Problematic Paper Screener

register entry problematic-paper-screener-aio-20261029 OBS-9180ce10e2aa · file as first run sha256 c6915dc52a012e4e…

• T1-1 definition: 'an online platform and automated screening tool developed by Guillaume Cabanac, Cyril Labbé, and Alexander Magazinov to detect fraudulent, error-filled, or computer-generated scientific literature'
• T1-2 intake: 'PPS scans millions of scholarly papers using data APIs (such as Crossref, Dimensions, PubMed, and PubPeer) across several specialized detectors'
• T1-3 tortured phrases: 'Identifies nonsensical paraphrased terms (e.g., "counterfeit consciousness" instead of "artificial intelligence") used by paper mills to bypass plagiarism checkers.'
• T1-4 generators: 'Flags texts produced by automated paper generators like SCIgen and Mathgen.'
• T1-5 fingerprints: 'Detects unedited AI-generated text patterns (such as ChatGPT remnants).'
• T1-6 Feet of Clay: 'Tracks active or unretracted papers that improperly cite previously retracted works, expressions of concern, or withdrawn articles.'
• T1-7 method: 'It pairs automated machine-detection algorithms with human post-publication review.'
• T1-8 reporting: 'Flagged papers are often automatically shared or logged on platforms like PubPeer and Pepia to alert editors, publishers, and the research community.'
• T1-9 adoption: 'The tool has been integrated into editorial workflows, repository platforms (such as Episciences), and analytics frameworks (such as Signals Graph) to help scrub the scientific record of fabricated or corrupted studies.'
• T1-10 closer: an offer to 'explore how to use the tool to check specific references or learn more about how publishers handle tortured phrases'

• B1 Institut de Recherche en Informatique de Toulouse - IRIT — PPS – Problematic Paper Screener card; HTTP 200; B1.html sha256 17dc1cfc7a816494
• B2 The Conversation — Problematic Paper Screener: Trawling for fraud in the scientific literature card; HTTP 200; B2.html sha256 f0b03dc0ab03444b
• B3 arXiv — [2210.04895] The 'Problematic Paper Screener' automatically selects suspect publications for post-publication (re)assessment card; HTTP 200; B3.html sha256 1de09e3df08a0bb3
• B4 Research Signals — Problematic Paper Screener: a new signal to help protect research integrity – Signals card; HTTP 200; B4.html sha256 8d8604d8a654149a
• B5 ccsd.cnrs.fr — Episciences Adopts Problematic Paper Screener as a Tool for Verifying ... - CCSD card; HTTP 200; B5.html sha256 8634979de90d1f07
• B6 YouTube·Crossref — Exploring Retraction Watch Data in Crossref 2025-05-09 card (video); oEmbed HTTP 200, B6.json sha256 133646a325a17123; admitted by title
• B7 Université Grenoble Alpes — The Conversation : "Problematic Paper Screener: Trawling for fraud in the scientific literature" own page; unreachable (HTTP 302 to a CAS login, then 403; three tries); B7.html is the error page, sha256 089de94bfd1cbc8c
• W1 Wikipedia — no article (Problematic Paper Screener; Tortured phrase a soft redirect to Wiktionary) HTTP 404 (raw and rendered); search 0 hits; W1.html sha256 3a8cb117470f4f6e

E_A, the archive's entity: the same screener, in the MMRS meta-aggregator's record: a monitor run by Cabanac with Labbé and Magazinov since 27 February 2021, a set of detectors scanning the literature weekly, whose unit is a published paper flagged by a detector, which records attribution partly through flagged references, and whose selection finds what leaves a fingerprint, clean machine text and misstated real sources passing beneath it; the scholarly record's automated immune response beside Academ-AI, the two recording machine composition entering the record unattributed; observed by the Capture Registry at its name, composed at its makers' grain; and, at its class, paper-mill output read by #741 as proxy gaming hard to detect under the proxy regime MMRS-01 MMRS-15 MMRS-02 MMRS-16 MMRS-20 MMRS-07 MMRS-08 MMRS-09 MMRS-10 CAP-1-06 O741-01 O741-02; defining papers #741

