How Do You Use AI for Citations Without Inventing Sources?

TL;DR:

• Chatbots such as ChatGPT can invent references that look real. Published studies found fabrication rates from about 18% to 55% of citations, and many of the "real" ones still had wrong details.
• The safe order: find real sources first, save them in a reference manager such as Zotero or Mendeley, read them, and only then use AI.
• Use AI to format, tidy and summarise sources you have already checked. Never ask it to supply sources for you.
• Check every DOI. Open it at doi.org or look the reference up on Crossref before it goes into your thesis.
• If you use AI in your work, say so in the format your university and referencing style require, and keep your prompts.

Who this guide is for

This guide is for Masters and PhD students, and anyone writing a proposal, thesis or journal article, with Kenyan universities in mind. You need no technical skill.

It stays on one job: getting your references right. For finding papers, see AI tools for literature review. For what is allowed overall, see using AI in postgraduate research without crossing the integrity line.

Why do chatbots invent references?

A chatbot writes by predicting the next likely words. Unless a search tool is switched on, it does not look anything up. A reference has a predictable shape (author, year, title, journal, volume, pages, DOI), so the model can produce one that looks perfect with nothing real behind it.

A September 2025 paper by Adam Tauman Kalai and colleagues, "Why Language Models Hallucinate", argues that training and testing reward guessing over admitting doubt. In their words, language models "are optimized to be good test-takers". A confident guess scores better than "I don't know", so the model guesses.

What does the published evidence say?

Peer-reviewed studies have tested how often chatbots invent references. Each tested one model at one point in time. GPT-4 did better than GPT-3.5, yet every model tested still produced fabricated or faulty references.

Note one detail from the 2025 study. Fabrication rose on narrow topics: for one disorder, specialised reviews had 46% fabricated citations against 17% for general ones. Many Kenyan theses ask a narrow, local question, so in our view the risk is higher for them.

This is already causing real damage

• A New York court, June 2023. In Mata v. Avianca, Judge P. Kevin Castel fined two lawyers and their law firm a total of USD 5,000 over filings citing six court decisions that did not exist. ChatGPT had produced them.
• An Australian government report, October 2025. Deloitte agreed to repay the final instalment of an AU$440,000 contract after its review for the Department of Employment and Workplace Relations was found to contain invented references and a made-up quote from a court judgment.
• Courts worldwide. Researcher Damien Charlotin's public database lists 2,289 court and tribunal decisions where AI-generated content, such as fake citations, was an issue (updated 10 November 2026). One is listed under Kenya.

If trained lawyers and consultants get caught, a tired student at 2 am can too.

What does a fake reference look like?

Fake references borrow real pieces: real authors on papers they never wrote, real journals with volumes that do not exist. Watch for these signs:

• The DOI does not open, or it opens a different paper.
• The title matches your topic almost too perfectly.
• The authors are real experts, but the paper is not on their publication list.
• The journal exists, but that volume and page range do not hold that article.
• You cannot find the title in Google Scholar, your library database or Crossref.
• The details change when you ask the chatbot about the same paper again.

One sign means stop and check. Two or more, and treat it as fake until you hold the paper.

The safe workflow: real sources first, AI second

This is the order we teach. AI is kept away from the job it does badly (supplying sources) and used for jobs it does well (tidying and explaining text you give it).

Step 1: Find real sources in real places

Search your university library's e-resources, Google Scholar, journal websites and Crossref's free search at search.crossref.org. The safe use of AI here is asking for search terms, then running the searches yourself.

Some AI search tools link to real papers; our literature review guide compares them. Open the source anyway. APA's 2025 guidance says the same: when AI acts as a search engine, you must still verify its sources.

Step 2: Save every source in a reference manager

A reference manager such as Zotero or Mendeley stores each source's details and writes citations for you. It invents nothing, because it only holds what you put in. Zotero's "Add Item by Identifier" button builds a full record from a DOI, ISBN, PubMed ID, arXiv ID or ADS Bibcode, which saves typing errors.

