How to use AI in your thesis without fake citations

Two uses of AI in academic work sound alike and are completely different. One is a reading tool. The other is how references nobody can find end up in a bibliography.

Where the line is

The tool doesn't decide it. A search engine, a reference manager and a language model all sit between you and a source. The line is whether the source exists and whether you read it.

On the defensible side:

• Finding which chapter of a 600-page book discusses a concept.
• Comparing how three authors define the same term, with each passage in view.
• Summarizing a paper you already read to check that you understood the argument.
• Sorting forty papers by methodological position.

On the side that causes trouble:

• Asking for "five references on this topic" and pasting them into your bibliography.
• Attributing to an author a position the system described, without opening the text.
• Using a generated summary as if it were a reading of the original.

Why references that don't exist show up

A bibliographic reference is a very regular pattern: author, year, title, journal, volume, pages. A language model learns that shape very precisely from millions of examples.

What it doesn't learn is which specific combination matches a real work. When you ask for references on a topic, it produces plausible combinations: authors who really publish in that area, journals that really exist, sensible years, titles that sound like papers in that field. The result has a valid format and invented content.

That is why asking the same system to confirm the reference exists doesn't work. It generates another prediction, not a check.

A method that holds up

1. Gather first, ask second

   Download the PDFs you are going to use. Work on that closed set, not on a model's memory. Everything you cite should come from something you have.

2. Require a location in every answer

   Each statement should come with a document and a page. If you can't reach the passage, don't cite it.

3. Read the passage before citing it

   A summary can be faithful or it can shift the nuance. From the summary alone you can't tell.

4. Check the reference in a catalog

   Library, journal repository or DOI. If it only exists when you ask for it, it doesn't exist.

5. Disclose your use according to your institution's policy

   Rules vary by university and program. Yours is the one that applies.

The question that shows how the system is working

Ask it something that is not in your documents.

If it answers with its usual confidence, it is generating from memory and you shouldn't cite it. If it says it couldn't find that in your material, it is reading what you gave it, and that is the minimum condition for its answers to be verifiable.

Where the time savings are

The heavy part of a literature review isn't writing. It is locating: finding where in forty texts a concept is discussed, how they contradict each other and what evidence each one brings.

That is where a tool that reads your own corpus and points to the exact page changes the real working time, without touching the part where you have to read, evaluate and decide.