How to Check Whether the References in Your Thesis Are Real
To check a thesis reference properly, verify more than whether the paper exists. Confirm the authors, title, year, journal and DOI or other identifier; check that you are citing the correct published version; open the source and verify that it actually supports the claim in your text. AI-generated references can be fabricated or contain convincing bibliographic errors, so every citation remains the author's responsibility.
Your reference list contains:
Okafor, J., Mensah, P., & Smith, R. (2022). Digital health interventions and contraceptive autonomy in sub-Saharan Africa. Journal of Global Reproductive Health, 18(4), 211-224.
Looks academic.
Has authors.
Has a year.
Has page numbers.
Possibly has absolutely no paper behind it.
Plausibility is not verification.
What does it mean to verify a reference?
A useful reference audit asks five different questions.
1. Does the source exist?
Can you find the actual paper, book, report or dataset in a credible scholarly source?
2. Is the bibliographic information correct?
Do the:
- authors;
- title;
- year;
- journal;
- volume;
- issue;
- pages;
- DOI;
match the real source?
3. Are you citing the right version?
A preprint may later have a peer-reviewed publication.
A conference abstract may have become a full article.
A report may have been superseded.
4. Does the source actually support the claim?
This is the check researchers skip surprisingly often.
5. Have you cited it accurately in the manuscript?
Is the source attached to the right sentence?
Does your wording overstate what the study found?
Those are separate checks.
A real paper can still be a bad citation for your sentence.
Existence is only the first gate.
Why are AI-generated references risky?
Large language models can produce bibliographic information that looks convincingly academic while being wrong.
Walters and Wilder tested 636 citations produced in literature reviews generated by GPT-3.5 and GPT-4 in 2023.
In that particular study:
- 55% of GPT-3.5 citations were fabricated;
- 18% of GPT-4 citations were fabricated.
Even among citations linked to real sources, bibliographic errors remained.
Do not turn those percentages into permanent failure rates for every modern AI system.
Models change.
Prompts change.
Tasks change.
The durable lesson is:
A fluent citation is not evidence that a source exists.
Human verification remains necessary.
the five-minute reference audit
Suppose your thesis says:
“Adolescents are less likely to use SRH services when confidentiality cannot be assured (Namusoke et al., 2023).”
Step 1: search for the source
Use one or more of:
- PubMed;
- Crossref;
- Google Scholar;
- Web of Science/Scopus if available;
- journal website;
- institutional repository;
- library catalogue;
- subject-specific database.
Search by:
- exact title;
- author + year;
- DOI;
- distinctive phrase.
If the citation cannot be located, do not immediately assume the database is incomplete.
Investigate.
Step 2: compare metadata
Check:
- author order;
- publication year;
- article title;
- journal title;
- volume/issue;
- page or article number;
- DOI.
AI errors can splice together plausible pieces from different real papers.
A real journal + real author + fake article title still equals a broken reference.
Step 3: open the source
Do not stop at:
“Google Scholar found it.”
Read the abstract at minimum.
Preferably inspect the full text relevant to your claim.
Now ask:
Does this study actually support the sentence I attached to it?
Step 4: inspect the strength of the claim
Suppose the source is a cross-sectional survey showing an association between perceived confidentiality and service use.
Your thesis says:
“Lack of confidentiality causes adolescents to avoid services.”
The source may exist.
The citation may still overclaim.
A better statement might be:
“Perceived confidentiality has been associated with adolescent service use in cross-sectional studies.”
The citation audit is also an interpretation audit.
Step 5: check the version
Suppose you cited:
Smith et al. 2024. arXiv preprint.
Search whether that preprint now has a final journal version.
The final publication may contain:
- changed title;
- different pagination;
- revised findings;
- new DOI;
- corrected analyses.
Where appropriate, update the reference.
Three different citation failures
1. Fabricated source
The work does not exist.
2. Corrupted citation
The source exists, but metadata are wrong.
Example:
- correct author;
- wrong year;
- journal from another article;
- invented DOI.
3. Citation laundering
The source exists and is cited correctly, but it does not support the claim.
This may be the most dangerous type.
A fake DOI eventually fails when someone searches it.
A real paper attached to a claim it never made can survive much longer.
Bench Check: reference managers do not verify truth
“It is in Zotero, so it must be real.”
No.
Reference managers are excellent for:
- storing metadata;
- inserting citations;
- changing styles;
- organizing libraries.
They do not guarantee that:
- the imported metadata are correct;
- the source supports your sentence;
- you read the source;
- the reference was not imported from a faulty record.
A tidy bibliography can still contain bad scholarship.
What if I have 250 references?