Step 3: Read the source yourself

Read at least the abstract, the method and the findings you plan to cite. If you cannot get the full text, ask your library. Never cite a paper for a claim you have not seen in it.

Step 4: Use AI only on what you have checked

Now AI is useful. Paste in details from sources you hold and ask it to format, compare or summarise. You supplied the material, so you can check every word it returns.

Step 5: Verify every DOI

A DOI (Digital Object Identifier) is a permanent link to one specific work. Put https://doi.org/ in front of it and open it. It should land on the exact paper you cited.

For a whole list, Crossref's free Simple Text Query tool (apps.crossref.org/SimpleTextQuery) takes up to 1,000 pasted references and tries to match each to a DOI. Check any reference without a match by hand. Many real reports and older books have no DOI, so a missing match alone does not prove a fake.

Zotero or Mendeley: which should you use?

Both do the core job well. Pick the one your supervisor or co-authors already use, because sharing libraries is easier on the same tool. Facts below are from the official product pages as of November 2026; the last row is our opinion.

One Zotero feature helps thesis writers: since 2019 it flags retracted items in your library, using Retraction Watch data, and warns you before you cite one. It works only for items with a DOI or PubMed ID.

Copyable prompts that keep AI honest

Each prompt gives AI material you have already checked and forbids it from adding anything. Replace the square-bracket parts.

1. Format verified references into APA 7

Format the references below in APA 7th edition style. I have checked every detail against the original source. Rules: do not add, remove or change any author, year, title, journal, volume, issue, page or DOI. If a detail APA needs is missing, write [MISSING] in its place. Do not suggest extra sources. References: [paste the details exported from Zotero or Mendeley]

2. Switch from APA to Harvard

Convert this APA 7 reference list into Cite Them Right Harvard style. Keep every detail exactly as given. Add "(Accessed: [date])" after each URL or DOI, using the date I give. List anything you were unsure how to convert at the end. List: [paste]

3. Match in-text citations against the reference list

Below are my chapter text and my reference list. List (1) every in-text citation that has no matching reference, (2) every reference never cited in the text, and (3) any citation where the year or spelling of the author differs between the two. Do not fix anything. Just list the problems. Chapter: [paste] References: [paste]

4. Summarise a paper you have open

Here is the abstract and findings section of a paper I am reading: [paste]. Summarise it in five bullet points for my literature review notes. Use only what is in this text. If something I might want (sample size, country, method) is not stated, write "not stated". Do not add any other studies.

5. Ask for search terms only

My research question is: [question]. Suggest 10 search terms and 3 Boolean search strings I can use in Google Scholar and my university library database. Do not name any papers, authors or references.

APA 7 and Harvard examples

Here is one real article in both styles. We checked its details in Crossref on 11 November 2026.

APA 7th edition (reference list):

Borden, S. L., & Tew, C. (2007). The role of journalist and the performance of journalism: Ethical lessons from "fake" news (seriously). Journal of Mass Media Ethics, 22(4), 300–314. https://doi.org/10.1080/08900520701583586

Cite Them Right Harvard (reference list):

Borden, S.L. and Tew, C. (2007) 'The role of journalist and the performance of journalism: ethical lessons from "fake" news (seriously)', Journal of Mass Media Ethics, 22(4), pp. 300–314. Available at: https://doi.org/10.1080/08900520701583586 (Accessed: 11 November 2026).

In-text citations:

The main differences: APA uses "&" inside brackets and a full stop after the year. Harvard puts the article title in quotation marks, writes "pp." before pages and adds the date you accessed an online source. In an APA reference list, name up to 20 authors; with 21 or more, list the first 19, an ellipsis, then the final author. In a Cite Them Right Harvard reference list, name every author, and use et al. only when there are more than 10. Whichever style you use, use only that one.

How do you cite AI itself?

You cite AI when it shaped your work and your university allows that use. APA and Harvard handle this differently, so follow your school's style.

