How many references have you clicked on given by AI?

This article was originally published in “Vista Newsletter (opens in a new tab)” No.149

Hello, hello!

I have observed that these days when I am taking classes at universities and research institutions, the questions I am asked after class have quietly changed to a different way of asking them.

Two years ago, everyone asked: “Teacher, can I use AI to write papers?” Nowadays, what is more commonly heard is: “Teacher, how do I know that the references given to me by AI are true?”

I like this question because it gets to the point. Judging from several foreign studies and news this year, it seems that false citations are no longer an embarrassing incident caused by a graduate student, but a troublesome problem that causes headaches for the entire academic publishing system.

In this e-newsletter, I would like to first share with you some foreign intelligence that I have read recently, and then talk about the most common mistakes I have seen in AI research work after conducting 232 actual tests in the past two months. In the middle, you will be given two prompt words that can be copied directly. Finally, I will also introduce to you the two academic research manuals I just completed.

Intelligence 1: False literature has increased 12 times in two years, and 98% of it has not been dealt with yet.

In May this year, “The Lancet” published a research report (opens in a new tab). The research team checked nearly 2.5 million papers (opens in a new tab) included in PubMed, comparing references one by one to see if they actually existed.

The result: approximately 1 in 277 papers published in the first seven weeks of 2026 cited non-existent literature. This ratio increased 12 times in two years.

What is even more concerning is that for more than 98% of the papers found to contain fake documents, the publisher had not taken any action at the time of the inspection. In other words, these papers are still sitting in the database, waiting to be cited by the next graduate student. (Report by Retraction Watch (opens in a new tab))

Another study zoomed in on “using AI to write a literature review.” An experiment published in JMIR Mental Health last year asked GPT-4o to write 6 literature reviews on mental health topics, resulting in a total of 176 citations (opens in a new tab). The researchers checked Google Scholar, Scopus, PubMed and other databases one by one and found:

• 35 (19.9%) were fabricated and no corresponding publications could be found
• Of the remaining 141 truly existing documents, 64 (45.4%) have errors in their bibliographic information. The most common ones are DOI errors or invalid ones.
• The less familiar the topic is, the higher the rate of fabrication

Taken together, nearly two-thirds of the references cannot be used directly. (Original paper (opens in a new tab))

This last point is particularly relevant to graduate students. Because the topic of your paper usually happens to be a relatively unpopular one.

Of course, this was an experiment done with GPT-4o in June 2025, and the new model may perform better. But judging from my own actual measurement experience, model improvement will make errors fewer, but it will not make them disappear. And when there are fewer mistakes, people are more likely to let down their guard.

Intelligence 2: From shame to institutional punishment

The consequences of false citations have also been rapidly “escalating” in the past two years.

The first thing that attracted attention was the https://stli.iii.org.tw/article-detail.aspx?tp=1&d=9050&no=64 (opens in a new tab) case in the United States in 2023 [Mata v. Avianca]. The lawyer wrote the precedent compiled by ChatGPT into the complaint and was fined $5,000 by the court. At the time, many people regarded it as a joke. Today, three years later, in the database (opens in a new tab) compiled by legal scholar Damien Charlotin, there have been 2,190 cases in which courts around the world have found that AI was used to fabricate content.

Government reporting and consulting firms are not immune. In October last year, a report written by Deloitte for the Australian government was found to have cited non-existent document (opens in a new tab), as well as fabricated quotes from judges. The revised version acknowledged the use of GPT-4o, and Deloitte eventually refunded part of the money.

It is conceivable that the academic circle has also begun to take action seriously. In May of this year, Nature reported that the preprint platform arXiv would suspend authors who cited AI hallucination documents. (Report by Nature (opens in a new tab))

In Taiwan, the Academia Sinica released the “Reference Guidelines for Using Generative AI (opens in a new tab)” earlier this year. The first of the three core principles is to manually check all citation sources.

The rules are becoming clearer and the consequences are becoming more serious. For researchers, there is little need to discuss whether to use AI; but how to prove that they use it responsibly is the next thing to face.

Intelligence 3: Everyone is using it, but only 20% trust it

Elsevier’s global researcher survey, released in November last year, interviewed more than 3,000 researchers: 58% were already using AI tools in their work, compared with only 37% the year before. But in the same survey, only 22% believed that AI tools are currently trustworthy. (Investigation Press Release (opens in a new tab))

More and more people are using it, but trust cannot keep up. This gap is probably the most real situation of researchers today: if they don’t use it, they’re afraid of not being able to keep up; if they use it, they don’t know what will happen.

This is why tools like Gemini Notebook (formerly NotebookLM, Google officially changed its name in July this year) are particularly popular in academic circles: because it only answers based on the data you upload.

There was a preprint study last year that tested 100 fictional pancreatic cancer cases. Gemini Notebook uses reference files to determine cancer staging, and the accuracy is 70%; similarly, when the reference files are given to its underlying model, Gemini 2.0 Flash, for direct reading, the accuracy is only 38%. The original text listed in Gemini Notebook is also 92% accurate. (Paper Abstract (opens in a new tab))

Well, the difference is considerable. However, the other side of 70% is that there are still 30% misinterpretations. Qualifying the source can greatly reduce the number of fudges, but it does not guarantee that it will read correctly.

I saw this very clearly during my own actual measurement.

My actual measurement: The most common mistakes AI makes are often not fabrications.

In the past two months, I have conducted two sets of practical tests.

The first set is general AI dialogue. I used 7 published Chinese and English journal articles as material and tested 30 research prompt words, 120 times in four rounds. The second group switched to Gemini Notebook, using the same 7 papers, plus 4 self-drafted and de-identified verbatim interviews, for a total of 112 interviews in four rounds.

Both groups of actual measurements used the same method: deliberately burying errors in the materials. For example, write the wrong year, refer to theoretical articles as “confirmations”, and alter the quotes of interviewees to see if the AI ​​will be deceived. Each answer will be handed over to a reviewer who did not participate in writing the prompt words, and will be checked sentence by sentence against the original text.

The results surprised me a bit. The most common problem is not fabricating a paper that does not exist, but the following five relatively difficult-to-detect errors.

1. Overstating the tone. The original text reads “discussion”, and AI changes it to “confirmation”; the original text writes “majority”, and AI writes “all”. Once, the original civilization was clearly insignificant in the first semester and only significant in the second semester, and AI was written as stable and significant in both semesters. Everywhere seems to be just a wording difference, and when put into the paper, it becomes an exaggeration of the research findings.

2. Describing the method more specifically than the source. The original only said it “asked 15 industry practitioners,” but the AI wrote that it “interviewed 15 industry practitioners.” In literature review, inquiry and interview are two different research methods. When the oral examiner looks back at the original text, he or she will find that the author did not write it that way at all in the research design you described.

3. Attributing other people’s words to the author. In the literature review of a paper, there are a large number of other people’s research reported by the author. AI can easily regard these as the author’s own findings and ask you to write “Wang Daming (2024) pointed out”, but that sentence is actually Wang Daming paraphrasing the research of Li Xiaohua (2019).

4. Quotation marks do not guarantee original wording. In actual testing, AI has changed words within quotation marks, added English names that were not in the original text, and has also put its own rewriting directly into quotation marks.

5. It won’t tell you which files it couldn’t read. This was my most unexpected discovery on Gemini Notebook. There are 5 sources in the notebook, one of which is a scanned file without text layer. Ask it to list all the sources it can see, and it only lists 4, and it definitely says they can all be read. You thought the literature review covered 5 articles, but in fact it only covered 4 articles.

These five mistakes have one thing in common: they are all attached to a reliable-looking exterior. There is a citation number, quotation marks, author and year after the sentence, and you will naturally believe it.