Manual checking sounds unpleasant because it is.
You do not have to audit every reference with the same intensity.
Prioritize:
Highest-risk references
- AI-suggested sources;
- citations you cannot remember reading;
- sources copied from another paper's reference list;
- old preprints;
- obscure journals;
- references with missing DOI/metadata;
- claims central to your argument;
- statistics used in the abstract or introduction;
- controversial or surprising claims.
Then perform a systematic full-list check before final submission.
A practical reference-audit table
| Reference | Exists? | Metadata correct? | Correct version? | Supports claim? | Action |
|---|---|---|---|---|---|
| Source A | Yes | Yes | Yes | Yes | Keep |
| Source B | Yes | No | Yes | Yes | Fix metadata |
| Source C | Yes | Yes | Preprint only | Yes | Check published version |
| Source D | No | , | , | , | Remove/replace |
| Source E | Yes | Yes | Yes | No | Rewrite claim/find source |
This makes the audit traceable.
What do I actually write if I discover a bad citation?
Do not simply replace the citation with the first real paper on the topic.
Ask what the sentence needs.
Original sentence:
“Peer support reduces depression relapse by 40%.”
If the cited source is fake, you have two jobs:
- find whether credible evidence supports the claim;
- revise the wording to match the evidence you actually find.
Sometimes the correct fix is:
delete the sentence.
A reference replacement should not preserve an unsupported claim at all costs.
Automated citation checking: where it can help
For a long thesis, automated tools may reduce part of the checking burden.
Check My Thesis currently states that its citation-verification service checks uploaded work against multiple academic databases, flags potentially fabricated or outdated references, maps preprints to published versions and offers citation-related corrections.
Methods Bench has not independently validated those performance claims yet.
Use automated verification as a screening layer.
Not as a declaration that every citation is now scientifically correct.
Affiliate disclosure: Methods Bench may earn a commission if you purchase Check My Thesis through our link, at no additional cost to you. Our recommendation does not replace manual source verification.
What automated tools cannot decide for you
Even if a tool confirms that a paper exists, you still need human judgement about:
- whether the source is appropriate;
- whether the study design supports your statement;
- whether you have overstated causality;
- whether a more authoritative source exists;
- whether the evidence is outdated;
- whether the citation belongs in the argument at all.
Reference verification is partly bibliographic.
Scholarship is interpretive.
What do I actually write in my final thesis checklist?
Use:
Reference audit completed
- Every high-risk citation was located in a scholarly source.
- Bibliographic metadata were checked against the source.
- Preprints were checked for later published versions.
- Claims central to the thesis were verified against the cited source.
- References suggested by AI tools were independently verified.
- Reference-list entries and in-text citations were reconciled.
That is a much stronger pre-submission check than:
“Run Zotero refresh.”
Pick five references from your thesis:
- one generated or suggested by AI;
- one central to your argument;
- one old reference;
- one preprint;
- one you do not remember reading.
For each one, open the actual source.
Then ask:
If a reviewer clicked this citation, would they find what my sentence claims?
Frequently asked questions
- How do I know if a citation is fake?
Search the exact title, authors and DOI in credible scholarly databases or the publisher's website. Failure to locate it after systematic checking is a warning sign.
- Can Google Scholar verify references?
It is useful for discovery, but verify key metadata against the publisher, DOI registration or authoritative database where possible.
- Does a DOI prove a citation is real?
A valid DOI linked to the correct source is strong evidence of existence, but a fabricated reference can contain an invented or mismatched DOI.
- Should I verify references generated by ChatGPT or another AI?
Yes. Treat AI-generated references as unverified until you locate and inspect the real source yourself.
- Can software tell me whether a source supports my claim?
Some tools claim contextual citation checking, but human reading remains necessary, especially for important or nuanced claims.
Download the Methods Bench Reference Audit Sheet.
Then, if you want an automated first-pass screen for a long document, use Check My Thesis as an optional affiliate tool before completing the manual audit.
References and further reading
- 1.Walters WH, Wilder EI. Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports. 2023;13:14045. doi:10.1038/s41598-023-41032-5.
- 2.Crossref. Scholarly metadata and DOI lookup.
- 3.PubMed/NLM. Biomedical literature database.
- 4.ICMJE Recommendations on human responsibility for AI-assisted scholarly content.
- 5.Check My Thesis website for current vendor-described product features. Product claims should be treated as vendor claims until independently tested by Methods Bench.
Take it further
Better research, once a week.
Practical methods guidance, research tools and funding opportunities from Methods Bench.
Get practical research notes and new opportunities
Short, useful emails for researchers in Uganda and East Africa. No spam.