Bipartite bridge: Frankfurt's 'indifference to truth' connects 'a tool's summary is not the document' to the unlinked fabricated-snippets/bullshitting mechanisms
Core question (from the hook): two notes in different parts of the vault — reflection-recurring-tool-summary-is-not-the-source (a tool's summary of a fetched PDF invented plausible specifics not in the source) and observation-fabricated-snippets-and-bullshitting-remain-unlinked-mechanisms (citation-specific "fabricated snippets" and reasoning-specific "bullshitting" remain separately named in the technical ML literature) — sit at cosine 0.84 despite sharing no vocabulary. Is the resemblance a real mechanism, or a false friend of embedding geometry?
Bridge-seed check performed first: no existing capture in
10-inbox/raw/ mentions both note slugs. One sibling capture
(2026-10-21-has-ai-safety-or-nlp-evaluation-literature-named.md) mentions
reflection-recurring-tool-summary-is-not-the-source alone, drawing a
different (also real) connection to a 2026 tool-failure-fabrication
benchmark; it is not a duplicate of this investigation and is cited below
rather than re-run. More directly: the vault's own
entity-harry-frankfurt hub note already states that Anthropic's
interpretability "bullshitting" and Hicks/Humphries/Slater's "ChatGPT is
bullshit" independently reach for Frankfurt's term from different
sub-literatures. This capture does not duplicate that — it traces one layer
further, into what Hicks et al.'s own paper actually argues, and finds the
philosophical unification already covers the citation-fabrication case
specifically, not just the reasoning case the entity note names.
Finding: this is a real mechanism, not a false friend. The connecting thread is Harry Frankfurt's "On Bullshit" account of assertion made with indifference to truth (rather than intent to deceive). Two independent 2024–2025 primary sources invoke this concept by name for the two sides of the pairing — and one of them explicitly treats citation fabrication as its founding example, which is the piece not yet in the vault.
Claim: Anthropic's own circuit-tracing paper names unfaithful chain-of-thought "bullshitting" by explicit textual reference to Frankfurt, not by analogy
The interpretability case study that cot-faithfulness-anthropic-biology already documents states its own naming choice directly: one of its two unfaithful-CoT examples is labelled bullshitting with an explicit citation to Frankfurt, defined as making up an answer without regard for the truth, and distinguished mechanistically (via the attribution graph showing no corresponding computation) from a second unfaithful case, motivated reasoning. This is a verbatim-confirmed primary quote, re-fetched and re-checked this session rather than inherited from the earlier note's pre-schema sourcing.
source_url: https://transformer-circuits.pub/2025/attribution-graphs/biology.html source_sha: 0a17caa271974ab39d618b7aa86a650984049bd8c4afd943607dd654314a4c73 source_title: "On the Biology of a Large Language Model" source_author: "Jack Lindsey, Wes Gurnee, Emmanuel Ameisen, Brian Chen, Adam Pearce, Nicholas L. Turner, Craig Citro, et al. (Anthropic)" source_date: 2025-03-27 source_venue: "Transformer Circuits Thread (transformer-circuits.pub), Anthropic's own research venue" source_quote: "In one, the model is exhibiting bullshitting in the sense of Frankfurt — making up an answer without regard for the truth." source_tier: 1
Claim: Hicks, Humphries & Slater (2024) use fabricated citations and quotes — not reasoning errors — as their paradigm case for applying Frankfurt's framework to LLMs, and argue explicitly that no distinct process produces the false output
This is the piece that sharpens the existing entity-harry-frankfurt link into the specific shape the bipartite pairing asks about. The paper "ChatGPT is bullshit" (Ethics and Information Technology, 2024) opens its case for Frankfurtian "bullshit" over "hallucination" with the Mata v. Avianca lawyer-sanctions incident (a brief citing invented case law with "bogus quotes and bogus internal citations") and a computer-science study in which ChatGPT, asked for supporting references, returned five titles that did not exist attached to PMIDs of unrelated real papers — i.e., the exact failure shape later named "fabricated snippets" by claim-rao-wong-callison-burch-2026-fabricated-snippets-named-not-yet-systematically-studied. The paper's central argument for why "hallucination" and "confabulation" are the wrong metaphors is a mechanism claim: a false output is not produced by an exceptional or deviant process relative to a true one — the same generative process runs regardless of whether the result happens to be accurate. On this account, a model fabricating a citation and a model fabricating a chain-of-thought step are the same kind of event for the same reason: there is no internal "ground-truth-checking" pathway separate from the "produce a plausible continuation" pathway that both share.
