Best AI Tools for a PhD Dissertation: A Practical Guide

👁 119 reads

For your PhD dissertation, don’t choose one AI tool for everything. Build a toolkit: use Semantic Scholar or Elicit to find research, Zotero to manage references, NotebookLM (now Gemini Notebook) to work with your own PDFs, and a general-purpose language model such as ChatGPT, Gemini or Claude to help with code and editing your own writing. Each tool has a specific job, and you need to know how to check its results.

This article is part of our AI in education series. Information and prices are current as of October 12, 2026. Feature descriptions and prices come from the providers’ websites, linked in the sources. To see how often models invent research references, read our test in Which AI should you use for an undergraduate dissertation?

AI tools for each stage of your PhD

This table suggests how to divide the work; it is not a ranking of model quality. Prices are shown in the currency used by the provider.

If your institution uses Google Workspace for Education, check your university account. Google says that files and questions in Gemini Notebook on such an account are not reviewed by humans or used to train models. The rules differ for personal accounts, which matters if you’re working with unpublished data.

How to build your research workflow step by step

1. Describe your research problem in your own words. Write down what your dissertation aims to explain, what it deliberately leaves out and what material it uses. Ask a model to identify unclear points in this description, not to write an outline.
2. Map the literature starting with a paper you know. Begin with a key publication in your field, then review its references and the papers that cite it. Natural-language searches in Semantic Scholar or Elicit can point you towards useful leads, but they don’t show what they missed. Keep a record of searches that returned no useful results, too.
3. Add every paper you find to Zotero. Your reference list shouldn’t live in your chat history. Install Zotero add-ons carefully: Zotero’s documentation warns that plugins have full access to the application and your computer, so only use ones from trusted developers.
4. Build an evidence table instead of summarising the whole field. Use one row per publication, with columns for the authors’ question, material, method, finding relevant to your work, limitations, page number and your notes. “Not reported in the text” is a valid result; a blank cell should prompt you to check the source. For a systematic review, follow the PRISMA 2020 checklist. It describes reporting items; it does not certify any software.
5. Analyse data in a reproducible environment. First define the unit of observation: does each row represent a person or a measurement? Then ask for code. Test the script on synthetic data with the same structure. Judge the result by running the code, not by the chat’s assurance that its calculations are correct.
6. Write from your own findings. A model can help put an idea into words if you can defend that idea. Changing “association” to “cause” is a scientific change, not a stylistic one. Work in excerpts and keep a copy of your draft from before revisions. For help organising chapters, see our guide to writing an outline with AI.
7. Practise your defence. Ask the model to list objections a critical examiner might raise. If your answer to a question about your choice of method is “the AI suggested it”, you need to revisit your rationale.

Fabricated references and plagiarism checks

Fabricated citations are a major risk in academic work. In a study by Walters and Wilder published in Scientific Reports in 2023, 55% of citations from GPT-3.5 and 18% from GPT-4 were fabricated. In an analysis published in JMIR Mental Health in November 2025, GPT-4o fabricated 19.9% of 176 citations; for under-researched topics, the rate rose to 28-29%. Check every reference from a chat in a database instead of asking the same model whether it is real. Read more about the issue in our article on AI hallucinations.

How to test a tool before paying

Before you pay for a plan, try it on material you know well.

1. Choose a publication or data excerpt with a challenge typical of your work, and prepare your own reference answer.
2. Set your criteria before the test: Is the reasoning supported by the text? Has the model preserved the degree of certainty in each claim? Does the code produce a result that matches a worked example?
3. Run the task and mark anything omitted or added.
4. Point out an error and see how the tool responds.
5. Check the export: after your subscription ends, will you still have clear documentation?
6. Record your decision and repeat the test if the model or plan changes.

Show your supervisor the answers you rejected, too. It’s a useful starting point for agreeing how to document your AI use. Rules differ between universities and doctoral programmes, so follow the policies that apply to your institution.

A prompt you can copy

The prompt below is an English translation of the original Polish prompt template, not a new test or a prompt tested in English.

Work only with the following paragraph from my dissertation in [field].
1. List the claims and the evidence supporting them.
2. Mark places where a conclusion goes beyond the material described here [brief description of the data].
3. Suggest edits without adding facts, numbers or sources.
4. Show any change in meaning (for example, changing association to causation) separately for me to decide.
5. Do not supply publication titles or authors’ names from memory.
Paragraph: [paste your own text]

Frequently asked questions

Which AI is best for a PhD dissertation?

There’s no single winner. Research search engines such as Semantic Scholar and Elicit are better suited to finding literature; Zotero is for managing references; and a general-purpose model can help with code and editing. Choose a tool for a task whose results you can verify.

Can you complete a PhD using AI for free?

Yes. You can start with a free toolkit: Semantic Scholar, Elicit Basic, Zotero and Gemini Notebook on the Standard plan. You only pay for higher limits, such as Elicit Plus at 11 USD/month with annual billing.

Can AI find a research gap?

It can suggest a possible gap. But a topic’s absence from a model’s answer doesn’t prove that nobody has researched it. Confirm any proposed gap by searching the databases yourself.

Do you need to tell your supervisor if you use AI?

Agree on this at the start of your research. Many universities require you to describe the tools you used and what you used them for, but the specific rules depend on your institution and doctoral programme.

What should you do if AI invents a paper?

Remove it and check the other references in the same answer. Don’t replace it with a similar-sounding title unless you’ve read the paper and confirmed that it supports your claim.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources

• Semantic Scholar - FAQ (Semantic Scholar, accessed October 12, 2026)
• Elicit - pricing (Elicit, accessed October 12, 2026)
• Zotero - Plugins (Zotero, updated May 5, 2026)
• Zotero - Storage (Zotero, accessed October 12, 2026)
• Gemini Notebook - plans and limits (Google, accessed October 12, 2026)
• NotebookLM changes its name to Gemini Notebook (Google, July 16, 2026)
• PRISMA 2020 checklist (PRISMA, accessed October 12, 2026)
• Walters and Wilder - fabricated ChatGPT citations (Scientific Reports, 2023)
• Fabricated GPT-4o citations in reviews (JMIR Mental Health, November 2025)

• Read next: Which AI should you use for references?
• Which AI should you use for a master’s dissertation?
• Using NotebookLM for studying

Kinetyra