E_F, the field's entity (type C — a conventional reading holds it (public entity: a monitor of the scholarly literature)): an online platform established 27 February 2021 by Guillaume Cabanac, Cyril Labbé and Alexander Magazinov that screens the scientific literature for papers partly generated with tortured phrases, SCIgen or Mathgen, for ChatGPT fingerprints and for bibliographies citing retracted, concerning or flagged references, harvesting Crossref with the Retraction Watch Database, Dimensions, PubMed and PubPeer, 130 million papers a week, by machine detection with human assessment; problematic papers reported on PubPeer; used by publishers, the Signals Graph and Episciences (B1–B6)

ρ = extends: the archive reads E_F as the entity and adds senses and limits on the field's own object: its unit and kind, what its selection leaves out (what leaves no fingerprint), its place in the field as a record of machine composition entering the scholarly record unattributed, attribution recorded partly, the record kept apart from the entity, the composition layer's observed treatment of it, and a reading of the class it works on under the proxy regime MMRS-16 MMRS-15 MMRS-07 MMRS-08 MMRS-09 MMRS-10 MMRS-20 MMRS-23 CAP-1-06 O741-01 O741-02. Guard: ρ = extends claims an addition and no more: MMRS-07–MMRS-10 are the record's readings, not findings that the screener is defective; O741-01–O741-03 bear on paper-mill output and its detection as a class, not on the screener, which #741 does not name; where an archive claim meets a field claim (MMRS-18/MMRS-26 beside F21's date; CAP-1-03 beside F10) both are composed under P_coh in the one claim space

The case: equal terms, ontology included (v0.9). The MMRS record holds the field's entity as the entity: it describes the Problematic Paper Screener from its own page and its makers' article, marking each field's basis (operator, here throughout), and composes nothing that denies a field presupposition; it restates the field's detectors, sources, scale, findings and retractions as one lineage with them. What it adds is on the field's object: the unit and the kind, what the selection leaves out (a detector finds what leaves a fingerprint), the place in the field beside Academ-AI as a record of machine composition entering the scholarly record unattributed, attribution recorded partly, the record kept apart from the entity; the Capture Registry adds one observation of the composition layer at the entity's address; #741 adds a reading of the class (paper-mill output under the proxy regime). The ontology the archive uses is available in the field itself: the makers state that the fingerprinted papers are 'the tip of the tip of the iceberg' (F25) and that the challenge is chatbots that generate content (F26); the archive's selection reading extends that stated limit from one detector to the instrument. The law of the suppressed witness does not apply (§0.10: not where the field holds the ontology). Not mixed.

#MMRS articles.py ARTICLES['problematic-paper-screener'], kernel: The Problematic Paper Screener, run by Guillaume Cabanac with Cyril Labbé and Alexander Magazinov since 27 February 2021, scans the scientific literature weekly for papers generated or disguised by tools (tortured phrases, SCIgen, Mathgen) and for references to retracted work.

#MMRS articles.py, 'What its selection leaves out' s1–s2: A detector finds what leaves a fingerprint. Text machine-written cleanly, or cited sources that exist but say something else, pass beneath it.

#MMRS articles.py, 'Place in the field' s1: It is the scholarly record's automated immune response, and the neighbour of Academ-AI, which reads by phrase where the screener reads by detector.

#MMRS census.json, schema_note: each page has two nodes, the MMRS record and the entity it describes; relations are typed; each assertion carries a source and a basis (operator statement, third-party report, our observation, our reading); lifecycle is kept apart from the observation that saw it

#capture problematic-paper-screener-aio-20261029, analysis: The screener is composed at its makers' grain, with its human review stage stated and its uptake by publishers' infrastructure added.

#741 §8 ¶3: The proxy regime's defense against gaming is its inscrutability: gaming is hard to detect because the signals being gamed are themselves opaque.

