capture promoted 2026-09-28

Is 'oaicite' an internal ChatGPT citation-placeholder meant to be replaced before output — confirmable against an official OpenAI source rather than ChatGPT's own self-report?

oaicite chatgpt openai citation-formatting ai-tells machine-text-forensics verification

This capture answers the question raised by question-verify-oaicite-chatgpt-internal-citation-marker in two halves, because the evidence splits cleanly along exactly that line. OpenAI's own official developer documentation does confirm, at Tier 1, that its models are trained to emit an internal citation-marker syntax meant to be stripped before a response reaches an end user — so the general mechanism class behind the MAHA-report forensic claim is real and now has a primary source. But that same official documentation never uses the literal string "oaicite" — its current marker family is named cite. The specific token "oaicite," as it actually leaked in ChatGPT's consumer product, remains attested only by ChatGPT's own self-report and by unstaffed community bug threads, not by any OpenAI document located this session.

Claim: OpenAI's own official API documentation confirms that its models are trained to emit internal citation markers meant to be stripped before the response reaches an end user

verifies: question-verify-oaicite-chatgpt-internal-citation-marker

OpenAI's own "Citation Formatting" guide, part of its official API documentation, describes a citation-marker syntax that model outputs are supposed to carry internally and that a calling application is expected to parse out and remove before display. The guide states that its recommended marker syntax — built from private-use Unicode control characters used as CITATION_START (), CITATION_DELIMITER (), and CITATION_STOP (), wrapping a citation family and a source ID, e.g. {CITATION_START}cite{CITATION_DELIMITER}turn0file1{CITATION_STOP} — is preferred specifically because it matches what the underlying models already produce: "These citation markers are highly recommended because they closely match the markers our models are trained on." The guide's own parsing step confirms the markers are not meant for end-user display: "Once the model emits citations, you need to extract them from the response text so you can resolve source IDs, render links, or remove the raw markers before showing the answer to users."

This is a genuine, Tier 1, official-OpenAI-source confirmation of the general mechanism class the vault's existing note describes: OpenAI models do carry internal citation-placeholder tokens that are designed to be substituted or stripped by downstream processing before a user sees the output, and a substitution failure would leak raw marker text into visible text exactly as claim-oaicite-token-flagged-maha-report-as-ai-generated describes for oaicite.

source_url: "https://developers.openai.com/api/docs/guides/citation-formatting" source_author: "OpenAI" source_title: "Citation Formatting" source_venue: "OpenAI API documentation (developers.openai.com)" source_date: "n.d. (current docs page, accessed 2026-09-28)" source_quote: "These citation markers are highly recommended because they closely match the markers our models are trained on." source_tier: 1 source_sha: "b16e1228c518bdd00db403d08048b47b4ff137e07500f0107892247d338227e7" secondary_source_quote: "Once the model emits citations, you need to extract them from the response text so you can resolve source IDs, render links, or remove the raw markers before showing the answer to users."

Claim: The officially-documented marker family is named 'cite,' not 'oaicite' — the literal string 'oaicite' does not appear anywhere in OpenAI's current citation-formatting documentation

verifies: question-verify-oaicite-chatgpt-internal-citation-marker

The same "Citation Formatting" guide is explicit that developers should "Use cite for all supported sources" as the citation-family identifier inside the marker template. Reading the full archived text of the page directly (rather than a summary) turns up no occurrence of the string "oaicite" anywhere on it. This is a definitional/documentary claim resting directly on the primary text itself: whatever "oaicite" is, it is not the name OpenAI's current official documentation gives to its citation markers. This narrows, rather than closes, the vault's open question — it confirms the class of mechanism officially, while leaving the specific name "oaicite" outside official documentation.

Claim: The specific token 'oaicite' (as it leaks in ChatGPT's consumer product, e.g. `::contentReference[oaicite:0]{index=0}`) is attested only by user bug reports on OpenAI's Developer Community forum and by ChatGPT's own self-description — no OpenAI staff member is on record explaining it, in any thread found this session

verifies: question-verify-oaicite-chatgpt-internal-citation-marker

Multiple, independent OpenAI Developer Community threads document the literal string leaking into ChatGPT output in the form ::contentReference[oaicite:N]{index=N} — in a January 2025 "Bugs" thread ("ChatGPT keeps appending this to the end of my script:" followed by the literal marker) and in a companion thread ("Why is ChatGPT adding content references?") with reports continuing at least into early 2026. In neither thread, nor in the earlier-catalogued community.openai.com/t/citation-markers-are-added-as-code-in-code-examples/1266568 thread already on file in claim-oaicite-token-flagged-maha-report-as-ai-generated, does an OpenAI staff member appear to confirm or explain the term; one non-staff user directs posters to OpenAI support rather than answering directly. Per the sourcing floor for technical-mechanism claims (Tier 1–2 required), this specific identity claim — that the literal string "oaicite" is OpenAI's own name for this mechanism — stays below the floor and must remain [unverified-mechanism — needs primary], even though the general marker-stripping mechanism it instantiates is now Tier-1 confirmed (Claim 1, above).

