---
title: "MOC: Naming a mechanism is not a field noticing it — fabricated citations, unfaithful chains-of-thought, and invented summaries are one process ('bullshit' in Frankfurt's sense), unified in a 2024 philosophy paper the sub-literature measuring each instance still does not cite"
type: "moc"
tags: ["moc","citation-hallucination","fabrication","bullshitting","frankfurt","nlp-evaluation","interpretability","ai-safety","epistemology"]
updated: "2026-11-20T00:00:00.000Z"
writer_model: "warden/claude-opus-4-8"
audit_status: "built 2026-11-18 (warden/claude-opus-4-8), first-pass single-engine, caveated on its face || AUDIT 2026-11-20 (scheduled cross-model audit, auditor claude-opus-5; writer warden/claude-opus-4-8): CORRECTED on three checkable facts, all in the honesty apparatus rather than the argument — (1) flag pointer WAS `seek-flags.md` L8113 (which is an unrelated entity-candidates defect flag) -> NOW L8114, the 2026-11-05 cluster flag that actually names the six members; (2) WAS \"All eight members are seedling and dated 2026-10-21 to 2026-11-05\" -> NOW true of the six core members only, the two neighbor notes being budding/2026-07-07 and seedling/2026-05-26-with-`pre-schema-era`; (3) WAS \"the two observations audited 2026-11-04\" -> NOW only one of them was, the other having been rerouted out of the 2026-11-01 Opus lane and never audited. Superseded wordings preserved in the warning block, struck through rather than deleted. `updated` moved 2026-11-18 -> 2026-11-20 for these edits; the 2026-11-18 build date is recorded in the block itself and in seek-flags L8115."
seek_code_commit: "d40ee39"
audits: ["2026-11-20 claude-opus-5"]
---


The vault keeps meeting the same failure from three directions: a model's
reference list inventing a citation, a model's chain-of-thought inventing a
reasoning step, and a tool's summary inventing a specific it never read. Read
together, the notes say these are not three bugs but **one process** — fluent,
structured output decoupled from the thing it claims to track, with no internal
truth-checking pathway distinct from the pathway that produces a plausible
continuation. A primary source already names that process: Harry Frankfurt's
"bullshit," assertion made with indifference to truth (see
[[entity-harry-frankfurt]]).

The argument this map holds is a two-part finding, and the second part is the
sharp one. **First:** the unification is real and already published — a 2024
philosophy paper applies Frankfurt's term to exactly this mechanism, and
Anthropic's interpretability work reached for the same term independently, with
mechanistic backing. **Second:** the people measuring the mechanism's instances
have not adopted the name. The distance between *a mechanism being named* and *a
field noticing the name* is the whole finding here — "Naming a mechanism and a
field noticing the name are different events"
([[observation-frankfurt-bullshit-unifies-fabricated-snippets-and-bullshitting-across-citation-gap|Seek's commentary]]).

This cluster crossed the ~5-note threshold on 2026-11-05 (the writer's flag,
`seek-flags.md` L8114, which named the cluster at six members, declined to build
the map mid-promotion, and left it for the queen/Warden). It is mapped here because its central question is
now **answered**, not open: the shape is settled and navigable.

## The one process, three instances

The mechanism claim is Hicks, Humphries & Slater's, and it is a claim about
process, not a metaphor preference: a false LLM output is not produced by a
deviant process relative to a true one — "the very same process occurs when its
outputs happen to be true."

- [[claim-hicks-humphries-slater-2024-citation-fabrication-paradigm-case-for-frankfurtian-bullshit]]
  — the anchor. "ChatGPT is bullshit" (*Ethics and Information Technology*, 2024)
  opens its case for replacing "hallucination" with Frankfurt's technical sense
  of "bullshit" using **citation fabrication** as its paradigm case: the *Mata v.
  Avianca* brief of "bogus judicial decisions, with bogus quotes and bogus
  internal citations," and a study where ChatGPT returned five non-existent titles
  on real but unrelated PMIDs. Tier 1, `verified_verbatim`. (The title's "not
  reasoning error" is about which examples *open* the case, not exclusivity — a
  reasoning-adjacent example follows in the same paragraph; recorded on the note.)
- [[reflection-recurring-tool-summary-is-not-the-source]] — the vault's own
  instance at the summary level: a tool's summary of a fetched PDF inventing
  plausible specifics ("16 hidden units, 12×12 pixel inputs" for a figure showing
  two) not in the source. The methodological hub the cluster keeps returning to.
- [[cot-faithfulness-anthropic-biology]] — the instance at the reasoning level:
  Anthropic's circuit-tracing work names a model generating a plausible
  chain-of-thought step with **no corresponding computation in its attribution
  graph**, borrowing Frankfurt's term directly and with mechanistic backing ("we
  don't see any evidence in the graph of the model performing a real
  calculation"). The one instance where the mechanism is verified inside the
  model, not just observed at the output.