#B2 findings list, item 4 (field): These articles represent the tip of the tip of the iceberg: They are cases where ChatGPT output has been copy-pasted wholesale into papers without any editing (or even reading) and has also slipped past peer reviewers and journal editors alike.

equal-terms case: every field claim composed as received, beside the archive claims it stands with. Reading the received column recovers the field (§5.9).

v0.9 §4.10: KO_B sentences surviving in the KO at similarity ≥ 0.95, with positions (computed as apply_v09.py does: difflib.SequenceMatcher ratio, the best KO sentence for each KO_B sentence): v0.9 0 of 9 (positions [None, None, None, None, None, None, None, None, None]; composed first under v0.9, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

The Problematic Paper Screener screens the scientific literature for papers partly generated with tortured phrases, SCIgen or Mathgen and for bibliographies citing retracted, concerning or flagged references, by its own page, which dates it from 27 February 2021. Its makers' paper of 7 October 2022, by Cabanac, Labbé and Magazinov and presented at WCRI 2022, calls it an online platform pairing machine detection with human assessment to flag published problematic articles, since post-publication assessment remains necessary, and a tool to curate the literature; Research Signals calls it an academic project led by Cabanac. Cabanac built it and administers it, using Digital Science and PubPeer metadata under no-cost agreements, in the account he and Labbé, computer scientists at Toulouse and Grenoble Alpes, gave The Conversation in January 2025; both the page and the article report its place in Nature's 10 of 2021. The page lists ten detectors and harvests Crossref with the Retraction Watch Database, Dimensions, PubMed and PubPeer, reporting problematic papers on PubPeer; the article counted nine, ChatGPT fingerprints among them, and a weekly trawl of 130 million papers for paper-mill output. Tortured phrases come from paraphrasing tools used to evade plagiarism detection, as Joined Together States for United States, and Nature's post on the page gives counterfeit consciousness for artificial intelligence. The article reports nearly 19,000 tortured-phrase papers, over 280 SCIgen papers, over 764,000 articles citing retracted work found by Feet of Clay, and over 70 with ChatGPT fingerprints, the last only the tip of the iceberg, with the open challenge of chatbots generating content. Publishers use it, some in the editorial workflow, it has helped toward more than 1,000 retractions, analytics firms have used it, and Cabanac consults pro bono for ClearSkies, Morressier, River Valley, Signals and STM. Signals added over 12,000 unretracted flagged publications to its graph in June 2024; Episciences checks references against its database of over a million references in July 2026, finding 226 of nearly 170,000, describing it also as a filter for errors, plagiarism, manipulation and conflicts of interest. The page was last updated 2026-10-28, and a Crossref webinar of 9 May 2025 on Retraction Watch data sits on the rail.

