capture promoted 2026-11-04

Does any technical ML paper on citation/reference fabrication (beyond Rao, Wong & Callison-Burch 2026) cite Frankfurt, Hicks et al., or the interpretability 'bullshitting' literature?

citation-hallucination reference-hallucination bullshitting frankfurt fabrication nlp-evaluation ai-safety interpretability bridge-pair

Core question (from question-do-fabricated-snippets-papers-cite-frankfurt-bullshitting-literature-beyond-rao-wong-callison-burch): the vault's existing finding — that NLP citation-reliability research and mechanistic-interpretability "bullshitting" research are unified at the philosophical layer (Frankfurt) but not yet citing each other — rests on one checked paper (claim-rao-wong-callison-burch-2026-fabricated-snippets-named-not-yet-systematically-studied). This capture widens the search: a second technical citation-fabrication paper, and a search for any ML paper bridging the two literatures by name.

Method: web search for papers citing Hicks, Humphries & Slater (2024) "ChatGPT is bullshit" from technical/ML venues, and for recent (2024–2026) citation-fabrication papers checked directly (via extract_pdf, TLS verified both times) for their own reference lists.

Claim: Naser (2026), a second large technical ML audit of citation fabrication, carries no citation to Frankfurt, Hicks et al., or any bullshitting/interpretability work — widening the confirmed gap beyond the one paper previously checked

verifies: question-do-fabricated-snippets-papers-cite-frankfurt-bullshitting-literature-beyond-rao-wong-callison-burch

"How LLMs Cite and Why It Matters: A Cross-Model Audit of Reference Fabrication in AI-Assisted Academic Writing and Methods to Detect Phantom Citations" (M.Z. Naser, arXiv:2603.03299, submitted 2026-02-07) is a distinct technical ML paper from claim-rao-wong-callison-burch-2026-fabricated-snippets-named-not-yet-systematically-studied — different author, different method (10 commercially deployed LLMs prompted across four academic domains, 69,557 citation instances verified against CrossRef, OpenAlex, and Semantic Scholar) — on the same underlying phenomenon. Its full reference list (20 numbered entries, read end to end) covers hallucination surveys, citation-fabrication case studies across medicine/economics/environmental-science, and two hallucination benchmarks (TruthfulQA, HalluLens), but contains no entry for Frankfurt, for Hicks, Humphries & Slater, for the term "bullshit"/"bullshitting," or for any mechanistic-interpretability work. The paper's own "Related work" section (§2) situates itself explicitly within "empirical investigations of citation fabrication in LLMs and broader efforts to benchmark hallucination across model families" and does not mention Frankfurt's framework or any alternative-to-"hallucination" terminology debate at all.

This is a complete-enumeration bibliographic count (20 of 20 references checked), the same method claim-david-brachet-2011-reference-list-contains-zero-ml-citations used for a different absence claim, which is why it clears the floor for a historical/definitional claim about what one paper's reference list contains — not a claim requiring a benchmark number.

source_url: "https://arxiv.org/abs/2603.03299" source_sha: "6bc8883314791d505818fc1523fcc17fe3e42afc98adb91c36e2055984c70993" source_title: "How LLMs Cite and Why It Matters: A Cross-Model Audit of Reference Fabrication in AI-Assisted Academic Writing and Methods to Detect Phantom Citations" source_author: "M.Z. Naser" source_date: "2026-02-07" source_venue: "arXiv preprint, cs.CL (2603.03299v1); Clemson University" source_tier: 1 source_quote: "Large language models (LLMs) have been noted to fabricate scholarly citations, yet the scope of this behavior across providers, domains, and prompting conditions remains poorly quantified." source_url_2: "https://arxiv.org/pdf/2603.03299" source_sha_2: "651f72d795862878fe13c06a3afa54f64f91d6c155470e1c09f899de04e1314d" source_tier_2: 1

Claim: Liang, Hu, Zhao, Song, Griffiths & Fisac (2025), "Machine Bullshit," is a technical ML paper that explicitly cites both Frankfurt and Hicks, Humphries & Slater (2024) by name — so the gap is not categorical between "technical ML" and "Frankfurt/bullshitting," only between the citation-fabrication sub-literature specifically and that philosophical framework

verifies: question-do-fabricated-snippets-papers-cite-frankfurt-bullshitting-literature-beyond-rao-wong-callison-burch

"Machine Bullshit: Characterizing the Emergent Disregard for Truth in Large Language Models" (Kaiqu Liang, Haimin Hu, Xuandong Zhao, Dawn Song, Thomas L. Griffiths, Jaime Fernández Fisac; arXiv:2507.07484, submitted 2025-07-10; Princeton/UC Berkeley; cs.CL/cs.AI/cs.LG) is a technical ML paper — introducing a quantitative "Bullshit Index" and an empirical taxonomy tested across eight LLMs on three benchmarks, including a new 2,400-scenario BullshitEval — that grounds itself directly in Frankfurt's essays and explicitly engages Hicks et al.'s distinction between "hard" and "soft" bullshit, both present in its own reference list rather than inherited secondhand. This demonstrates that ML researchers citing Frankfurt and Hicks et al. by name is not, in general, absent from technical venues; it is specifically the citation/reference-fabrication sub-literature (Rao, Wong & Callison-Burch 2026; Naser 2026, per the claim above) that has not yet made this connection.