APA Style (updated September 2025)

APA updated its 2023 ChatGPT guidance with a three-part series on its Style Blog on 9 September 2025. The key points, as summarised by the University of British Columbia Library:

• The author is the company that makes the tool. So the in-text citation is (OpenAI, 2025) or OpenAI (2025).
• A general reference to a tool: OpenAI. (2025). ChatGPT (GPT-5) [Large language model]. https://chatgpt.com/
• A specific chat you can share by link: AI company. (Year, Month Day). Title of chat [Generative AI chat]. Tool name or model. URL of the shared chat.
• AI output is not a personal communication in APA, because no person wrote it.
• Keep your prompts. Long AI responses can go in an appendix or supplementary file.
• You do not cite AI built into everyday software, such as Copilot inside Word, which APA treats like spellcheck.
• AI is never an author, and you remain responsible for the accuracy of anything it produced.

Cite Them Right Harvard

Cite Them Right differs. If the AI output is only available to you (a private chat), you cite it as a personal communication and describe it in your text. The form is: Name of AI (Year of communication) Medium of communication to Receiver of communication, Day Month of communication.

For example: OpenAI ChatGPT (2026) ChatGPT response to Jane Wanjiru, 4 November.

If the AI material is published online, you cite it like that medium, with the AI company as the author.

A sample disclosure statement

A citation shows where text came from. A disclosure statement shows how you worked. Here is a sample to adapt:

"I used ChatGPT (GPT-5, OpenAI) to format my reference list in APA 7th edition and to check my in-text citations against it. I did not use AI to find sources, analyse data or write findings. I checked every AI output against the original sources. My prompts are available on request."

Write only what is true. A vague or inflated statement is its own integrity problem.

What do Kenyan universities and supervisors expect?

Rules differ by university, school and supervisor, and are still forming. This is what we could verify as of November 2026.

• There is no binding national rule yet. The Commission for University Education (CUE) convened about 54 universities in Nairobi on 18 to 20 August 2026 to start drafting an AI framework for higher education. No framework had been published when we checked.
• Some universities have their own guidelines. Strathmore University's Center for Teaching Excellence says AI "should not be used to complete assignments unless expressly permitted". Where permitted, students must acknowledge it, keep their prompts and responses, and "conduct a critical check for facts and sources".
• Many have not published one. A student guide updated in August 2026 reports that most public universities have not published formal AI policies. Where none exists, general academic integrity rules apply: AI text presented as your own is dishonesty.
• Referencing styles are set by your school. Strathmore's graduate research guide (revised May 2022), for example, asks for APA in IT, business and social sciences, and MLA in humanities such as literature and history. In-text references must "correspond exactly" to the reference list.
• Similarity checks are strict. Kenyatta University's anti-plagiarism policy, as described on its library website, sets an overall similarity limit of 16%, with no more than 1% from any single source. Strathmore's guide accepts a general score below 25% unless the school sets its own level.

Your supervisor's rules come first. Before using AI in your thesis, ask these in writing and keep the reply:

1. May I use AI tools, and for which tasks?
2. Which referencing style and edition should I use?
3. How should I declare AI use: in the methods chapter, a preface, an appendix or a separate form?
4. Do you want to see my prompts?

Our page on AI for thesis writing in Kenya has more on agreeing the rules with your supervisor. For the structure of the thesis itself, see how to write a Master's thesis.

A 10-minute reference check before you submit

Run this on every chapter before it goes to your supervisor.

• [ ] Every reference came from a database, library or publisher page, not from a chatbot's memory.
• [ ] I have opened every source and seen the claim I cite it for.
• [ ] Every DOI opens the right paper at doi.org.
• [ ] References without a DOI are found in a library catalogue or on the publisher's site, with full details.
• [ ] Every in-text citation has a matching reference, and every reference is cited.
• [ ] Author spellings and years match between the text and the list.
• [ ] One style, one edition, used throughout.
• [ ] No retracted papers (Zotero will flag those with a DOI or PubMed ID).
• [ ] My AI use is declared the way my university asks, and my prompts are saved.

Questions people ask

Can ChatGPT give me real references?