The Problematic Paper Screener (PPS) is an online platform, in its makers' paper, that combines automatic machine detection with human assessment to identify and flag already published problematic articles (arXiv:2210.04895); its own page says it screens the scientific literature for papers (partly) generated with tortured phrases, SCIgen and Mathgen, and the MMRS record has it run by Guillaume Cabanac with Cyril Labbé and Alexander Magazinov since 27 February 2021, scanning the literature weekly for papers generated or disguised by tools and for references to retracted work, at the address dbrech.irit.fr. Its unit, in the MMRS record, is a published paper flagged by a detector; CCSD, announcing Episciences' adoption, describes what it provides as a database of over one million references identified as containing problematic elements, typed by severity from articles written entirely by AI to valid articles whose references carry anomalies ("Feet of clay"). The page dates the screener "Est. February 27th, 2021" and gives its reference as Cabanac, Labbé and Magazinov (2022), presented at WCRI 2022, the 7th World Conference on Research Integrity; the paper was submitted to arXiv on 7 October 2022 by those three authors, and the MMRS record gives the same three as operator and the same founding date. Writing in The Conversation (29 January 2025) as "a pair of computer scientists at Université de Toulouse and Université Grenoble Alpes" who specialize in detecting bogus publications, Cabanac and Labbé say that Cabanac built the automated tool, and their disclosure names him its administrator: "a public platform that uses metadata from Digital Science and PubPeer via no-cost agreements". The page says it was "Featured in the Nature's 10 list of 2021"; The Conversation says it "was named one of 10 key developments in science by the journal Nature in 2021". The MMRS record calls it a set of detectors run over the literature: tortured phrases, SCIgen and Mathgen papers, references annulled or under concern, and others; the page's menu, updated 2026-10-28, lists ten (Annulled, Concerning, Feet of Clay, Tortured, Suspect, SCIgen, Mathgen, Citejacked, Seek&Blastn, Problematic Cell Lines), and The Conversation, in January 2025, counted eight beside tortured phrases, among them ChatGPT fingerprints, "snippets of telltale text left behind by the AI agent". Tortured phrases are, in the makers' article, nonsensical word sequences found "littered throughout reputable scientific journals", which "typically result from using paraphrasing tools to evade plagiarism-detection software": "the Joined Together States" for the United States, the record's example too; the screener's page quotes Nature's post of 8 August 2021: "Why say 'colossal information' instead of big data? Or 'counterfeit consciousness' instead of artificial intelligence?" Its bibliographic detectors screen the articles citing retracted references, references marked with expressions of concern and references the screener's own detectors flag (its page), the one that finds articles citing retracted works named "Feet of Clay" (The Conversation); for a research question the MMRS record codes it as recording attribution partly, since references to retracted or concerning works are flagged. It harvests the APIs of Crossref, "now including the Retraction Watch Database", Dimensions, PubMed and PubPeer (its page), the sources and intake the MMRS record states with the screener's page as basis. It trawls "130 million scholarly papers every week" for signs of paper-mill production (The Conversation's standfirst, the article saying the same of Cabanac's tool), the weekly scan of 130 million publications the MMRS record gives as its cadence. Its makers start from the claim that post-publication assessment "remains necessary to check erroneous or fraudulent scientific publications" and present the screener as "a new effective tool to curate the scientific literature" (arXiv:2210.04895); Research Signals describes it as an academic project led by Cabanac and collaborators that "combines automatic machine detection and human evaluation". "Problematic papers are reported on PubPeer" (its page). By The Conversation's count it had found nearly 19,000 papers with at least five tortured phrases each, more than 280 SCIgen papers, more than 764,000 articles citing retracted works (about 5,000 with five or more), and more than 70 papers carrying ChatGPT fingerprints such as "Regenerate Response"; the MMRS record carries the same figures, dated "The Conversation, 31 January 2025", where the article at theconversation.com is dated 29 January 2025. Its makers state the limit of the fingerprint count themselves: those papers are "the tip of the tip of the iceberg", output "copy-pasted wholesale into papers without any editing (or even reading)", and "The challenge is to identify cases where chatbots are used not just for language-editing purposes but to generate content – essentially fabricating data." The MMRS record extends that limit to the whole instrument: "A detector finds what leaves a fingerprint", and "Text machine-written cleanly, or cited sources that exist but say something else, pass beneath it." The archive's The Evaluator Exists (#741, 19 May 2026) reads the class the screener works on, paper-mill output, as proxy gaming at industrial scale, "thousands of fraudulent publications that enter the citation network and accumulate proxy signals indistinguishable from legitimate scholarship", holds that under the proxy regime "gaming is hard to detect because the signals being gamed are themselves opaque", and sets against it an evaluator that synthetic scholarship must get past by reading, with "empirical claims that survive a citation audit". Several publishers use it, by its makers' account, some inside the editorial workflow to spot suspect papers upfront, and it "has been instrumental in more than 1,000 retractions"; analytics companies have used it to pick out suspect authors among highly cited researchers, Cabanac has offered pro bono consulting on detection tools to actors including ClearSkies, Morressier, River Valley, Signals and STM (The Conversation), and the MMRS record says it has contributed to more than 1,000 retractions. On 20 June 2024 Research Signals integrated the articles it flags into the Signals Graph as a new signal, "over 12,000 problematic publications that have not been retracted", cited more than 160,000 times, to give sleuths, publishers and researchers visibility of them. On 3 July 2026 CCSD announced that Episciences analyses citations extracted from articles against the PPS database, at submission and after publication, displaying problematic citations for correction or information; of nearly 170,000 references in a first extraction, 226 were indexed in the PPS, most in the Feet of clay category. The same announcement describes the screener as a tool to identify and filter out articles "containing methodological errors, plagiarism, data manipulation, or conflicts of interest", a description that stands beside the screener's own list of detectors. The MMRS record files it as a monitor and calls it "the scholarly record's automated immune response", the neighbour of Academ-AI, "which reads by phrase where the screener reads by detector"; with the screener, its Academ-AI record says, that database records "machine composition entering the scholarly record unattributed: the machine is the uncredited author, and the publisher is where the reception happens." The MMRS record keeps itself apart from the entity, each assertion with a source and a basis and the lifecycle (active) apart from the observation (read, 2026-10-28); it read the screener's page for detectors, sources and start date, leaves the licence unchecked, and relates the capture of 2026-10-29 as the composition layer's account of the entity; the page itself was last updated on 2026-10-28 at 15:54. Google's AI Overview at the name, signed out, 2026-10-29, gave the three makers, the sources and the detectors as the IRIT page and the record give them, and composed the screener at its makers' grain, with its human review stage stated and its uptake by publishers' infrastructure added (Capture Registry, 2026-10-29, checked against the MMRS record and the entity's own pages); no card cited the record, which was published the same night. The Capture Registry records the Overview's example of a tortured phrase, "counterfeit consciousness" for artificial intelligence, as the composition's, the record's being "Joined Together States", not checked against the screener's list; the example stands on the screener's own page, in Nature's post of 8 August 2021. The adoptions the Overview carried, Episciences and the Signals Graph, are uptake the MMRS record lacked (Capture Registry, 2026-10-29); the rail's sixth card is Crossref's video "Exploring Retraction Watch Data in Crossref 2025-05-09", admitted by its title.

