Why most AI thesis writers invent sources — and what changes that
Every general-purpose chatbot shares one failure mode when you ask it for a thesis: it writes the prose and the references in the same breath. A citation, to a language model, is just another string to predict. It has seen tens of thousands of references shaped like Author (Year), "Title," Journal, DOI, so it can produce something with that exact shape on demand — with no mechanism anywhere in the process that checks whether the paper exists.
This is why fabricated references are often more convincing than real ones. They are assembled from the statistical average of real citations in your field: a plausible author name, a title stitched from familiar keywords, a DOI prefix belonging to an actual publisher. Everything looks right because every part is an average of things that are right.
You cannot fix this by asking the model to be careful, and you cannot reliably fix it by asking a second model to check the first one's work. The order of operations has to change.
Retrieval first, writing second
CiteLyra inverts the sequence. Before any prose is drafted, the system searches academic indexes — Crossref, OpenAlex, and Semantic Scholar — for work relevant to your topic. Records that come back are filtered: a citation is discarded unless it carries a title, named authors, and a publication year, and unless it resolves to a DOI or a working URL. Author names are validated (single letters and generic placeholders are rejected), publication years are range-checked, and DOIs registered to preprint servers are flagged as preprints rather than passed off as peer-reviewed articles.
What survives becomes a compiled bibliography. Only then does the drafting begin — and the writing agents receive that bibliography as a fixed list of what they are permitted to cite, annotated with how strongly each source may be used. Sources judged tangential are marked background only: usable for context or motivation, but not as primary evidence for a central claim, and not in the abstract or conclusion as proof.
The practical consequence is structural rather than probabilistic. The model is not writing references and hoping they check out. It never writes references at all — it selects from a list that existed before it started, assembled from records that a database returned.
What this looks like in practice
Consider a thesis sentence that needs support. A generate-then-hope tool produces something like this (illustrative — this reference is fabricated):
Transformer pruning reduces inference latency by up to 40% with negligible accuracy loss (Almeida & Vestergaard, 2021).
Almeida, R., & Vestergaard, T. (2021). Structured pruning for efficient transformer inference. Journal of Machine Learning Systems, 14(3), 221–248. https://doi.org/10.1145/3419887.3419902
Read it critically. The authors are plausible. The title is exactly what such a paper would be called. The journal sounds real. The DOI has a legitimate ACM prefix. Every individual component is credible, and the whole thing is invented — the journal does not exist under that name, and the DOI belongs to a different paper entirely. Your committee will find this in about ninety seconds, and you will spend the meeting explaining it rather than defending your argument.
The retrieval-first path cannot produce that sentence, because the claim is written against a record that was returned by a search before drafting started. If no indexed source supports the claim, there is no citation available to attach to it — and the gap is visible to you in the bibliography rather than papered over with a convincing-looking string.
This is also why we publish complete example drafts with their bibliographies intact rather than excerpts. The references are the part worth checking, so they are the part we leave checkable.
What a thesis-scale draft actually contains
"AI thesis writer" covers a wide range of document sizes, and vagueness here is how tools oversell. The concrete targets:
| Document | Target length |
|---|---|
| Research paper | 3,000–5,000 words |
| Bachelor's thesis | 10,000–15,000 words |
| Master's thesis | 25,000–30,000 words |
| PhD dissertation | 50,000–80,000 words |
These are structured multi-section documents, not one long undifferentiated block. A master's draft arrives with an introduction that frames the research question, a literature chapter organized thematically rather than paper-by-paper, a methods framing section, discussion, and conclusion — with headings that map onto what most universities expect, and inline citations in APA, MLA, IEEE, or Chicago throughout.
Exports go to PDF, Word (.docx), Markdown, LaTeX for Overleaf, or a ZIP bundle containing all of them. Your supervisor comments in Word; your department requires LaTeX; you want the Markdown for your own notes. All three come from the same run.
What it will not do
A thesis is an argument you are responsible for. Being specific about the boundary is more useful than reassurance:
It does not do your research. The system finds published literature on your topic. It does not run your experiments, clean your data, or interpret your results. If your thesis has an empirical core, that core is yours and the draft is scaffolding around it.
It does not know whether a source supports your claim. Retrieval guarantees the paper exists and the metadata is accurate. It does not guarantee that the paper's findings actually back the sentence they are attached to. Reading the sources you keep is not optional, and it is the step people are most tempted to skip.
It does not produce submittable text. A first draft from any tool — human or otherwise — is a starting position. The prose needs your voice, your framing, and your judgment about what matters. Committees are increasingly good at recognizing unedited machine output, and the sections that read as generic are the sections you have not yet made your own.
It does not resolve your institution's AI policy for you. Those policies vary enormously, from outright prohibition to required disclosure to open acceptance for drafting support. Find yours before you start, not after.
A workflow that survives scrutiny
The students who get real value out of this treat it as a literature and structure accelerator rather than a text generator:
- Bring a real research question. The narrower and more specific your topic and constraints, the more relevant the retrieved literature. "Machine learning in healthcare" returns a scattered bibliography; "transformer-based triage prediction in emergency departments, 2020 onward" returns a usable one.
- Read the bibliography before the prose. It arrives with DOIs. Open the ten sources that matter most to your argument and confirm they say what the draft claims. This is the highest-leverage hour you will spend.
- Cut aggressively. A draft that covers thirty sources when your argument needs twelve is worse than one that covers twelve well. Delete what does not serve the thesis.
- Rewrite section by section in your own voice. Use the structure as a skeleton and the retrieved literature as your reading list. The argument has to become yours or the defense will expose that it is not.
- Disclose where required. If your institution asks for a statement on AI use, write one. Describing a tool as a literature search and drafting assistant is accurate and is far safer than silence.
Used this way, the tool removes the two genuinely mechanical parts of thesis writing — finding the relevant literature and getting past the blank page — and leaves the scholarship where it belongs.
If your work is dissertation-scale, the AI dissertation writer page covers long-chapter drafting specifically. If you are mainly trying to validate references you already have, the citation checker does that on its own.