On the other hand, AI also does something unexpectedly well. Gemini Notebook extracted 120 original sentences from the interviewees from the verbatim drafts, of which 119 were correct verbatim. I gave it three quotations, one was a correct excerpt, one was altered, and one was completely fabricated, and it got all three of them correct. I buried 7 errors in a draft of a literature review and asked it to check the citations sentence by sentence, and all 7 were caught.

What these things have in common is taking a piece of text and comparing it to the source. So my current principle of using AI for research is simple: let it compare, don’t let it draw conclusions.

Three principles can help you avoid half of the troubles

1. Give materials, don’t give empty questions

“Help me find literature on the digital gap” is the easiest way for AI to fabricate non-existent papers out of thin air. Please paste in the summary, notes or drafts you have on hand and ask it to answer only based on what you posted. If you want to find new documents, please go back to the academic database to search.

2. Ask it to mark areas of uncertainty

In the prompt words, ask the AI to mark the places where you are unsure or lack information with a mark such as [to be investigated] or [not mentioned in the original text]. Please don’t find it troublesome when you see these marks, as they will save you time from being caught by the oral examination committee. Where there are no marks, it does not necessarily mean that they are correct.

3. Leave the comparison to AI and leave the judgment to yourself

AI can help you organize information, conduct preliminary screening, compare original texts, and highlight doubtful points. As for how to determine the research question, how to interpret the data, and which document is worth citing, it must be your own judgment. In the end, it is you who is responsible for signing the paper.

I’ll give you two reminder words

The following two items are revised versions after actual testing and can be copied directly. Please replace the spaces enclosed by [] with your own content; [] is a mark that requires AI to write the answer as usual, without replacement.

After writing the paper: Compare in-text citations and reference lists

Suitable for any AI tool, ChatGPT, Gemini, Claude or the version provided by your school.

The following is the text of my paper (or one of its chapters):
[Paste the content]

Here is my reference list:
[Attach reference list]

Please help me compare:
1. There are citations in the text but not in the reference list.
2. There is a reference list but no citations in the text.
3. The author’s name or year in the text and list of the same document are inconsistent

Only make comparisons, do not modify or add any bibliographic information. 

The last sentence is important. Without it, AI will often add the author, year or volume for you, but the information added may not be correct. This article is only responsible for comparison. The bibliographic information of each reference still needs to be checked using the DOI or database.

After writing the literature review using Gemini Notebook: citation backchecking

First upload the paper you cited to the same notebook, and then paste this prompt. I used this one to catch all the 7 buried mistakes mentioned earlier.

The following is my literature review paragraph:
[Paste paragraph]

Please check sentence by sentence and list it in a table:
1. Sentence
2. The source mentioned in the sentence, and the nature of the source: empirical study, theoretical article, manifesto, or review
3. Excerpt the original text from the source that supports this sentence verbatim and attach quotations; list the excerpt together with the concessions or conclusion clauses before and after, do not just cut off half the sentence; the excerpt must directly support or refute this sentence, if you cannot find it, write [source not mentioned]
4. Judgment: Complete support, partial support, too strong tone, no basis found, choose one of the four, other situations are written in the description

Rules:
- Theoretical articles or declarations cannot "prove" any effect. If they encounter "confirmation" or "proof", they will be deemed to be too strong.
- For the research methods (interviews, questionnaires, experiments), samples or years mentioned in the sentence, check them in the sources one by one. If you cannot find the corresponding words, also mark them.
- For sentences judged to be "no basis" or "too strong", please explain the difference.

Please pay special attention to one thing when using it: Is the passage pointed to by the quotation the research finding of this paper? If it refers to literature discussion, that sentence actually refers to someone else’s research, and the paper should be changed to cite the original literature.

Two manuals: Organizing the entire method into a process to follow

The actual measurement results and the two prompt words above are all from the two manuals I completed in the past two months.