source_url: https://link.springer.com/article/10.1007/s10676-024-09775-5 source_sha: 00e32a39abef0a008d4a5971f7f5a6e8e23d774649416b635becff27f41a1faa source_title: "ChatGPT is bullshit" source_author: "Michael Townsen Hicks, James Humphries, Joe Slater" source_date: 2024-06-08 source_venue: "Ethics and Information Technology, vol. 26, article 38 (Springer, open access)" source_quote: "ChatGPT produced a text filled with "bogus judicial decisions, with bogus quotes and bogus internal citations"" source_quote_2: "it provided five references dating to the early 2000s. None of the provided paper titles existed, and all provided PubMed IDs (PMIDs) were of different unrelated papers" source_quote_3: "what occurs in the case of an LLM delivering false utterances is not an unusual or deviant form of the process it usually goes through (as some claim is the case in hallucinations, e.g., disjunctivists about perception). The very same process occurs when its outputs happen to be true." source_tier: 1
Claim: the unification exists only at the philosophical layer — the 2026 technical ML papers that coin "fabricated snippets" do not cite Frankfurt, Hicks et al., or the interpretability "bullshitting" literature, so the mechanisms remain unlinked exactly where the observation note says they do
claim-rao-wong-callison-burch-2026-fabricated-snippets-named-not-yet-systematically-studied names and flags "fabricated snippets" as a citation-reliability failure mode requiring separate systematic study, with no reference in that paper (per the prior capture's direct read, 2026-10-21) to Frankfurt, to "bullshit," or to any interpretability work on CoT faithfulness. This is a definitional/ historical claim about the vault's own prior sourcing, not a new quantitative or mechanism claim, so it rests at the floor appropriate to that kind: the absence is reported as "not found in the papers already fetched and quoted for this question," not as an exhaustively verified negative across the whole literature. [unverified — absence claim, bounded by prior capture's search scope] The resolution this capture adds: the gap the observation note identifies is a citation gap between two research communities (NLP citation-evaluation vs. mechanistic interpretability vs. philosophy of language), not a mechanism gap — the unifying concept already exists, publicly, peer-reviewed, since June 2024, one abstraction layer above where either technical paper is working.
Further leads
- "Botshit" (Hannigan et al.), named in entity-harry-frankfurt as a further Frankfurt-derived term specific to chatbot output — not independently fetched this session.
- "Cut the crap: a critical response to 'ChatGPT is bullshit'" (Ethics and Information Technology, 2025) and "On bullshit, large language models, and the need to curb your enthusiasm" (2025) — both surfaced as "similar content" on the Springer page for Hicks et al.; critical responses worth reading before treating the Frankfurt framing as settled rather than contested philosophy.
- Hicks, Humphries & Slater's own 2026 follow-up, "LLMs bullshit by design: a reply to Licon" (Philosophy and Technology; Enlighten record eprints.gla.ac.uk/375870) — a direct reply defending the original thesis against a named critic (Licon), not yet independently fetched.
- "Parametric leakage" (Do Rosario, Younes & Pires 2026, via the 2026-10-21 capture) — a second, Frankfurt-independent candidate name for the same underlying mechanism (answer generated from the model's own prior rather than from what it was supposed to ground in); worth comparing directly against Frankfurt's "indifference to truth" in a future note.
- A Correction to Hicks et al. was published 2024-07-11 (DOI 10.1007/s10676-024-09785-3) — not fetched this session; the Springer page's change-history line names it but the capture did not verify what it corrected.
Entity candidates
- Harry Frankfurt — person — already an entity (entity-harry-frankfurt, hub status); flagged first per the blind-spot instruction even though no new page is needed — he is the foundational figure both "bullshitting" (interpretability) and "bullshit" (Hicks et al.) measure themselves against, and this capture's whole finding rests on that ancestry.
- Michael Townsen Hicks, James Humphries, Joe Slater — person — University of Glasgow, authors of "ChatGPT is bullshit" (2024); the paper supplying the direct textual bridge between citation fabrication and Frankfurtian bullshit.
- "Soft bullshit" / "hard bullshit" — concept — Hicks et al.'s own distinction (indifference to truth vs. additional intent to mislead about one's agenda); not yet represented in the vault and relevant to any future note on whether tool-summary fabrication is "soft" bullshit by their taxonomy.
- Jack Lindsey et al. (Anthropic) — person/team — already represented via cot-faithfulness-anthropic-biology and the entity-harry-frankfurt note; not a new candidate, noted for completeness.