Is ChatGPT feedback enough for your thesis?

As a deadline approaches, the obvious move is to paste the whole thesis into a chatbot and ask "what is missing?". This guide explains where that works, where it can mislead you, and when another kind of tool may be worth it. It is about the general behaviour of general-purpose chatbots, not any one product; these tools are updated constantly, so the details of the version you use may differ.

What general chatbots do well

To be fair, these tools genuinely help at many points in writing a thesis.

• Explaining concepts: they can explain the assumptions of a statistical test, the core ideas of a theory or the difference between two methods in plain language, with as many examples as you need.
• Brainstorming: they are well suited to generating alternative research questions, possible limitations, or counter-explanations to address in your discussion.
• Quick feedback on a passage: in seconds they can tell you whether a short paragraph is clear or a transition feels abrupt.
• Flexible dialogue: if you do not like an answer, you can rephrase, push back with "why?", or narrow the topic. Most dedicated tools do not offer that flexibility.

In short, a chatbot is like a study partner who answers every question patiently. The trouble starts when you expect it to read your entire thesis the way a committee would.

Where it struggles with a full thesis

Attention may be uneven across a long document

Even when a chatbot accepts a long text, it may not evaluate every chapter with the same care. Feedback often concentrates on prominent parts such as the introduction and conclusion, while an inconsistency in the middle of the methods chapter, or a table in chapter three that contradicts the text, can slip through. A general, positive summary does not mean the text was actually read from start to finish.

References are judged from memory

A chatbot may comment on a publication in your reference list based on what it remembers from training data, without looking it up in any database. When it suggests sources, it can produce publications that do not exist: in one experiment, 19.9% of citations generated by GPT-4o pointed to publications that do not exist (Linardon et al., 2025). That figure may change as models change, but the principle holds: what the tool says about a source does not show that the source exists or supports your claim. The reference check guide explains how to verify yours.

Quotes are not verified against your text

When the tool says "in chapter two you wrote…" and shows a sentence, that sentence may not appear in your text word for word; it may be a close paraphrase. If the criticism rests on that reconstructed sentence, you can waste time fixing a problem that is not there.

Feedback is not anchored to locations

The answer arrives as a conversation. A point like "in places the interpretation goes beyond the findings" may be right, but it is hard to act on without knowing which page or paragraph. Which problems are serious and which are minor is usually not stated either.

A tendency to rewrite your text

Chatbots can be inclined to hand back "improved" sentences even when you did not ask for them. That is a real academic integrity issue: universities and graduate schools have different rules on AI use, and some restrict AI-generated text in a thesis or require it to be declared. Check your own institution’s current guidelines. The AI for thesis writing guide covers which uses are generally considered safe.

Confidentiality

When you upload an unpublished thesis to a service, how the text is stored and whether it is used to train models depends on that service’s settings and your account type. Check the tool’s data settings before uploading, and consider not sharing chapters that contain participant information or data covered by ethics approval at all.

How to get better results from a chatbot

These limits do not mean you should avoid chatbots altogether. How you ask largely determines what you get.

1. Give it one chapter at a time. Start with your research question and a one-paragraph summary of the thesis, then send each chapter in a separate message, so attention stays on a single chapter.
2. Ask for questions, not rewrites. For example: "Read this methods chapter as a committee member would. Do not rewrite anything; just list the questions the committee might ask and the reason for each."
3. Ask for locations. Say "for each point, quote the relevant sentence from my text exactly", then search for that sentence in your own file. If you cannot find it, do not trust the point.
4. Ask for priorities. Asking it to sort problems into serious, moderate and minor makes it easier to decide where to start.
5. Verify every reference yourself. Open and read every publication the tool suggests or comments on, through a database or its DOI.
6. Ask for the opposing view. "What other explanations could there be for this finding?" lets you see the committee’s questions in advance. More in the thesis defense questions guide.

When a dedicated review tool makes sense

If you want to understand a concept, clarify a paragraph or generate ideas, a general chatbot is often enough. If you want to check the whole thesis before submission or your defense, compare your references with real records, and know exactly where each finding sits in your text, a tool built for that job may save effort. Whichever you choose, the best check is still a reading by someone who knows the field but has never seen the text; for the steps, see the thesis check guide.

tezkeys is built for that second job: it reads your thesis chapter by chapter and then evaluates the whole, marks findings on your own text with location and severity, compares your references with Crossref records, and searches for every quote it shows in your text verbatim, dropping any it cannot find. It does not write your text for you. The report is AI-generated, can be wrong, and does not replace your advisor. The reference pre-check is free; see what a report looks like in the sample report.