How to Check Whether an AI Citation Supports the Claim
A real link can accompany an overstated conclusion. Follow a writing-study example to check the source, measure, scope, and wording before reusing a claim.
An AI-written proposal says, “Research proves AI makes every employee 40% more productive,” followed by a link to a real university website. Opening the link is only the first check. You also need to establish whether the page supports that sentence.
The sentence above is an invented overclaim for this exercise, not an output attributed to a particular assistant. Its problem is useful to practice: an authentic source, a real number and an unsupported generalization can appear together.
Start with the sentence you intend to reuse
Copy the complete claim into a note, including words such as “every,” “proves,” “average” and “up to.” These words determine what evidence would be needed. Do not reduce the check to whether the page contains “AI” and “40%.”
For the exercise, open MIT's 2023 report on a writing experiment by Shakked Noy and Whitney Zhang. Check the publisher, date and study being discussed before searching for the number. A page describing another experiment is not interchangeable merely because it reports a similar percentage.
Separate three questions: Does the source exist? Does it contain relevant evidence? Does that evidence justify the proposed wording? A successful link check answers only the first.
Find the measure and read its surrounding explanation
MIT reports an experiment involving 453 college-educated professionals doing occupation-specific writing tasks. Access to ChatGPT reduced task completion time by 40% and improved evaluated output quality by 18%. The report also explains that the assignments did not require the company-specific context or precise factual accuracy that some real work needs.
That is narrower than a claim about every employee's overall productivity. In particular, less time on selected tasks and more productive across an entire job are different measures. A number cannot move between them simply because both sound like improvement.
Use a small comparison note:
| Part of the proposed claim | What to check |
|---|---|
| “Every employee” | Who participated, and whether the result is an average |
| “More productive” | What was measured: time, rated quality, output volume, or something else |
| “40%” | Which measure and comparison the percentage belongs to |
| “Proves” | Whether the wording goes beyond the study's design and scope |
The percentage is a starting point for inspection, not the whole evidence. Read the paragraph before and after the relevant passage, then look for a methods or limitations section when your decision depends on it.
Write only what the inspected source supports
A narrower sentence is:
MIT reported that access to ChatGPT reduced completion time by 40% in a 2023 experiment on selected professional writing tasks.
Keep the source link beside that sentence. This wording identifies the account used and limits the result to the experiment. It does not forecast savings for your organization or establish how current models compare.
For an internal adoption proposal, the next step would be to test your own task, including checking and revision time. You do not need to discard a useful study because its scope is limited. You need to avoid making it answer a question it did not investigate.
This walkthrough uses the publicly readable MIT account, checked on 14 September 2026. It does not claim a review of the full journal paper. If the underlying paper is necessary for your decision, obtain it and inspect the relevant methods rather than silently treating a news summary as the full study.
Ask AI to help locate evidence, then inspect it yourself
AI can help separate a compound claim and identify passages to read. Give it the exact claim and accessible source text:
Check whether this source supports the claim below.
Split the claim into separate factual parts.
For each part, identify the relevant source section and a short passage.
Mark it supported, partly supported, contradicted, or not established.
Explain any difference in population, task, measure, date, or certainty.
Propose a narrower sentence using only the supplied source.
If you cannot read the source, say so. Do not infer its content from its title.
Claim: [paste the complete sentence]
Source: [paste the accessible text and its URL]
The returned passage is another item to verify. Find it in the original and read around it. A plausible quotation that you cannot locate does not close the check. Nor does a second assistant agreeing with the first.
Keep the final sentence, URL, section or page, access date, and reason for your judgment together. If the evidence only partly supports the sentence, narrow the wording. If the source cannot be read, mark the claim unresolved and keep it out of the proposal until you can verify it.
