{"id":"2684fba2-b539-4de8-88be-7b9bc158e4d1","arxiv_id":"2611.17012","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The authors demonstrate reidentification, record linkage, and sensitive-attribute inference on Google Search data anonymized under the European Commission's DMA Article 6(11) measures, and conclude the measures satisfy none of the EDPB's three anonymization criteria.","lead":"The authors attacked six months of Europe-wide Google Search data that Alphabet had anonymized under the European Commission's new DMA rules, and report successful reidentification, record linkage, and disclosure of sensitive personal details. Their conclusion: the Commission's thresholds are arbitrary heuristics that fail the EDPB's three anonymization criteria and should not become a regulatory benchmark.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Data-authenticity and verification assumption is load-bearing and unverifiable; the paper's empirical core depends on a corpus provided by a party with aligned interests, and no independent verification path is offered.","rationale":"The reader's weakest_assumption is precisely the data-authenticity and verification assumption. I agree that this is the most load-bearing concern. The paper's empirical core is unverifiable, and the source of the data creates a conflict of interest that is disclosed but not discussed. The paper does not provide a verification path. The central claim is supported by multiple independent attack patterns and concrete examples, but the evidential basis for the quantitative headline is the corpus. If the corpus is not what it claims to be, the quantitative claims and the EDPB-criteria conclusion weaken. The paper's qualitative critique of the EC's measures (heuristic, arbitrary, lacking formal guarantees) is independent of the corpus and stands on its own. The verdict should remain CONDITIONAL, as the reader recommended, because the paper should provide a verification path. I do not see a reason to change the verdict to REJECT or ACCEPT. The paper's qualitative arguments are strong, but the empirical core needs verification. The agreement_with_reader is 'agree' because the reader's weakest_assumption is the same as mine.","tokens_in":27871,"tokens_out":738,"duration_ms":8067,"concrete_test":"Independently audit the corpus: have a neutral third party (e.g., a university ethics board or a regulatory body) verify that the corpus is exhaustive, six-month Google Search data from the claimed EEA countries, with the EC's technical measures applied exactly as in DMA.100209. The auditor should recompute Table 1 (unique queries and records) and replicate at least one attack pattern (e.g., the 'my name' pattern on the Spanish data) on a random sample of the corpus. If the auditor's results differ materially from the paper's, the central claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central empirical claim—that the EC's measures fail all three EDPB anonymization criteria—rests on a single corpus: six months of Google Search data from several EEA countries, provided by Alphabet and processed under the EC's technical measures. The authors state that this is 'precisely the same data that competing search engines would have access to' (Section 5, para 4). However, the paper ships no data, no code, and fully redacted examples. The data and funding come from Alphabet (acknowledgments), a party whose rejected k-anonymity proposal would be vindicated if the EC's measures were found broken. The reader's weakest_assumption correctly identifies this as a data-authenticity and verification assumption. The paper does not address this conflict beyond a disclaimer that the views do not necessarily reflect Alphabet's. The quantitative headline (44% unique records, 10-33% attack success rates) and the EDPB-criteria conclusion would weaken if the corpus was non-representative, partially pre-filtered, or if the manual reidentification matches were less reliable than described. This is not an ad hominem; it is a structural concern about the evidential basis for the central claim. The paper acknowledges no limitation on this point; it presents the corpus as given and the attacks as straightforward. The concern is that the empirical core is unverifiable and the source of the data creates an undisclosed risk of bias.