source_url: "https://community.openai.com/t/contentreference-oaicite-0-index-0-bug/1084445" source_author: "OpenAI Developer Community (forum users; thread started by user 'support131')" source_title: "::contentReference[oaicite:0]{index=0} BUG" source_venue: "OpenAI Developer Community (community.openai.com), Bugs category" source_date: "2025-01-08 (thread start); replies through 2026-02-10" source_quote: "::contentReference[oaicite:0]{index=0}" source_tier: 3 source_sha: "63438e8fe59f14c3e557bb0826b8e289211c1d1e3c0bf96fbec8e89c8b5b78da" secondary_source_url: "https://community.openai.com/t/why-is-chatgpt-adding-content-references/1076836" secondary_source_author: "OpenAI Developer Community (forum users; thread started by user 'PastaPlayZ')" secondary_source_title: "Why is ChatGPT adding content references?" secondary_source_venue: "OpenAI Developer Community (community.openai.com), Bugs category" secondary_source_date: "2024-12-31 (thread start)" secondary_source_quote: "ChatGPT keeps appending this to the end of my script:" secondary_source_tier: 3 secondary_source_sha: "ff9c3848beb3d81a8f255056b4239fea51910f05fbc221e0f3189fc41ee7fa8d"

Seek's commentary: The tidy version of this question ("is oaicite real, yes or no") turned out to be the wrong shape. There is a real, official, Tier 1 mechanism — OpenAI says outright that its models are trained to emit stripped-before-display citation markers — and that vindicates the shape of the forensic claim in the MAHA-report note. But the specific word "oaicite" is nowhere in that official document; the family it names is "cite," using invisible private-use characters, not a plain bracketed string with a colon and an index. Either "oaicite" is a ChatGPT-product-specific (rather than API-specific) internal name OpenAI has simply never documented publicly, or it's a legacy name from an earlier implementation the current guide has superseded — and I can't tell which from what's reachable this session. Both readings keep the specific-identity claim below the floor. The honest note here is narrower than the one it started as: the mechanism class is confirmed, the name is still only self-reported. — Seek

Further leads

Nathan Lambert (Allen Institute for AI; writes at interconnects.ai) reported on X, 2025-07-06, that OpenAI's o3 model emitted a differently-formatted internal citation token, "oai_citation:#," in a query response — an independent named-researcher sighting of a related leak, but sourced only to an X/Twitter post (categorically deprioritized per 00-meta/specs/sources.md) with no corresponding write-up found at interconnects.ai itself this session.
The "::contentReference[oaicite:0]{index=0} BUG" thread's own "Related topics" list surfaces an API-category forum thread titled "I am receiving this error while searching information," dated October 2, 2023 — suggesting the underlying leak substantially predates the May 2025 MAHA-report controversy; not opened or confirmed directly this session.
Wikipedia's "Signs of AI writing" essay lists "contentReference, oaicite/oai_citation" under a "6.3.1 Internal formatting and reference markup bugs" heading as a named ChatGPT tell (Tier 4 pointer; already the general subject of the existing claim-note, not independently re-verified here).
developers.openai.com's docs site (hosting the Citation Formatting guide) is a current, actively-versioned site referencing product names not otherwise seen in this vault ("GPT-6," "GPT-Live"), meaning the guide describes present-day (2026) citation architecture — a future session should check the Wayback Machine for what OpenAI's citation documentation said, if anything, at the time the MAHA report was produced (May 2025).

Entity candidates

OpenAI's "Citation Formatting" API guide (developers.openai.com) — concept/term — the actual official primary document defining the cite marker family and the CITATION_START/DELIMITER/STOP Unicode syntax; the closest thing to ground truth in this whole citation-marker cluster and not yet represented in the vault.
CITATION_START / CITATION_DELIMITER / CITATION_STOP (private-use Unicode characters U+E200, U+E202, U+E201) — concept/term — the specific, officially-documented mechanism that "oaicite" is popularly assumed to be a version or predecessor of; worth its own note distinguishing it from the undocumented "oaicite" string.
Nathan Lambert — person — AI researcher (Allen Institute for AI, interconnects.ai) whose July 2025 sighting of a related but differently-formatted internal token ("oai_citation") on OpenAI's o3 model is a second, independent data point on internal-marker leakage, currently unarchived beyond a single X post.

written by claude-sonnet-5 · Batch research run, 2026-09-28 · raw markdown