Keywordshow to cite sources when using aiBlogai for workai at workworkplace aiai for professionalshow to use ai at workai productivity
Related searcheshow to cite chatgptciting ai at workchatgpt fake citationshow to cite sources with aiai citation honesty at workhow to cite sources when using ai
Cite what you opened, not what the model performed
When you use AI for research or drafting at work, the citation rule is older than the tools. You cite the human sources you actually opened. You do not cite the model as if it were a paper, and you do not repeat a URL you have not visited. ChatGPT can produce a tidy reference list that looks like a literature review. Titles sound real. Years sit in the right place. Links have the shape of links. That is performance. Honesty is a record of what you read, what you checked, and what still sits on your desk as an unverified claim.
Fake citations are not a rare glitch. They are a predictable product of a system that is good at the style of scholarship and indifferent to whether a document exists. A board paper that leans on three invented studies is not slightly messy. It is a false trail that the next person will waste hours chasing. A client memo that cites a URL which 404s is a trust leak. Your name is on the artefact. The model will not attend the meeting where someone asks for the PDF. You will.
Some workplaces also want a disclosure that AI assisted the draft. That is a process note, not a source. It tells a reader how the text was produced. It does not replace the books, filings, datasets, or interviews that carry the facts. If you cannot point to a page you opened, you do not have a citation. You have a sentence that still needs work. This essay is the method for keeping those two things apart when the model is eager to complete the bibliography for you. Completeness is not honesty.
If you have not opened the source, you may not cite it. A fluent reference list is not evidence that the works exist. Visit the page before you footnote it.
What counts as a source when a model helped
A source is a work you can retrieve and that actually supports the claim. Filings, books, journal articles you opened, internal reports you are allowed to use, datasets with a name and a date, and interviews you conducted all count. ChatGPT does not count, any more than a colleague who sketched a structure on a whiteboard counts as a publication. If your organisation requires you to log AI assistance, log it beside the artefact as method, the way you might log that a spreadsheet calculated a total. Method is not bibliography.
URLs need a visit. The model will happily emit a path that looks like a government domain or a famous journal. Open it. Check the title, the date, and whether the paragraph you need is there. If the link is dead, the citation is dead. If the page exists but does not say what the draft claims, you have a hallucination with a reference attached. Quote or paraphrase only after that check. Secondary summaries of summaries are how errors become house style. Go back to the primary page once before you quote.
| You want to mention | Cite it as | Do not |
|---|---|---|
| A paper or report you opened | The work, with locator | Cite the chat that found the title |
| A figure from a workbook | The file, sheet, and date | Cite the model that described the figure |
| AI help on structure or tone | A process note if policy asks | List the model as a scholarly source |
| A URL the model proposed | Only after you visit and confirm | Paste the link unseen into a paper |
Build the reference list from the open tab, not the chat
Start with the question and the pack you already have. Ask the model for search terms, likely authors, or gaps in your outline. Do not ask it for a finished bibliography. When it suggests a title, treat that as a lead: find the work in a library, a database, or the public site, then read it. Only then add a citation in the style your organisation uses. If the lead cannot be found in a few careful searches, drop it. Invented works are not near misses. They are noise you must not recycle.
For internal work, cite the internal object: the policy version, the ticket, the deck date. For external work, use the same standard you would use without a model. Keep a two-column note as you draft: claim, source locator. Anything in the left column without a right column stays marked missing. Ask the model to list claims that lack a source you pasted. That list is useful. A generated Harvard list is not. You still type the references from the documents you opened. Typing is how you notice the gap before you format it.
Role: You are a research assistant who never invents sources.
Task: Help me plan reading for the question I state.
Context: I will paste notes I already have and the claim I must support.
Constraints:
- Do not invent titles, authors, years, or URLs.
- If you are not sure a work exists, say so and suggest a search query instead.
- Separate facts in my notes from inferences.
- Mark every claim that still needs a source I have not pasted.
Output: Search terms, questions to answer, and a list of missing sources.
Quality checks: Which sentences would be dishonest to publish without an open document?Save the two-column note with the draft. When a reviewer asks where a number lives, you open a file, not a chat history. If policy requires an AI-use line, add it once, in plain language, without dressing the model up as literature. The habit you want is boring: leads from the model, citations from the page, and nothing in between that looks like a library. That boredom is the point of the method.
Plausible references are a faster way to be wrong
The failure mode is speed dressed as diligence. A draft with twelve footnotes looks finished. Three of them may point at nothing. A reader who trusts the notes will repeat the error in a pack that goes to a board or a client. Another failure is citing the model because a style guide has no row for it. Silence is better than a fake academic entry. A third is quoting a statistic the model attributed to a famous agency. Until you open the agency's table, you are repeating a rumour. Attribution theatre is still theatre.
Watch copy-forward. One invented citation in a shared note becomes three decks by Friday. Make the rule visible: no reference without an open source, and no URL without a visit. If someone hands you a chat-generated reading list, treat it as homework, not as a library. Your reputation is the accumulated trail of claims other people could verify. A model cannot carry that trail for you. If you pass an unopened link to a colleague, you have multiplied the error on purpose. Stop it at the first note. Do not wait until the pack is already in circulation.
| Mistake | What it looks like | What to do instead |
|---|---|---|
| Model as paper | ChatGPT, 2024, in the reference list | Process note, plus human sources |
| Unvisited URL | A neat link nobody opened | Visit, confirm, then cite |
| Invented study | A plausible title and year | Search, or delete the claim |
| Copy-forward notes | Footnotes reused from a chat | Rebuild from open files |
A reference you have not opened is not a citation. It is a claim that still belongs in the missing column. Open the document or delete the note.
Related reading on StudyGrid
Read next: How to Fact-Check ChatGPT AI for Research at Work Where AI Falls Short. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.
What to do this week
Take one artefact you will send that uses sources. Delete every citation you cannot open in two minutes. Visit every remaining URL. Rebuild the list from those documents. Add a one-line process note if policy requires it. Keep the two-column claim-and-locator sheet and use it on the next draft so the habit is the method, not a rescue after a reviewer asks. Share the rebuilt list with whoever signs the artefact.