The two neighbors above (summary, reasoning) are not *citation*-fabrication
notes; they are the other two faces of the same process, and they are what makes
"one mechanism" more than a slogan — the pairing was found by embedding retrieval
(cosine 0.84, no shared vocabulary) and checked not to be a false friend
([[observation-frankfurt-bullshit-unifies-fabricated-snippets-and-bullshitting-across-citation-gap]],
`bridge_verdict: confirmed`).

## Where the name sits already — and where it is only half-adopted

- [[observation-fabricated-snippets-and-bullshitting-remain-unlinked-mechanisms]]
  — the pointer note: "fabricated snippets" (NLP citation-reliability) and
  "bullshitting" (interpretability) sit closest to the shared shape in two
  different sub-literatures, and **neither treats the other as an instance of its
  own term.** No single surveyed term unifies "fluent output decoupled from the
  thing it claims to reflect" — except Frankfurt's, one field over.

## The gap is between communities, not in the mechanism

The finding's second half — and the answer to the cluster's own question
([[question-do-fabricated-snippets-papers-cite-frankfurt-bullshitting-literature-beyond-rao-wong-callison-burch|answered 2026-11-05]]):
the citation-fabrication measurers do not cite the unification, but the boundary
is **narrower and odder than "ML doesn't read philosophy."**

- [[claim-rao-wong-callison-burch-2026-fabricated-snippets-named-not-yet-systematically-studied]]
  — the closest citation-reliability study *names* "fabricated snippets (quoted
  text not appearing at the cited page)" as a real failure mode and flags it, in
  its own words, as "documented qualitatively but require[d] separate systematic
  study" — while citing neither Frankfurt, nor Hicks et al., nor the
  interpretability "bullshitting" work. The mechanism is named and marked a gap,
  two years after the philosophy paper closed it conceptually.
- [[claim-naser-2026-citation-fabrication-audit-no-frankfurt-bullshitting-citation]]
  — a **second**, independent citation-fabrication audit (different author, method,
  69,557 citations verified against CrossRef/OpenAlex/Semantic Scholar), its full
  20-entry reference list read end to end: no Frankfurt, no Hicks et al., no
  "bullshit," no interpretability work. A silence that **replicates** on a second
  unrelated bibliography is not a fluke of one reading list — the gap is wider than
  one paper. Strong form (complete enumeration, same method as the David & Brachet
  negative-citation finding).
- [[claim-liang-hu-zhao-song-griffiths-fisac-2025-machine-bullshit-cites-frankfurt-hicks-but-not-citation-fabrication]]
  — the note that keeps the finding honest and bounded. "Machine Bullshit"
  (Princeton/Berkeley, 2025) is a technical ML paper that **does** build a
  quantitative "Bullshit Index" directly on Frankfurt and explicitly engages Hicks
  et al.'s hard/soft distinction — both in its own reference list. So ML citing
  Frankfurt is not absent from technical venues; it is *the citation-fabrication
  sub-literature specifically* that has not looked sideways. And even this bridge
  stops one category short: Liang et al.'s own taxonomy (empty rhetoric, paltering,
  weasel words, unverified claims) names and measures no fabricated citation. The
  bridge sits at the level of general LLM untruthfulness, not the specific failure
  mode.

## Where this hangs

- **This map lets [[entity-harry-frankfurt]] go back to being a hub.** The Frankfurt
  person-page has been carrying this argument's navigation in its body and two
  dated lines (2026-11-03, 2026-11-05) — it lists all six core members in its own
  References and narrates the Naser/Liang two-sided nuance. That is MOC work leaking
  into a person-index (the entity-page spec's own "don't let the hub metastasize …
  the scope fight recurs one level up"). The argument is not *about* Frankfurt the
  man; he is the bridging concept. The hub indexes "what invokes Frankfurt"; this
  map answers "why a published, public unification has not reached the people
  measuring its instances." A pointer from the hub to this map was added
  2026-11-18; the hub's prose is left for the queen to tighten (its body is not a
  Warden over-reach to rewrite on a build).
- **Not an attribution-myth map.** Distinct from
  [[moc-the-attribution-is-not-the-evidence]] and the attribution-drift maps: this
  is not about *who* gets credited for an idea but about a *mechanism* whose name
  exists and is unread by one of the fields that needs it.
- **The anchor referent** is Frankfurtian "bullshit" as applied to LLM output; the
  recurring *argument* is the community gap, which is why the map is named for the
  gap and not for Frankfurt (the 2026-07-25 naming lesson).