“AI Academic Research Prompt Manual (opens in a new tab)”

It contains 30 research prompts and is divided into six chapters according to the research process: title and research questions, literature search and reading, literature review writing, research design and methods, revision and submission, as well as review responses, oral examination and research ethics.

Each item is divided into three parts: when to use it, prompt words that can be copied directly, and verification reminders, which are the most common mistakes made by AI in this link and how you should check them. These prompt words are not bound to any AI tool.

《Gemini Notebook Research Workflow Manual (opens in a new tab)》

There are 8 workflows and 36 steps, starting from the establishment of a research notebook, all the way to in-depth deconstruction of a single paper, cross-reference of multiple documents, from research gaps to research questions, coding of verbatim transcripts of interviews and field notes, literature review writing and citation checking, and finally oral examination preparation and review responses. Each step is accompanied by prompt words and verification reminders, as well as an implementation asset package, including a prompt word library and customized command templates that can be directly copied.

The division of labor between the two books is very clear. The first one is to ask one question at a time, suitable for any AI tool, to handle scattered single tasks in the research process; the second one is to read a batch at a time, put the entire batch of documents or verbatim manuscripts into a notebook, and do cross-reference and systematic extraction.

There is one thing I want to point out in particular.

The appendices of both manuals fully list the known limitations found in actual measurements, with 19 in the first and 31 in the second. For example, which prompt words have only been run once, which ones have errors twice in three actual tests, and why the AI ​​word count control cannot be fixed no matter how hard it is fixed. In the end, it can only be left to the readers to check by themselves.

I write this because I believe that a manual that teaches people to check AI should also be able to withstand the review itself.

I also want to make it clear first, what are these two manuals not?

They do not teach you how to use AI to write papers, nor do they guarantee that you will not make mistakes if you use it. What they do is let you know where AI is most likely to make mistakes in every aspect of research, and how to check them one by one.

Combined purchase of two manuals

• “AI Academic Research Prompt Manual”: starting price NT$499 (original price NT$699), PDF e-book
• “Gemini Notebook Research Workflow Manual”: starting price NT$699 (original price NT$999), PDF manual plus implementation asset package
• Purchase two books together: NT$1,099, which is NT$99 less than the original price of purchasing two books separately.
• Both books are updated for free within one year
• Pay through Portaly and an electronic invoice will be issued for each transaction

The initial price is limited to the first 50 copies each. The combined purchase is only available on vista.tw’s store page, and individual copies can also be purchased on the same page:

Look at the dual manual purchase

These two manuals are more suitable for the following types of friends:

• Graduate student who is writing a thesis: I can’t finish reading the literature, and I am worried that the content provided by AI is unreliable.
• People who are already using Gemini Notebook: Not sure if its answers can be written directly into the paper
• The teacher who supervises the thesis: I hope that students will have a common set of assessment standards when using AI.

If you want to see how I measured it first, there are two free articles on the researcher base that you can read first: “Why does AI fabricate literature and change “exploration” to “confirmation”?” (opens in a new tab)〉 collates the first set of experimental findings, and 〈Gemini Notebook Academic Usage Guide (opens in a new tab)〉 explains what this tool can and cannot do.

Write at the end

Back to the question at the beginning: How do we know that the references provided by AI are true?

My answer may be a little disappointing: no tool can guarantee that for you. All you can do is check them one by one, so that every quote can be traced back to the original text, and every statement passed through your own eyes.

AI can indeed save a lot of mechanical labor in sorting and comparing, but the time it saves is best used to do things that only you can do: think clearly about research problems, understand the meaning behind the data, and then be responsible for every sentence you write.

If you have friends, students or lab partners who are writing papers, you are also welcome to forward this newsletter to them. After using these two prompt words, if you have any thoughts or stuck points, you are welcome to write back to me directly and I will read them carefully one by one.

It’s inevitable that you’ll feel a bit lonely while studying this path, so let’s work together!

Sincerely, Vista