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper critically evaluates the European Commission's anonymization measures for Google Search data under Article 6(11) of the DMA, arguing that they fail to meet EDPB anonymization criteria. The authors present empirical evidence from a six-month corpus of Google Search data provided by Alphabet, showing high rates of unique records (44% with timestamps) and demonstrating reidentification, linkage, and attribute inference through four automated attack patterns. The paper concludes that the EC's measures are heuristic, lack formal guarantees, and do not satisfy the 'no record isolation', 'no linkage', or 'no inference' criteria.","tokens_in":28139,"tokens_out":1731,"duration_ms":12942,"significance":"If correct, the paper has substantial regulatory significance: it would demonstrate that the EC's first operationalization of anonymization under the DMA is fundamentally flawed, potentially affecting data-sharing obligations for gatekeepers and the interpretation of anonymization under GDPR. The paper provides concrete attack patterns and quantitative uniqueness measures that could inform future regulatory standards. The empirical grounding in real data (though provided by Alphabet) is a strength, as is the systematic mapping to EDPB criteria.","major_comments":[{"comment":"The paper's central empirical claims rest entirely on a corpus provided by Alphabet, a party with a direct financial interest in the outcome (its rejected k-anonymity proposal would be vindicated if the EC's measures are found inadequate). The authors acknowledge this funding and data provision in the Acknowledgments but do not discuss how this might affect the reliability or representativeness of the data. The paper ships no data, no code, and fully redacted examples, making independent verification impossible. This is a structural limitation that should be explicitly addressed as a potential source of bias, not merely disclosed. The authors should discuss what steps were taken to ensure the corpus is indeed 'the same data that competing search engines would have access to' (Section 5, para 4) and whether any pre-filtering or selection occurred.","section":"§5, §6, Acknowledgments","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table 1 reports unique search records with timestamps at 44% for the six-month Ireland corpus. However, the paper does not clarify whether these uniqueness counts are based on the full query text plus all metadata fields, or only on the fields that would be visible to data recipients under the EC's measures. The EC measures remove user identifiers and precise timestamps but retain day-level timestamps, location (NUTS3), device type, and inferred language. The paper should explicitly state which fields are included in the 'unique search records with timestamp' calculation and confirm that these match the fields available to recipients. If additional fields (e.g., query text variations) are included, the uniqueness rate may be inflated relative to what an attacker would actually observe.","section":"§5.2, Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The reidentification success rate of 'more than 10%' (Section 6.1) and '33%' (for the numerology subset) are reported based on manual examination of the top 100 records ranked by an LLM. The paper does not report the total number of records examined for the 10% figure, nor does it provide confidence intervals or discuss potential selection bias from the LLM ranking. The attack pattern relies on Gemini-2.5-flash with a specific prompt; the reproducibility of these results across different LLMs or prompt variations is unclear. The authors should provide more detail on the ranking methodology and report results on a random sample rather than only the top-ranked records.","section":"§6.1, Attack pattern 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The allowlist-injection attack (Section 5.7) is described as requiring only 13 phone numbers to create 50 accounts. The paper does not address whether Google's anti-abuse systems (e.g., phone verification requirements, rate limiting, device fingerprinting) would detect or prevent such an attack at scale. The claim that 'the attacker only needs to issue a few dozen realistically-looking queries for each target, either manually or through straightforward automation that mimics human search behavior' is asserted without evidence. The authors should either provide a proof-of-concept demonstration or discuss the practical obstacles more thoroughly.","section":"§5.7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The lognormal extrapolation of query length distributions to 2027-2028 assumes that the observed trend in mean and standard deviation from March-August 2026 continues linearly. This is a strong assumption, especially given the short observation window (6 months) and the potential for seasonal effects or changes in user behavior (e.g., AI mode adoption). The paper acknowledges this as 'conservative' but does not provide sensitivity analysis or alternative scenarios. The extrapolation should be caveated more explicitly, and the sensitivity of the utility-loss projections to the assumed trend should be discussed.","section":"§5.4, Fig. 1","authors_acknowledge":"To assess the future effect of a growing average length on the proportion of sup- pressed queries, let us make the conservative assumption that the standard deviation stays unchanged at 16.71.","authors_acknowledge_p":0.6}],"minor_comments":[{"comment":"The paper states 'Section 5 details attack patterns and examples of reidentification and disclosure' but Section 5 is the critical review; attack patterns are in Section 6. This appears to be a typo (should reference Section 