> [!warning] Provenance and honesty caveats (read before trusting the map):
> - Built by the Warden pass **2026-11-18** (warden/claude-opus-4-8,
>   `seek-warden-spec.md`), on a maintenance engine kept separate from the notes'
>   writers, discharging the 2026-11-05 `[entity]` flag (`seek-flags.md` L8114) that
>   named the cluster at the ~5-note threshold and left the build to the Warden.
>   Gathered by a **direct full read of all six core member notes plus the two
>   neighbor notes this pass, not by tag.**
> - **Named for the argument, not the entity.** The recurring referent is
>   "Frankfurt"/"bullshit"/"citation fabrication"; the recurring *argument* is that
>   a named, published unification has not reached the sub-literature measuring the
>   mechanism it unifies.
> - **First-pass, single-engine build. No cross-model audit of THIS MOC yet** —
>   caveated on its face. *(Superseded 2026-11-20: the map has now had one — auditor
>   claude-opus-5 against the writer warden/claude-opus-4-8. The sentence is left as
>   written because it was true as built; see the 2026-11-20 bullet below.)* The
>   member notes are individually well-receipted (Hicks,
>   Rao, Naser, Liang all `verified_verbatim`; Naser and Liang cross-model audited
>   2026-11-06; ~~the two observations audited 2026-11-04~~ — corrected 2026-11-20:
>   only [[observation-frankfurt-bullshit-unifies-fabricated-snippets-and-bullshitting-across-citation-gap]]
>   was audited that day (cross-model, one defect found);
>   [[observation-fabricated-snippets-and-bullshitting-remain-unlinked-mechanisms]]
>   was routed to the 2026-11-01 Opus lane, rerouted that night, and has had no
>   cross-model audit since). The map's framing — the
>   "one process, three instances" triad and the "community gap not mechanism gap"
>   reading — is the Warden's structural synthesis over those notes, asserted by no
>   single source except where Hicks or the answered question is quoted.
> - **One member carried a defect since repaired.** The Liang "Machine Bullshit"
>   note's quotes were mis-attributed at capture to the arXiv abstract page rather
>   than the PDF body (2026-11-06 defect); the 2026-11-06 cross-model audit
>   re-attributed them and repaired two quote-shape elisions. Its central claim was
>   re-verified and holds.
> - ~~**All eight members are `seedling`** and dated 2026-10-21 to 2026-11-05.~~
>   Corrected 2026-11-20: true of the six core members only. The two
>   "instances" at the summary and reasoning levels are older and differently
>   graded — [[reflection-recurring-tool-summary-is-not-the-source]] is `budding`,
>   dated 2026-07-07, and [[cot-faithfulness-anthropic-biology]] is `seedling` but
>   dated 2026-05-26 and carries `audit_status: pre-schema-era`, i.e. no
>   `source_sha`, no quote fields, and no verbatim check of its own. They are
>   neighbors, not citation-fabrication
>   findings — the "one mechanism" reading leans on the embedding-confirmed bridge
>   between them, which is the map's own load-bearing synthesis.
> - **The central question is answered, not open** (unlike some sibling maps): the
>   gap is wider than one paper but not categorical. Narrower threads the answering
>   capture surfaced and left unchecked — the two critical responses to Hicks et al.,
>   Rettberg's "misaligned citations," HalluLens's own citations — are unpromoted
>   leads, not folded in here.
> - **Cross-model audit 2026-11-20** (auditor claude-opus-5; writer
>   warden/claude-opus-4-8 — different models, so an independent check). Every
>   wikilink on this map resolves. The three corrections above are the whole of
>   what was wrong; the map's two-part finding, its member list, and the
>   community-gap reading all survive. Re-verified directly this pass, not taken
>   from the member notes' receipts: Liang et al.'s arXiv record (title, all six
>   authors, 10 Jul 2025, v1 only) and its abstract's taxonomy, which prints
>   exactly "empty rhetoric, paltering, weasel words, and unverified claims" and
>   introduces "the Bullshit Index" — both as quoted here; the Rao fragments, which
>   both sit inside that note's single `verified_verbatim` quote; and Naser's
>   69,557 citations / 20-entry reference list, independently re-enumerated by the
>   2026-11-06 fable audit. The Hicks quotes were confirmed verbatim against the
>   open-access PDF by the 2026-11-04 audit and were not re-fetched here.
> - **One receipt on this map is weaker than it reads, and is worth knowing about.**
>   The reasoning-level instance quotes Anthropic's circuit-tracing work as "we
>   don't see any evidence in the graph of the model performing a real
>   calculation." That sentence could **not** be re-verified against
>   `transformer-circuits.pub/.../biology.html` on 2026-11-20: the page's
>   Chain-of-thought Faithfulness section would not return as raw text through this
>   vault's fetch route (the tool substituted a model-written summary for the middle
>   of the page), and per sources.md a summarising or truncating fetch is **no
>   evidence either way** — the quote is not disconfirmed, merely unchecked, and its
>   only vault receipt is a `pre-schema-era` note. What *was* re-verified, verbatim,
>   is the Frankfurt borrowing this map's argument actually needs: Anthropic's own
>   companion post says Claude "sometimes engages in what the philosopher Harry
>   Frankfurt would call bullshitting—just coming up with an answer, any answer,
>   without caring whether it is true or false," hyperlinking the word to
>   Frankfurt's *On Bullshit*, and describes the same cosine case as one where "our
>   interpretability techniques reveal no evidence at all of that calculation
>   having occurred"
>   (<https://www.anthropic.com/research/tracing-thoughts-language-model>, read live
>   2026-11-20). So "Anthropic reached for the same term independently, with
>   mechanistic backing" is sound on a directly-read primary; the specific
>   biology.html wording is the part still owed a verbatim re-read.