Sometimes, but you cannot tell which by looking. Studies found about 18% to 55% of chatbot citations were fabricated, depending on the model. Treat every chatbot reference as unconfirmed until you have opened the source.

Can I ask AI to find references for my topic?

Ask it for search terms instead, then search Google Scholar and your library yourself. Some AI search tools link to real papers, and those are safer, but you must still open every source before you cite it.

Is it cheating to use AI to format my references?

Usually not, if your university and supervisor allow AI use, because formatting does not create the ideas or findings. Check your university's rules and declare the use if they ask. The citations must still match real sources you have read.

What should I do if I have already submitted a fake reference?

Tell your supervisor as soon as you find it, and replace it with a real source you have read, or remove the claim. Owning it early is far better than an examiner finding it. Then check the rest of your list.

Do I need to cite AI if I only used it to fix grammar?

It depends on your university and style. APA's 2025 guidance says you do not cite AI built into software, such as Copilot in Word, but that if you use a separate AI tool to edit your own writing, you disclose it in the author note or method section and cite the tool. Your university's own rules may ask for more, so check with your supervisor.

Sources

All sources last checked 11 November 2026.

• Walters, W. H. and Wilder, E. I. (2023), "Fabrication and errors in the bibliographic citations generated by ChatGPT", Scientific Reports 13, 14045
• Bhattacharyya, M. et al. (2023), "High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content", Cureus 15(5): e39238
• JMIR Publications via EurekAlert, press release on Linardon et al. (2025), JMIR Mental Health 12: e80371 (17 November 2025)
• Linardon, J. et al. (2025), "Influence of Topic Familiarity and Prompt Specificity on Citation Fabrication in Mental Health Research Using Large Language Models: Experimental Study", JMIR Mental Health 12: e80371
• PsyPost, report on Linardon et al. (2025), including the 46% and 17% binge eating disorder figures
• Kalai, A. T., Nachum, O., Vempala, S. S. and Zhang, E. (2025), "Why Language Models Hallucinate", arXiv:2509.04664
• Legal Dive, sanctions in Mata v. Avianca (22 June 2023)
• The Register, "Deloitte refunds Aussie gov after AI fabrications slip into $440K welfare report" (6 October 2025)
• Damien Charlotin, AI Hallucination Cases Database
• University of British Columbia Library, "Citing generative AI" (summary of APA Style's September 2025 guidance)
• Newcastle University Library, Cite Them Right Harvard style sheet (2025)
• Purdue OWL, APA reference list: articles in periodicals
• Purdue OWL, APA in-text citations: author/authors
• Purdue OWL, APA reference list: author/authors
• Zotero home page
• Zotero documentation, adding items to Zotero
• Zotero Storage plans
• Zotero blog, "Retracted item notifications with Retraction Watch integration" (14 June 2019)
• Mendeley Reference Manager
• Mendeley Cite
• Crossref, retrieving metadata
• Crossref, Simple Text Query
• DOI Foundation, key facts on the DOI system
• Google Scholar help
• Africa International University, report on CUE policy dialogue on an AI framework for higher education (August 2026)
• Strathmore University Center for Teaching Excellence, student guidelines for the use of generative AI tools
• Strathmore University, Guide to Research for Postgraduate Theses, Dissertations and Projects (revised 10 May 2022)
• Kenyatta University Library, review of the anti-plagiarism policy
• McTaba Labs, "AI and academic integrity at Kenyan universities: what the policies say" (updated 26 August 2026)

Where FuKazee fits in

FuKazee runs the ScholarShift Accelerator, a four-week live online programme for Masters and PhD students who want to use AI properly without putting their academic integrity at risk. Researchers in our cohorts have come from universities in Kenya, Ghana and Scotland.

We assume nobody in the room knows AI yet, we use no jargon, and we show the free version of every tool without pushing one product. The reference workflow in this guide is part of what we teach, alongside literature review, data analysis and writing. Nobody becomes an expert in four weeks. You leave with a method you can keep using until you graduate.

See how the programme works on the ScholarShift Accelerator page.

To ask about the next cohort, email us at lead@fukazee.com