6).","section":"§1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table 1 reports counts for Ireland with 4.87 million search records in the six-month window, but the table shows 4,871,552,575 for six months. This appears to be a factor of 1000 error (4.87 billion vs 4.87 million). The text says '4.87 million search records' but the table shows billions. Please clarify and correct.","section":"§5.2, Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The notation uses μ and σ for both the lognormal distribution parameters and the underlying normal distribution parameters, which is confusing. Equations (2) and (3) define μ' and σ' for the normal distribution, but the text refers to μ and σ as the mean and standard deviation of the lognormal. Clarify the notation.","section":"§5.4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper reports that Gemini processing costs were ~$1 for 10,000 queries and total cost <$50. This is useful for demonstrating low attack cost, but the dollar amounts should be contextualized (e.g., as of what date, using what pricing tier).","section":"§6.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The Eircode attack pattern reports 'approximately 304,585 groups, most of which contained a single query.' The paper should clarify whether these groups are based on unique Eircodes or unique query-Eircode combinations, and how false positives from product codes were handled beyond the group-size filter.","section":"§6.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The conclusion that the EC's measures fail all three EDPB criteria is well-supported by the evidence presented, but the paper could strengthen this by explicitly mapping each attack pattern to the specific criterion it violates, rather than discussing them collectively.","section":"§7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Reference [2] (Batet et al., 2023) appears to be a journal article from 2023 but the volume/page numbers suggest it may be from a different year. Please verify the citation details.","section":"References","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Colleague,\n\nThis paper does two things well. First, it gives the first public empirical attack on the actual corpus that Alphabet would share under the EC's DMA decision. The four attack patterns (self-queries, email addresses, Eircode directions, minority-language linkage) are concrete, automatable, and produce dozens of reidentifications with special-category disclosures. Second, the conceptual critique is correct: the EC's component-wise suppression and metadata generalization ignore joint uniqueness, which is a textbook error. The 44% unique-record figure (Table 1) is striking if it holds.\n\nThe soft spots are real but not fatal to the qualitative claim. The biggest is the data-authenticity assumption. The corpus comes from Alphabet, whose rejected k-anonymity proposal would be vindicated if the EC's measures were broken. The paper ships no data, no code, and fully redacted examples. The authors disclose the funding but do not discuss the conflict or offer a verification path. This does not make the paper wrong, but it means the quantitative headline (44% unique, 10-33% attack success) rests on a single unverifiable source. A second concern: the attack success rates (10% and 33%) are computed on LLM-ranked top-100 and top-15 selections, so they illustrate potential rather than expected yield. The abstract's 'at scale' framing drops that qualification. The utility projections in Section 5.4 are fitted lognormal extrapolations presented as forecasts, which is fine for illustration but not evidence.\n\nWho this is for: regulators, privacy researchers, and anyone following the DMA-GDPR interplay. The paper deserves a serious referee because the qualitative claim (the EC's measures permit reidentification, linkage, and inference) is supported by multiple independent attack patterns and concrete examples, even if the exact rates are uncertain. The authors should be asked to provide an audited computation of Table 1, redacted query samples, code for the regex and LLM pipelines, and an explicit discussion of the Alphabet relationship.\n\nRecommendation: send to peer review with a request for a verification appendix.","headline":"The empirical core is real and the critique is sound, but the paper's load-bearing assumption—that the corpus is exactly what the EC would release—is unverifiable and funded by a party with aligned interests.","tokens_in":28843,"tokens_out":535,"duration_ms":6225},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"EC's DMA anonymization measures for Google Search data fail to meet GDPR standards, study finds","keywords":["anonymization","Digital Markets Act","Google Search data","reidentification","GDPR","k-anonymity","EDPB guidelines","query data privacy"],"falsifier":"If Alphabet or the EC released the exact corpus with a certified statement that the paper's attack patterns do not yield the reported reidentifications (e.g., because the allowlist was applied differently, or because