A practical source-checking guide
Does the source cited by AI exist? Check the link and quotation
When preparing work or study material in Armenia, a link in an AI answer can look like proof. But the existence of a page and support for a particular claim are separate questions. We compare one fictional claim with a real official page and keep a record of the decision.
Isolate one claim, open the original source, check its title and publisher, and find the passage that is supposed to support it. Record the URL, checking date, relevant meaning and decision. If the page cannot be opened or the passage cannot be located, do not present the claim as verified.
Start with one precise claim
Our invented teaching claim is: “An AI answer containing a source link no longer needs human verification.” We attach the OpenAI Safety best practices page as a fictional attribution. This is an authored test candidate, not an actual ChatGPT response and not a quotation from OpenAI.
The question is limited: does that document give this advice? One page cannot measure the accuracy of every model or establish that every AI answer is wrong. Separating the claim first gives us a specific relationship to check: a statement and its alleged source.
Open the page and establish its identity
On 28 October 2026, we opened https://developers.openai.com/api/docs/guides/safety-best-practices. A public GET returned HTTP 200 and the final URL stayed the same. Reading the official document confirmed the title Safety best practices and the Human in the loop (HITL) section.
The publisher is OpenAI’s official documentation site. An individual author and an update date were not established in the retrieved document; we do not fill those fields by guessing. Our checking date records our observation, not the page’s publication date.
Build an evidence card around the relevant passage
The relevant section contains the short phrase “human review outputs”. It recommends human review before outputs are used in practice wherever possible, with access to the original information needed for verification. It particularly stresses high-stakes applications and code generation.
Our card preserves the route to that meaning, rather than only the link. Another reader can return to the same section and examine the conclusion.
Go to the human-review section
- Claim: having a link removes the need for human verification. This is our fictional teaching claim.
- Source title: Safety best practices.
- Publisher: OpenAI Docs. Individual author: not named in the retrieved document.
- URL: https://developers.openai.com/api/docs/guides/safety-best-practices.
- Checked: 2026-10-28.
- Location: Human in the loop (HITL) section.
- Evidence meaning: a human reviews outputs wherever possible and can compare them with the original material.
- Decision: reject the claim and attribution; the source advises the opposite.
Record the decision and a supported replacement
The page exists and is official, but that does not support our fictional sentence. The claim said review was unnecessary; the retrieved passage recommends human review. We therefore do not publish the sentence as advice or a quotation from OpenAI.
A supported replacement is: “OpenAI’s API safety guide recommends human involvement in reviewing outputs wherever possible.” This is a narrow paraphrase, not a direct quotation. Guidance for API applications does not establish a universal legal requirement for all ChatGPT users.
Keep accept, reject and unverified separate
Accept a narrow claim when the relevant meaning is present in the source and its conditions are preserved. Reject an incorrect attribution or a statement contradicted by the passage, as in this example. Leave a claim unverified when the URL cannot be opened, the passage cannot be found or the information is insufficient.
An inaccessible page does not by itself prove that the information is false or the source never existed. A plausible title, familiar author or fluent summary does not verify a particular sentence either. For a consequential decision, seek another primary source or a qualified person’s clarification.
Save enough information to repeat the check
Reuse these blank fields: claim, source, author or publisher, URL, checking date, section, relevant meaning and reason for the decision. If the material changes, record the new checking date and wording. Seeing a search-result snippet alone is not a completed source check.
Our actual action here was opening and reading the official source. We did not run a model experiment, send an API request or establish that the source will remain unchanged. In your own records, distinguish completed actions, proposed steps and results that remain unverified.
Key takeaways
- A working link does not establish support for a claim.
- Locate the relevant passage and preserve its conditions.
- Keep an evidence card and explain the accept, reject or unverified decision.
- Never present a fictional example or an unperformed experiment as an observed result.
Frequently asked questions
Did ChatGPT produce the incorrect claim?
No. We wrote it as a fictional teaching example. The official page attributed to it was actually opened and checked.
Does HTTP 200 mean a source has been verified?
It establishes that the page was technically accessible at that moment. You still need to check its identity, relevant passage and whether it supports the meaning of the claim.
What if there is no author or date?
Record the visible publisher and state which fields were not established. Keep your checking date separate from the source’s publication date.
Can I simply ask AI whether its source is correct?
You can ask it to clarify the reference, but a second answer is not independent evidence. Open the source and compare the relevant passage with the claim.
Source
Checked: