A citation is a place to check, not proof that a sentence is supported. To catch a plausible but wrong AI answer, give the model a specific source excerpt, break its answer into claims, and check each claim against the words in that source; then correct or qualify any claim that goes beyond the evidence.
Prerequisites and intended result
Use ChatGPT with a short passage you can inspect yourself. You can paste the passage directly or attach a supported text or document file; ChatGPT supports TXT, PDF, and DOCX uploads, subject to plan, workspace, and account limits. OpenAI’s file-upload guide describes the available formats and limitations.
The goal is a small evidence table: each factual claim gets a supporting phrase, a status, and a correction when needed. “Not stated” is useful: it means the passage does not establish that claim. It does not prove the claim is false in the world.
Work through one answer against its source
The following passage and answer are invented for practice. The answer sounds plausible, but it adds a condition and a deadline that the passage never states.
Example input: Paste this source and task into ChatGPT:
Source S1 — Example travel policy excerpt
“Employees may request reimbursement for approved business travel by submitting an expense report with receipts. Reports should be submitted promptly after the trip. Reimbursement is subject to manager approval.”
Task: Check the answer below against Source S1 only. Split it into individual factual claims.
For each claim, quote the shortest supporting phrase, mark Supported, Contradicted, or Not stated, and explain briefly.
Do not use outside knowledge or infer a deadline. If the passage does not establish a claim, say so.
Then write a corrected answer using only supported information.
Answer to check:
“Employees must file within 30 days. They can be reimbursed for any business trip as long as they include receipts. Manager approval is optional.”
Expected output: The table below is an illustrative output, not a claim that this prompt was run in ChatGPT.
| Claim in the answer | Evidence in S1 | Status | Correction |
|---|---|---|---|
| Reports must be filed within 30 days. | “promptly after the trip” | Not stated | The passage says to submit promptly; it gives no 30-day deadline. |
| Any business trip qualifies if receipts are included. | “approved business travel”; “with receipts” | Contradicted | The passage limits this to approved travel and requires receipts. |
| Manager approval is optional. | “subject to manager approval” | Contradicted | Reimbursement is subject to manager approval. |
Corrected answer: “Employees may request reimbursement for approved business travel by submitting an expense report with receipts. Reports should be submitted promptly after the trip, and reimbursement is subject to manager approval.”
The prompt makes the model distinguish “not stated” from “contradicted.” That distinction matters: the passage does not say there is no 30-day rule elsewhere; it simply does not support that deadline. The corrected answer should not silently fill gaps with general knowledge.
Check every claim before accepting the correction
Compare the generated table with the passage yourself. For each row, ask: does the quoted phrase appear in the source, and does it support the whole claim? In the second row, “with receipts” supports a receipt requirement, but it does not erase the separate requirement that the trip be approved.
Verify the result: Accept the corrected answer only if it preserves “may request,” “approved business travel,” receipts, “promptly,” and manager approval. Reject any revised wording that changes “may” to “must,” invents a fixed deadline, or treats approval as optional. If the answer includes a quote, locate the exact words in the source.
When checking a web answer, follow each citation and inspect the page, date, and relevant passage. ChatGPT search citations may point to incomplete, outdated, or incorrect results, so a citation alone does not establish support; OpenAI recommends checking that the linked source backs the answer. Searching the web with ChatGPT explains how to open citations and review Sources.
Recover from unsupported claims
If ChatGPT returns a polished summary instead of a claim-by-claim table, ask it to split the answer into atomic claims and include a status and exact evidence phrase for each one. If it invents a quote, deadline, or policy condition, point to the sentence it misread and ask for a corrected table using only the supplied excerpt.
If a citation leads to a page that does not support the claim, search for a more authoritative source or specify the source and relevant date in a new search. ChatGPT can sound confident while giving an incorrect fact or fabricated citation, so confidence is not a substitute for checking the evidence. Does ChatGPT tell the truth?
This method is unsuitable when the excerpt is incomplete, illegible, or too ambiguous to resolve the question; mark the claim “Not stated” and obtain the missing text or an authoritative source.
For quick current facts, you can manually select View all tools > Search in ChatGPT, when available. Search is different from generation based only on learned information: without search enabled, the answer may rely on what the model learned during training; with search, it can use web information and citations. Searching the web with ChatGPT and Does ChatGPT tell the truth? describe these distinctions.
For research that requires combining many sources, Deep Research can produce a documented report with citations, but account and workspace access may affect availability. Deep research in ChatGPT
Frequently Asked Questions (FAQ)
Q1) Does a citation prove that the answer is correct?
No. A citation gives you a source to inspect, and search results or citations can be incomplete, outdated, or incorrect. Open the page and check whether it supports the specific wording of the claim.
Q2) What if the source does not answer part of the question?
Mark that claim “Not stated” and leave it out of the corrected answer unless you find another reliable source. A missing statement does not prove the claim is false; it means this source cannot establish it. Keep the additional source separate so readers can tell which evidence supports which claim.
Q3) Should I use web search or Deep Research?
Use web search for a quick current fact and inspect the cited pages. Deep Research is designed for multi-step tasks that need evidence gathered and combined across sources, and it returns a structured report with citations or source links. Availability and connected sources can depend on your account, region, and workspace settings.