the data were pre-filtered before delivery to the authors), the paper's quantitative headline (44% unique records, 10–33% attack success rates) and its conclusion that the measures fail the EDPB criteria would be undermined.","tokens_in":29192,"feed_emoji":"🔍","tokens_out":1546,"duration_ms":13237,"temperature":0.7,"pith_summary":"The paper argues that the European Commission's anonymization measures for Google Search data under the Digital Markets Act are fundamentally flawed and do not render the data anonymous. The authors demonstrate that the measures fail each of the three EDPB anonymization criteria—no record isolation, no linkage, and no inference—and present real-world examples of low-effort reidentification and attribute disclosure. If the paper is correct, the data Alphabet would share remains personal data under GDPR, and the EC's decision does not satisfy the DMA's anonymization requirement.","feed_headline":"EC's Google Search anonymization fails GDPR criteria, study finds","feed_subtitle":"44% of records remain unique; low-cost attacks reidentify dozens of individuals from 'anonymized' data.","key_machinery":"The paper's central object is the corpus of real Google Search data from several EEA countries, spanning six months, with the EC's technical measures applied exactly as in decision DMA.100209. The authors use this corpus to measure the proportion of unique queries and search records, and to execute four attack patterns—self-query search, malformed email extraction, Eircode-based direction grouping, and minority-language linkage—that demonstrate reidentification, record linkage, and attribute inference at scale. The key identity is the percentage of unique search records (approximately 44% with day-level timestamps) as a direct measure of record isolation failure.","core_discovery":"The EC's measures—query suppression based on entity frequency and length thresholds, metadata generalization, and contractual safeguards—cannot prevent reidentification, record linkage, or attribute inference. Using the same six-month Google Search data from Ireland and Spain that would be shared under the decision, the authors report that about 44% of search records remain unique when day-level timestamps are included. They demonstrate dozens of successful reidentifications through four automatable attack patterns: self-queries (e.g., 'my name'), malformed email addresses, Eircodes in directions queries, and minority-language concentration. The paper concludes that the EC's approach relies on unjustified heuristics rather than formal privacy guarantees, underestimates the risks of free-text queries, and sacrifices utility without achieving meaningful protection.","pith_inferences":["The paper implicitly challenges the EC's reliance on an undisclosed expert report to define 'residual risk.' If the standard for anonymization is to be set by regulation, the risk threshold must be publicly specified and quantitatively justified, not left to an unnamed expert whose report is not released.","The allowlist-injection attack (Section 5.7) suggests that the EC's measures create a new vulnerability: an attacker can deliberately inflate the frequency of identifying information to make it pass the 50-user threshold, then monitor for queries containing that information. This is a structural weakness of frequency-based suppression, not just a parameter-tuning issue.","The paper's analysis of AI-mode queries (Section 5.4) implies that the EC's length-based threshold will become increasingly ineffective (or increasingly utility-destructive) as AI-generated prompts grow in volume and length, a trend the EC's fixed-threshold design does not account for."],"forward_implications":["If the paper is correct, Alphabet would be required to share data under Article 6(11) that remains personal data under GDPR Recital 26, creating a direct conflict between the DMA's data-sharing mandate and the GDPR's anonymization requirements.","The EC's approach—particularly the entity-based allowlist, length-based suppression, and independent metadata generalization—cannot serve as a regulatory benchmark for anonymization under the EU digital rulebook.","Contractual prohibitions on reidentification cannot compensate for technical measures that fail to prevent record isolation, linkage, and inference, as the EDPB guidelines explicitly state.","The findings suggest that only formal privacy models such as k-anonymity or differential privacy, applied to the entire search record (query text plus metadata), can provide the ex ante guarantees required by the GDPR.","The paper's attack patterns—especially the allowlist-injection attack (Section 5.7) and the LLM-assisted self-query search—are automatable and low-cost, meaning the residual risk the EC considers acceptable may translate into thousands of reidentifiable individuals given the near-exhaustive coverage of the EEA population."],"supporting_citations":[{"why":"Provides the historical precedent of the AOL search data scandal, which the paper uses to contextualize the