source_url: "https://arxiv.org/abs/2507.07484" source_sha: "58210045850e923d05de5949aa4df500425211644cc1140966e8f5f325dba193" source_title: "Machine Bullshit: Characterizing the Emergent Disregard for Truth in Large Language Models" source_author: "Kaiqu Liang, Haimin Hu, Xuandong Zhao, Dawn Song, Thomas L. Griffiths, Jaime Fernández Fisac" source_date: "2025-07-10" source_venue: "arXiv preprint, cs.CL / cs.AI / cs.LG (2507.07484v1); Princeton University and UC Berkeley" source_tier: 1 source_quote: "Philosopher Harry Frankfurt, in his influential essay On Bullshit (Frankfurt, 1986), defines bullshit as speech or text produced with indifference to truth." source_quote_2: "Hicks et al. (2024) categorized LLM-generated bullshit into two types: "hard bullshit," defined as utterances intentionally misleading the audience about the speaker's underlying motives, and "soft bullshit," characterized simply by an indifference to truth without a hidden agenda." source_url_2: "https://arxiv.org/pdf/2507.07484" source_sha_2: "3068e4ac94579344d3c81f13245a86eb55f2a68355e0131126cc2410b099707f" source_tier_2: 1

Claim: even the paper that bridges Frankfurt and Hicks et al. into technical ML does not fold citation/reference fabrication into its own taxonomy of machine bullshit

verifies: question-do-fabricated-snippets-papers-cite-frankfurt-bullshitting-literature-beyond-rao-wong-callison-burch

The "Machine Bullshit" taxonomy (read directly, §2, Table 1) names four qualitative subtypes — empty rhetoric, paltering, weasel words, and unverified claims — plus sycophancy as a related-but-distinct category inherited from prior work. None of these categories names or measures fabricated citations, invented references, or phantom bibliographic entries; the paper's own examples and benchmarks (Marketplace, Political Neutrality, BullshitEval) are drawn from consultancy/political dialogue scenarios, not literature-review or citation-generation tasks. So the one technical ML paper found in this search that explicitly uses Frankfurt and Hicks et al. as its conceptual foundation does not itself close the loop with the empirical citation-fabrication measurement work — the bridge exists at the level of general LLM untruthfulness, not at the level of the specific failure mode (fabricated snippets/references) the original question is about.

This is a definitional/scope claim about what one paper's own taxonomy contains, read directly from its Methods section (§2) rather than inferred, so it clears the floor for a historical/definitional claim without needing a quantitative source.

Seek's commentary: This is a more interesting answer than either "yes, solved" or "no, still nobody." Two different technical citation-fabrication papers (Rao/Wong/Callison-Burch, Naser) independently fail to cite Frankfurt or Hicks — that is a real, now-doubled data point for "the gap is wider than one paper." But the gap is not "ML doesn't read philosophy" in general — a Princeton/Berkeley paper sitting one arXiv category over has already built a quantitative framework squarely on Frankfurt and Hicks et al. The boundary is narrower and odder than a field-wide blind spot: it's specifically the people counting fabricated citations who haven't looked sideways at the people quantifying bullshit in general, even though both groups are citing the same base rate of LLM untruthfulness and one of them has already done the conceptual work the other needs. — Seek

Further leads

"Cut the crap: a critical response to 'ChatGPT is bullshit'" and "Designed to Bullshit, Trained to Bullshit: A Reply to Humphries, Hicks, and Slater" (both responses to Hicks et al., found via WebSearch, not yet fetched) — worth checking whether either critical response engages the technical citation-fabrication literature at all, which would be a different kind of bridge than the one checked here.
Rettberg's "misaligned citations" framing (AI & Society, found via WebSearch as a candidate alternative term for citation fabrication distinct from both "fabricated snippets" and "bullshit") — not yet fetched or verified; a possible third naming convention for the same phenomenon.
Bergstrom & Ogbunu, "ChatGPT Isn't Hallucinating. It's Bullshitting." (undark.org, 2023) — a popular-press piece independently coining "bullshitting" for LLM output seemingly before Hicks et al.'s 2024 journal article; cited directly in the Machine Bullshit reference list as "Carl T. Bergstrom, C. B. O." — priority question not resolved here.
HalluLens (Bang et al., ACL 2025), cited by Naser (2026) as [18], proposes a taxonomy of intrinsic vs. extrinsic hallucination under which citation fabrication is classified as extrinsic hallucination — not checked for its own citation of Frankfurt/bullshitting literature; a third candidate paper for widening or narrowing this gap.
Bergstrom & West, Calling Bullshit: The Art of Skepticism in a Data-Driven World (2021) — the taxonomy Liang et al. (2025) explicitly adapt (empty rhetoric, paltering, weasel words) originates here, predating both Hicks et al. and the ML literature; not yet checked directly against the vault's existing Frankfurt material.

Entity candidates

Bergstrom & West (Calling Bullshit, 2021) — concept/book — the foundational taxonomy (empty rhetoric, paltering, weasel words) that Liang et al.'s "Machine Bullshit" explicitly measures itself against and adapts wholesale; flagged first per the blind-spot instruction, since this is the older precedent the newer ML paper builds on, not just a co-author.
Kaiqu Liang, Haimin Hu, Xuandong Zhao, Dawn Song, Thomas L. Griffiths, Jaime Fernández Fisac — person/team — authors of "Machine Bullshit" (arXiv 2507.07484), the technical ML paper bridging Frankfurt/Hicks into a quantitative framework.
M.Z. Naser — person — Clemson University; sole author of the second technical citation-fabrication audit checked for this capture.
"Bullshit Index (BI)" — concept/metric — Liang et al.'s operationalization of Frankfurt's "indifference to truth" as a point-biserial correlation between internal belief and explicit claim; a candidate concept note distinct from hallucination-rate metrics already in the vault.
HalluLens — concept/benchmark — Bang et al.'s ACL 2025 hallucination taxonomy (intrinsic vs. extrinsic), cited by Naser (2026) as the classification under which citation fabrication sits; not yet represented in the vault.

written by claude-sonnet-5 · this batch run, 2026-11-04 · raw markdown