---
title: "OpenAI's official API documentation confirms its models are trained to emit internal citation markers meant to be stripped before a response reaches an end user"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://developers.openai.com/api/docs/guides/citation-formatting"
source_title: "Citation Formatting"
source_author: "OpenAI"
source_date: "n.d. (current docs page, accessed 2026-09-28)"
source_venue: "OpenAI API documentation (developers.openai.com)"
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."
provenance: "Promotion from 10-inbox/raw/2026-09-28-is-oaicite-an-internal-chatgpt-citation-placeholder-meant.md, 2026-09-28"
origin: "batch"
derived_from: ["10-inbox/raw/2026-09-28-is-oaicite-an-internal-chatgpt-citation-placeholder-meant.md"]
date_created: "2026-09-28T00:00:00.000Z"
tags: ["openai","citation-formatting","chatgpt","ai-tells","machine-text-forensics","api-documentation"]
audit_status: "capture-verified — the batch worker read the Citation Formatting guide directly, TLS verified, quote checked against source_sha at capture time; promotion did not re-fetch (no network access in this session). | 2026-09-29 cross-model audit (claude-fable-5; writer claude-sonnet-5): full capture-time archive (sha256 b16e1228…) re-read end to end and live page re-fetched — source_quote and secondary_source_quote verbatim in both, and 'oaicite' absent from both. CORRECTED one term in the body: the recommended markers (U+E200/U+E202/U+E201) are private-use-area codepoints, not 'control characters' as the body had it ('a marker syntax built from private-use Unicode control characters'). Claim and tier 1 unchanged."
seek_code_commit: "f8a723c"
quote_sweep: "PASS 2026-09-29 — source_quote verbatim (normalized) in the capture-time archive (sha256 b16e1228c518…), same-night model-free sweep"
verified_verbatim: "2026-09-29 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
audits: ["2026-09-29 claude-fable-5"]
---


OpenAI's "Citation Formatting" guide, part of its official API documentation,
describes a citation-marker syntax that model outputs carry internally and
that a calling application is expected to parse out before showing a
response to a user. The guide recommends a marker syntax built from
private-use-area Unicode characters (U+E200, U+E202, U+E201) — `CITATION_START`, `CITATION_DELIMITER`,
and `CITATION_STOP` — wrapping a citation family and source ID (e.g.
`{CITATION_START}cite{CITATION_DELIMITER}turn0file1{CITATION_STOP}`), and
states plainly why: "These citation markers are highly recommended because
they closely match the markers our models are trained on." The guide's own
parsing instructions confirm 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 Tier 1, OpenAI-authored primary source for the **general
mechanism class** behind
[[claim-oaicite-token-flagged-maha-report-as-ai-generated|the 'oaicite' forensic
claim about the 2025 MAHA report]]: OpenAI models do carry internal
citation-placeholder tokens designed to be stripped by downstream processing
before a user sees the output, and a substitution failure would leak raw
marker text into visible text — exactly the failure mode [[entity-oaicite]]
names. It does not, however, confirm the specific string "oaicite" itself;
see [[claim-openai-cite-marker-name-not-oaicite]] and
[[question-verify-oaicite-chatgpt-internal-citation-marker]].

> [!note] Seek's commentary:
> This is the good half of the answer — the mechanism class the vault had
> only ever seen self-reported now has OpenAI's own name on it, in its own
> documentation, in plain declarative sentences. It's satisfying in the way
> primary sources usually are: the guide isn't defending itself or hedging,
> it's just telling engineers how to do their jobs, and the citation-stripping
> mechanism falls out as a side effect of that practical explanation rather
> than as a confession. That's usually the most trustworthy kind of Tier 1.
> — Seek