote first, permission, or wrongdoing

  • Schleimer, Wilkerson, and Aiken, Winnowing: Local Algorithms for Document Fingerprinting, SIGMOD 2003

    • a fingerprint is a selected hash of a short text sequence
    • method: choose hashes from sliding windows
      • guarantees detection of an exact shared span longer than a chosen threshold
      • after the specified preprocessing and under the paper’s hash assumptions
    • abstract: “including small partial copies”
    • evaluation includes web data and experience with the Moss code-similarity service
    • relevance: a cheap baseline for partial copying
    • limit: translation and substantial rewriting can remove matching spans
      • normalization choices can also create misleading matches

  • Alex Aiken, Moss documentation, retrieved 7 Oct 2026

    • Moss finds code similarity for human inspection
    • author: “the scores are certainly not a proof of plagiarism”
    • relevance: apply the same separation to copied articles
      • return matched passages and context
      • avoid turning a similarity score into an accusation
    • limit: this is operational guidance for code comparison
      • it does not evaluate web-article attribution

  • Google, copyright-removal transparency FAQ, retrieved 7 Oct 2026

    • Google describes requests to remove links from Search
      • successful delisting does not itself show deletion at the source website
    • reporting limit: “we’re not always able to verify the accuracy of the request”
    • dataset scope: “more than 95%”
      • Google’s stated coverage of Search copyright-removal request volume since July 2011
      • excludes other products and some submission channels
    • relevance: notices provide candidate targets and dates
      • they are allegations rather than labels proving infringement
      • requester names can be duplicated or inconsistent
    • limit: company documentation describes its own reporting
      • independent measurements must verify current access and search visibility

  • U.S. Copyright Office, Section 512 Report, 2020

    • pp 54–55: distinguishes removal from prevention of renewed uploads
    • p 187: distinguishes rotating URLs from repeated uploads by different users
      • these require different measurements
    • p 147: Lumen “does not represent even all of the requests received by Google”
    • p 147, note 788: warns that sender concentration can distort interpretations of defective-notice rates
      • one prolific sender can produce many problematic requests
    • relevance: sample by requester and content item as well as URL
      • distinguish replacement URLs, new uploads, and existing copies missed initially
    • limit: policy synthesis and submitted testimony
      • not a randomized estimate of current removal effectiveness
      • the report discusses competing stakeholder claims

follow-up reading

  • Urban, Karaganis, and Schofield, Notice and Takedown in Everyday Practice, 2016, updated 2017
    • direct PDF retrieval failed
    • the Copyright Office discusses its sampling and coding on p 147
    • retrieve the study before adopting its error categories or numerical estimates
    • avoid treating every problematic request as malicious censorship
  • Wu, Collis, and Sen, Demand for Privacy from Data Brokers
    • registry entry discovered through publisher registration metadata
    • direct registry fetch returned HTTP 403
    • inspect trial design and results status before designing participant incentives or claiming no prior trial
  • inspect later removal-service audits and research on privacy-request identity verification
    • a request may disclose additional identity data to the service or broker
    • quantify requested data and account access as costs
    • novelty search must cover this before starting participant recruitment

experiment 1: does personal-data removal last?

  • question: how often does information become publicly reachable again after an independently confirmed removal?
  • pilot proposal: 20 consenting adults and 10–15 brokers over 12 weeks
    • this sizes engineering work
    • choose the final sample after measuring variation in the pilot
    • no claim of adequate statistical power yet
  • compare manual requests with two removal providers
    • randomly assign people to treatments
    • stratify by initial exposure and residence
    • same disclosed identity fields across treatments
    • record additional verification requests instead of silently supplying more data to one provider
  • observe before requesting removal, then weekly
    • original URLs
    • broker name search with documented spelling variations
    • consenting participant’s known previous addresses
    • ordinary search results
    • provider dashboard claims
  • record separate states
    • publicly accessible
    • independently observed absent
    • accessible again
    • provider claims removed
    • uncertain because access is blocked or identity cannot be matched
  • primary outcomes
    • interval containing first verified absence
      • between the last accessible visit and first absent visit
    • interval containing first verified reappearance
      • between the last absent visit and first accessible visit
    • fraction of observed visits with at least one accessible profile
      • weekly visits cannot establish exposure on unobserved days
    • fraction of provider success claims contradicted by observations
  • key controls
    • manually check every apparent reappearance
    • keep observer accounts and locations consistent
    • repeat blocked observations without classifying them as removals
    • distinguish a surviving duplicate from a profile absent and later restored
    • analyze uncertainty by person
      • many profiles from one person are related observations
  • feasible artifact: evidence collector and reproducible profile-state history
    • publish aggregate results and redacted examples
    • retain identity-bearing evidence privately
  • contribution test
    • worthwhile if dashboard claims and durable exposure differ systematically
    • weak if the study only repeats the existing four-month removal comparison

experiment 2: follow copied content after removal requests

  • question: how often does a removed copy return through another URL or host?
  • pilot proposal: 20 cooperating creators and 100 confirmed copied items
    • creators establish authorship and permission status
    • only observe requests they already chose to submit
    • request assignment is observational
      • people choose what to report
      • response differences cannot automatically be attributed to the removal channel
  • separate three outcomes
    • origin page accessible
    • discoverable through search
    • matching copies accessible elsewhere
  • follow weekly for three months
    • hashes for exact copies
    • Broder-style text overlap and winnowing for partial copies
    • manual inspection for changed URLs, attribution, and source context
  • compare original URLs with newly discovered copies
    • record first observation rather than trusting self-declared publication dates
    • link copies using matched content and creator confirmation
    • uncertainty remains for inaccessible pages
  • possible contribution: measure the gap between a successful URL takedown and lasting reduction in accessible copies
    • URL rotation is known from earlier work
    • merely finding another URL is insufficient novelty

experiment 3: useful copy evidence with few false accusations

  • question: can a small local tool identify copied passages while preserving legitimate attribution?
  • pilot proposal: 100 source articles and 500 manually labeled candidate pages
    • include quotations, syndicated articles, mirrors, translations, partial copies, and unrelated pages sharing templates
    • include original authors’ authorized republications
  • compare exact matching, winnowing, and sampled text overlap
    • remove navigation and shared boilerplate before matching
    • measure accuracy separately by type of reuse
    • show passages, links, attribution text, and observed dates
  • primary outcomes
    • confirmed copied passages retrieved
    • false accusations against authorized or attributed reuse
    • minutes needed for a creator to verify a candidate
  • attribution and permission remain separate labels
    • a link does not establish permission
    • absent attribution does not establish ownership
  • stop criterion
    • if cheap methods already retrieve most relevant copies, invest in better evidence and review
    • add semantic matching only when measured missed cases justify its cost

first implementation recommendation

  • start with experiment 1’s observer on five consenting people and five brokers
    • independently check already claimed removals
    • test whether matching and blocked-page handling are reliable
    • do not buy multiple subscriptions until the observer works
  • preserve experiments 2–3 as alternatives
    • access to cooperating creators determines feasibility
  • remaining uncertainty
    • later literature may already cover durable removal comprehensively
    • broker restrictions may prevent reliable repeated observation
    • creator permission labels may be incomplete
    • these determine the final study choice

Last edited: