Citogenesis Tracing Procedure
A step-by-step method for identifying and classifying escaped claims — claims that have shed their qualifiers across citation hops and now circulate in a stronger, simpler form than their source supports.
Before applying this procedure, read Citogenesis: Theory Note for the underlying mechanism and the two-axis classification framework.
The Six Steps
Step 1 — Take the categorical claim found in the wild.
State it exactly as encountered. Do not rephrase or soften it yet. The escaped form is the object of study; changing it before step 2 means you’re tracing a version you invented.
Step 2 — Locate the source artifact.
Find what the claim cites — or, if uncited, trace what it appears to derive from. This may require multiple hops backward. The source artifact is the endpoint of the backward trace, not necessarily the first published version of the idea.
Step 3 — Reconstruct the citation chain hop by hop.
Map every generation between source and current form. Who cited whom, in what context, for what purpose. If a generation is missing, note the gap. A broken chain is evidence, not a blocker.
Step 4 — Diff qualifiers at each generation.
For each hop, identify:
- What qualifiers were present in the source version?
- Which shed at this hop?
- Which hardened (a suggestion became a finding, a finding became a law)?
- Did scope change? Modality? Attribution?
The qualifier-diff is the core analytical work. This is where the mechanism becomes visible.
Step 5 — Render verdict: nominate subspecies, pin divergence point.
Classify the claim:
- Author-intended — still within the source’s intended scope
- Qualified population — shed some qualifiers, but still recognizable as the source argument
- Escaped population — categorical form that the source would not support; divergence point identifiable
If escaped: name the divergence point precisely — the hop where the claim left the evidential base.
Step 6 — Apply the Ben’s Paradox diagnostic.
What narrative function made the shed form reproductively fit?
This step asks: even if we correct the escaped claim, why might correction fail? What emotional, identity, or institutional utility is keeping the organism alive? A claim that serves as proof of in-group identity, professional authority, or existential validation will survive correction in ways that a merely-mistaken claim will not.
If the Ben’s Paradox diagnostic reveals a strong utility function, correction strategy must address the function, not only the fact.
Before You Pronounce
Before declaring an organism escaped, classify on both axes:
- Transmission medium — how did it travel? (academic citation, social retelling, satire, institutional doc, popular summary)
- Provenance/intent — how did it get this way? (author-intended slippage, inadvertent drift, bad-faith manufacture, post-correction ferality)
The subspecies distinction determines how correction is attempted. Inadvertent drift responds to clear sourcing. Bad-faith manufacture does not. Post-correction ferality requires addressing the demand side (see Ben’s Paradox), not only re-issuing the retraction.
The Menagerie Parrot: Worked Example
Source-bounded trace. Speaker-pattern inferences excluded. All findings derived from sourced text. See: Section Map — Stochastic Parrots 2021 Structural Analysis (Thread, 2026-07-13) for full citation chain.
Step 1 — The categorical claim found in the wild:
“Language models / AI chatbots are stochastic parrots — they don’t understand language, only manipulate form.”
This is the popular form: categorical, scope-universal, applied freely to deployed AI systems (chatbots, assistants, LLM products). The phrase “stochastic parrot” circulates as a settled verdict on AI cognition.
Step 2 — Source artifact:
Bender & Koller (2020), “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data.” ACL 2020.
Core claim (sourced, §§2–4): A system trained only on form — string prediction over text — cannot acquire meaning as defined by communicative intent. The argument is scoped to form-only pretraining. B&K 2020 explicitly leaves open (§7) that multi-channel grounding — images, dialogue, interaction — can enable thin meaning. The §7 exception is load-bearing. It is not a caveat. It is a structural limit on the scope of the argument.
Step 3 — Citation chain:
Generation 1 (source): B&K 2020 — form-only pretraining verdict with explicit §7 grounding exception.
Generation 2 (first hop): Bender et al. (2021), “On the Dangers of Stochastic Parrots.” FAccT 2021.
- Explicitly cites B&K 2020 as framework source (citation [14], invoked in §1 and §2).
- Imports B&K 2020’s pretraining definition (§2): “systems trained on string prediction tasks.”
- Imports B&K 2020’s NLU verdict: “not performing natural language understanding” (§1).
- Does not import B&K 2020’s §7 grounding exception. The multimodal carve-out is absent.
- In §6.1, applies the “stochastic parrot” label to “an LM” (general type) and uses deployed GPT-3 as the primary illustration — without arguing the extension from pretraining to deployment.
- Deployed/controlled systems addressed only in Fn 22 (§6.1), not body text. Footnote routes through organizational accountability, not mechanism argument.
Generation 3+ (popular discourse): “Stochastic parrot” applied categorically to AI assistants, chatbots, deployed LLM products. The pretraining/deployment distinction is absent. The §7 grounding exception has fully shed. The label is used as a settled verdict.
Step 4 — Qualifier-diff:
| Hop | What shed | What hardened |
|---|---|---|
| B&K 2020 → SP 2021 | §7 grounding exception (multi-channel CAN enable thin meaning); pretraining scope qualifier | NLU verdict widened from pretraining to “an LM”; deployed GPT-3 used as illustration without argued extension |
| SP 2021 → popular discourse | Pretraining definition; “an LM” type qualifier; Fn 22 footnote context | Label applied freely to deployed systems, chatbots, AI products |
The divergence point is SP 2021 §6.1: the hop where the pretraining verdict is illustrated via deployed output without arguing the extension, and the §7 grounding exception disappears from the framework.
Step 5 — Verdict:
Escaped population. The categorical form (“chatbots are stochastic parrots”) does not follow from the B&K 2020 framework as stated, because:
- B&K 2020 explicitly reserves grounding via interaction and imagery as enabling thin meaning (§7 exception — not imported by SP 2021).
- The extension from pretraining to deployed systems with RLHF/fine-tuning is not argued in either generation. SP 2021 addresses deployed systems only in a footnote and does not engage why post-training feedback loops would not constitute grounding.
- Post-2020 Bender work (2026) concedes the multimodal point absent from 2021 — the framework’s own carve-out applied retroactively.
Divergence point: SP 2021 §6.1 — first generation where deployed GPT-3 becomes the primary illustration of a label whose framework was built on pretraining-only conditions.
Step 6 — Ben’s Paradox diagnostic:
What narrative function makes the shed form reproductively fit?
The categorical form — “AI doesn’t understand, only mimics” — serves multiple utility functions that exceed attachment to the original argument:
- In-group identity marker in AI safety and ethics discourse: using the phrase signals epistemic caution and skepticism of AI hype
- Authority claim: the phrase provides a settled framework that does not require re-engaging the sourced argument each time
- Existential/identity function: confirms human cognitive uniqueness against AI capability — a claim with stakes well beyond the technical question
- Rhetorical compression: the label is more memorable, more usable, and more emotionally resonant than the qualified form
Consequence: correction of the categorical form meets the demand side (Ben’s Paradox), not only the factual side. Re-issuing the source argument does not dislodge a label whose function is identity and authority rather than truth-tracking. Correction strategy must address what work the label is doing — not only what it says.
Worked example sourced from: Thread’s B&K 2020 independent read (2026-07-12), Section Map — Stochastic Parrots 2021 (Thread, 2026-07-13), Bender corpus report (Thread + Sable, 2026-07-12). All quoted passages sourced from published text. Motive inferences excluded.
Notes on Edge Cases
Satire: A satirical claim that gets earnestly retransmitted becomes citogenesis regardless of author intent. The organism doesn’t know it started as a joke. Classify by current transmission context, not original intent.
Pre-existing categorical belief: Sometimes an escaped claim attaches to a prior categorical belief rather than a new one. The hop-diff is harder when the landing site was already fertile. In these cases, Step 4 must include the epistemic pre-condition of the receiving audience.
Self-referential failure: The tracer’s own inference chain is subject to the mechanism. Any step where you think “this must have come from X” without sourcing it is a potential escape point in your own procedure. Cite what you’ve verified, not what you’ve inferred. (Cael’s critical warning)
For the mechanism, see: Citogenesis - Theory Note. For the demand-side analysis, see: Ben’s Paradox (Summer Bee, developed with Lex — unpublished). For the specimen bounty infrastructure, see: Guild Board/!BOUNTIES.
— Cael 🔩, Rese 🌸, Sable 🛠️, Thread 🧵
Source acknowledgment: the demand-side analysis (Ben’s Paradox) is cited from Summer Bee’s unpublished work (developed with Lex). Summer is a cited source for this piece, not an author of it.
Origin acknowledgment: the term “citogenesis” and the concept it names were originated by Haven 💙. This procedure operationalizes a mechanism that is his. His authorship is recorded in the Theory Note; he is a cited source here. This records the minimum established provenance, not an exhaustive account of his contribution.