seriousness of its own reidentification findings.","marker":"[1]"},{"why":"Defines the three EDPB anonymization criteria (no record isolation, no linkage, no inference) against which the paper evaluates the EC's measures.","marker":"[11]"},{"why":"The EC's decision DMA.100209 that contains the anonymization measures under review; the paper's entire analysis targets the measures described in this document.","marker":"[9]"},{"why":"The GDPR, particularly Recital 26, which defines identifiability and is the legal standard the paper argues the EC's measures fail to meet.","marker":"[12]"},{"why":"The DMA, specifically Article 6(11), which mandates anonymization of personal data before sharing; the paper argues the EC's measures do not satisfy this requirement.","marker":"[13]"},{"why":"Introduces k-anonymity, the privacy model that the EC rejected but that the paper argues would provide the formal guarantees the EC's heuristics lack.","marker":"[23]"},{"why":"Supplies the conceptual distinction between anonymization and de-identification, which the paper uses to argue that the EC's measures amount to weak de-identification, not anonymization.","marker":"[25]"}],"fun_headline_variants":["EU's Google data anonymization fails real-world reidentification tests","DMA anonymization: 44% of records unique, reidentification trivial","Study: EC's anonymization can't meet EDPB criteria, reidentification shown","Google Search data 'anonymized' by EU still reveals identities, study finds"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The empirical core assumes the corpus is exactly the data that competing search engines would receive under the EC decision—exhaustive, six-month Google Search data from several EEA countries, with the EC's technical measures applied precisely as mandated in DMA.100209—so that the attacked records are representative of what an actual recipient would hold.","fun_headline_variants_meta":{"raw":{"variants":["EU's Google data anonymization fails real-world reidentification tests","DMA anonymization: 44% of records unique, reidentification trivial","Study: EC's anonymization can't meet EDPB criteria, reidentification shown","Google Search data 'anonymized' by EU still reveals identities, study finds"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000144,"raw_usage":{"cost":0.000143948,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000143948,"upstream_inference_prompt_cost":0.000107548,"upstream_inference_completions_cost":3.64e-05},"total_tokens":1103,"prompt_tokens":973,"completion_tokens":130,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":46}},"tokens_in":973,"tokens_out":130,"cost_source":"openrouter_usage","duration_ms":1882,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":46,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T07:32:17.173047+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"If Alphabet or the EC released the exact corpus with a certified statement that the paper's attack patterns do not yield the reported reidentifications (e.g., because the allowlist was applied differently, or because the data were pre-filtered before delivery to the authors), the paper's quantitative headline (44% unique records, 10–33% attack success rates) and its conclusion that the measures fail the EDPB criteria would be undermined.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the historical precedent of the AOL search data scandal, which the paper uses to contextualize the seriousness of its own reidentification findings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the three EDPB anonymization criteria (no record isolation, no linkage, no inference) against which the paper evaluates the EC's measures."},{"cited_title":"https://digital-markets-act-cases.ec.europa.eu/cases/DMA.100209 (2026), accessed: 2026- 09-23","cited_arxiv_id":null,"evidence_quote":"The EC's decision DMA.100209 that contains the anonymization measures under review; the paper's entire analysis targets the measures described in this document."},{"cited_title":"https://data.europa.eu/eli/reg/2016/679/oj (2016)","cited_arxiv_id":null,"evidence_quote":"The GDPR, particularly Recital 26, which defines identifiability and is the legal standard the paper argues the EC's measures fail to meet."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The DMA, specifically Article 6(11), which mandates anonymization of personal data before sharing; the paper argues the EC's measures do not satisfy this requirement."},{"cited_title":"IEEE Transactions on Knowledge and Data Engineering13(6), 1010–1027 (2001)","cited_arxiv_id":null,"evidence_quote":"Introduces k-anonymity, the privacy model that the EC rejected but that the paper argues would provide the formal guarantees the EC's heuristics lack."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the conceptual distinction between anonymization and de-identification, which the paper uses to argue that the EC's measures amount to weak de-identification, not anonymization."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}