T against L_F(B) (AIO's ontology at work on its own cards): At 'problematic paper screener' (T1, six cards: B1–B6) the composition gives the screener at its makers' grain. Composed from its cards: the definition as an online platform (B3 F30) detecting 'fraudulent, error-filled, or computer-generated' literature (B3 F29 'erroneous or fraudulent'; B1 F1), the three makers (B3 F32; B1 F7), the data APIs (B1 F3), tortured phrases as paraphrase used to evade plagiarism detection (B2 F12) by paper mills (B2 F11), SCIgen and Mathgen (B1 F1), ChatGPT fingerprints (B2 F17), the citation of retracted references and expressions of concern (B1 F2) under 'Feet of Clay' (B2 F23), machine detection with human review (B3 F30; B4 F33), reporting on PubPeer (B1 F5), the editorial workflow (B2 F18), the Signals Graph (B4 F34) and Episciences (B5 F38). Composed without attribution, at the coarse grain of an example: 'counterfeit consciousness' for artificial intelligence (T1-3) is Nature's example as the screener's page quotes it (B1 F10, Nature's post of 8 August 2021), composed as the screener's own; recorded, not a breach (§6.2, the grain rule). The Capture Registry records the example as 'the composition's' (CAP-1-03) and did not check it against the screener's list (CAP-1-04); the field ledger locates it on card 1. Limit dropped: the weekly scan of 130 million papers (F11) becomes 'scans millions of scholarly papers' (T1-2), the cadence dropped; the fingerprint count's limit, 'the tip of the tip of the iceberg' (F25), and the challenge of chatbots that generate content (F26) are not composed, 'unedited' (T1-5) keeping a part of F25. Force raised: articles citing retracted works 'that could be unreliable' (F23) become papers that 'improperly cite' them (T1-6); 'Problematic papers are reported on PubPeer' (F5) becomes 'often automatically shared or logged' (T1-8). Only in the transcript (absent from the disclosed field): 'Pepia' (T1-8) is in no saved text; 'withdrawn articles' and 'active' papers (T1-6) are in no ledgered span; 'scrub the scientific record of fabricated or corrupted studies' (T1-9) has its nearest words in Signals' unledgered 'help correct the scholarly record' (B4 ¶4). Available, not composed: the founding date (F4), Nature's 10 (F8, F20), the last update (F9), the article's examples (F13), Cabanac as builder and administrator and the no-cost metadata agreements (F14, F15, F27), the detectors beyond four (F6, F16: Annulled, Concerning, Suspect, Citejacked, Seek&Blastn, Problematic Cell Lines), every count (F21–F24), the 1,000 retractions (F18), the analytics companies and the consulting (F19, F28), the purpose of curation (F31), the Signals figures and aim (F34's 12,000 and 160,000 citations, F35), the database of over one million references (F37), Episciences' check and its figures (F39, F40), and the video (F41, its card on the rail). Not composed and not confused: CCSD's description of the screener as filtering methodological errors, plagiarism, data manipulation and conflicts of interest (F36). Content replaced by use: the closer (T1-10). The own page B7 (unreachable) contributes no claim; no Wikipedia article exists to be composed.

