The reference that looks perfect and is not real
Ask a chatbot for five sources on almost any topic and you will usually get five. They will look right. There will be an author with a plausible name for that field, a journal that genuinely exists, a year that fits the timeline of the debate, page numbers, and quite possibly a DOI in the correct format.
Some of them will not exist.
Not "will be out of date" or "will be the wrong edition". Will not exist. No such paper, by that author, in that journal, in that year. The DOI resolves to nothing or to something unrelated. The author may be real and may work in roughly that area, which makes it worse, because the reference passes every quick sniff test you might apply.
This is the single most damaging thing that happens to students using AI for research, and it is worth understanding why it happens before we get to the fix.
Why it happens
A language model produces text that is likely to follow the text before it. When you ask for a citation, what it produces is a very good imitation of what a citation for that topic looks like. That is a different task from retrieving a record from a database, and the model is not doing the second thing unless it has actually been given a search tool and has actually used it.
So a fabricated reference is not the system failing. It is the system doing what it does, applied to a question that needed a library. The output is fluent because fluency is the thing it is good at. Accuracy about the world is a separate matter entirely.
This is why "but it sounded so specific" is not reassurance. Specificity is cheap to generate. It costs the model nothing to add page numbers.
Handing in a reference you did not open is a serious matter at most institutions, and being able to say a model suggested it is not a defence. From a marker's point of view, citing a source that does not exist and citing a source you never read are the same problem.
The one rule
Never cite anything you have not opened.
That is the whole lesson in six words, and it survives every future improvement in these tools. Even when a model has web search and gives you a working link, you still open the link, you still check the paper says what the summary claimed, and you still take the citation details from the source itself rather than from the chat window.
Opening it does not mean reading all forty pages. It means: the thing exists, the author and year match, and the sentence you are about to attribute to it is actually in there.
Checkpoint
Models can generate references with plausible authors, journals, years and DOIs that do not exist, so the rule is simple: never cite anything you have not opened yourself.
Ask for search terms, not for sources
Here is the reframe that makes AI genuinely useful for research rather than dangerous. Do not ask it for references. Ask it for the words you should be searching with, and then do the searching in a real database, your library, or an academic search engine.
This plays to its actual strength. A model is very good at knowing that your everyday phrasing of a question maps onto a technical term you had never heard of, and that term is what unlocks the literature.
โ Weak prompt
Prompt
Give me 8 peer reviewed sources on the effect of remote working on team communication, with full references.
Output
A tidy numbered list of eight references, correctly formatted, with authors, journals, volumes, page ranges and DOIs.
Some of these will not exist, and you cannot tell which by looking. Worse, the list feels like progress, so you stop searching and start writing. This is how a bibliography full of ghosts gets built.
โ Good prompt
Prompt
I am researching how remote working affects communication inside teams. I am new to this field. Give me the technical terms and phrases researchers actually use for this, any named theories or frameworks I should know, the disciplines that study it, and six search queries I could run in a library database. Do not give me any references.
Output
A list of field-specific terms, a few named frameworks, the disciplines involved, and six search strings combining them.
Everything here is checkable the moment you search, and nothing can be silently invented in a way that ends up in your bibliography. You still do the finding, which is the part that has to be yours.
I am researching this topic: [describe it in your own everyday words].
It is for [an essay / a dissertation chapter / a work report] at [level].
Give me:
- The technical terms and phrases specialists in this field use for it
- Any named theories, models or frameworks I should know the names of
- Which academic disciplines study this, and how their angles differ
- Six search queries I could paste into a library database
- Three narrower versions of my question and three broader ones
Do not give me any references, citations or reading lists.
I will find the sources myself.
The other good use: explaining a paper you already have
The second safe use is the mirror image of the first. Once you have a paper in front of you, obtained from a real source, AI is excellent at helping you get through it. You are no longer asking it to know things about the world. You are asking it to work with a document you supplied, which is a much easier and much more reliable job.
Ask it what the paper's central claim is. Ask what the method was, in plain English. Ask which parts are the authors' own findings and which are summaries of other people's. Ask what a specific paragraph means. Then check the important answers against the paper, because summaries can still drift, and a summary that quietly overstates a finding will make you write something the source does not support.
Here is a paper I am reading. I am [level, and what I already know].
- State the central claim in two sentences, plainly.
- Describe what they actually did, without jargon.
- Separate their own findings from what they are reporting from others.
- List the limitations the authors themselves admit to.
- Quote the exact sentences that support point 1, so I can find them.
If anything is unclear or missing from what I pasted, say so
rather than filling the gap.
Paper:
[paste it]
Point five is the useful one. Asking for the exact supporting sentences gives you something you can check in seconds, and it makes drift obvious.
What to do when you have already got a list
If you have a list of references from a chat window, treat it as a list of guesses. Take each one and search for it directly in your library catalogue or an academic search engine. Search the title first, then the author. If neither returns anything, do not spend twenty minutes hunting. Assume it does not exist and move on. Then, for the ones that survive, open them and confirm they say what you were told they say.
A note on links. A working link is evidence the page exists, not evidence it says what the summary claimed, and not evidence that the source is any good. Provenance still matters. If you want to go deeper on checking sources and claims, AI Safety, Privacy and Verification covers it properly.
Checkpoint
Use AI for search terms and for explaining papers you already have, and do the actual finding yourself in a real database.
๐ Quiz
Question 1 of 3A model gives you a reference with a real-sounding author, a journal that genuinely exists and a correctly formatted DOI. What can you conclude?