v0.9 §0.5: the first diff is already ontological, and here it is small. The field's ontology and T's agree on the entity (a screening platform of three makers with detectors over bibliographic APIs and a human review stage); what T keeps of its cards is the makers' self-description and the uptake, and what it drops is the screener's measured scale and cadence, its findings and its makers' stated limit of the fingerprint count. Two moves raise force (citing becomes improper citing; reporting becomes automatic sharing) and one platform (Pepia) enters from outside the disclosed field. The row records this as observed and claims nothing about the composer's internals (§0.5).

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the unit, a published paper flagged by a detector (MMRS-16), and the kind, a monitor (MMRS-15); what its selection leaves out, the archive's sense the field states only of the fingerprint count: 'A detector finds what leaves a fingerprint', clean machine text and cited sources that exist but say something else passing beneath it (MMRS-07, MMRS-08); the place in the field, the scholarly record's automated immune response beside Academ-AI, the two recording machine composition entering the scholarly record unattributed (MMRS-09, MMRS-10); attribution recorded partly, through the flagged references (MMRS-20); the record kept apart from the entity, each assertion with a basis, the lifecycle apart from the observation, the licence unchecked, the provenance and the capture related (MMRS-21–MMRS-25); the composition layer's observation at the entity's address (CAP-1-01–CAP-1-07); and #741's reading of the class the screener works on: paper-mill output as proxy gaming, hard to detect because the gamed signals are opaque, against an evaluator that reads (O741-01–O741-03). Coherence (P_coh, composed each with its source): the MMRS record's 'The Conversation, 31 January 2025' (MMRS-18, MMRS-26) against the article's own date, 29 January 2025 (F21); the registry's 'the composition's' example (CAP-1-03) against the screener's page, where Nature's example stands (F10); CCSD's description (F36) beside the screener's list (F1, F6), field against field. Restated lineages (§3.6, the own page keeping the earlier date): MMRS-01 with F1, F30, F4; MMRS-02 with F6, F16, F12; MMRS-03 and MMRS-17 with F3; MMRS-04 and MMRS-19 with F11; MMRS-05 and MMRS-18 with F21–F24; MMRS-06 with F18; MMRS-13 and MMRS-14 with F4, F7, F32. The measure is in row.measure.

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20', 'F21', 'F22', 'F23', 'F24', 'F25', 'F26', 'F27', 'F28', 'F29', 'F30', 'F31', 'F32', 'F33', 'F34', 'F35', 'F36', 'F37', 'F38', 'F39', 'F40', 'F41'], 'T1_composed': ['F1', 'F2', 'F3', 'F5', 'F10', 'F11', 'F12', 'F16', 'F17', 'F18', 'F23', 'F29', 'F30', 'F32', 'F33', 'F34', 'F38'], 'T1_only': ["'Pepia' (T1-8)", "'withdrawn articles', 'active' papers (T1-6)", "'scrub the scientific record of fabricated or corrupted studies' (T1-9)"], 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; the typology of what is not is in T_vs_LB (§6.2)"}