Four Ways to Leverage AI in a PhD Dissertation

Meta description: A practical guide to 4 Ways to Use AI on a PhD Thesis: citations, code, fact-checking, and defence prep, with human oversight and clear limits.

Tags: AI, PhD thesis, academic writing, research tools, Towards Data Science

Featured image: Abstract friendly illustration of a thesis, reference links, code blocks, and question cards, with no real people, no logos, and no text.

The post titled 4 Ways to Use AI on a PhD Thesis appeared first on Towards Data Science and set out four practical uses: finding citations, consolidating code, fact-checking, and preparing for the defence. This guide is for doctoral candidates and early-career researchers who want to use AI as a structured assistant rather than a ghostwriter. AI here means tools that can generate text, explain code, summarise drafts, or suggest questions. The key point is that AI can speed up repetitive academic work, but the researcher still decides what is true, what is citable, what is reproducible, and what can be defended.

Find citations as a search task, not a generation task

The first use is finding citations. In a thesis, citation work is often scattered across reading notes, PDFs, reference managers, and older drafts. An AI assistant can help by turning a vague memory into a search plan: suggesting keywords, listing possible authors, generating queries for library databases, or comparing a draft reference list with a bibliography. It can also help format entries and spot missing fields.

The limit is that language models can produce plausible but false references. Every citation should be checked in the original source or a trusted bibliographic database. If a tool suggests a source, treat it as a lead, not evidence. Doctoral candidates should also check their university rules about disclosing AI use, because citation work sits close to source verification and academic integrity.

Consolidate code without hiding analysis decisions

The second use is consolidating code. PhD projects often leave behind notebooks, scripts, data cleaning steps, plots, and small utilities. AI can help summarise a long script, convert a notebook into a reusable function, draft a README, identify repeated logic, or generate tests for simple outputs. It can also help translate comments into documentation.

This can improve reproducibility, but it must be done carefully. Confidential participant data, proprietary code, and unpublished results should not be uploaded without checking institutional policy. The author should keep version control and verify that any AI-generated changes do not alter assumptions, random seeds, data filters, or analysis decisions. A cleaner script is useful only if the scientific choices remain transparent.

Fact-check claims before polishing prose

The third use is fact-checking. A thesis contains many kinds of statements: definitions, method claims, background facts, numerical results, and interpretations. AI can help by asking for each claim, its evidence, and its confidence level. It can flag overstatement, inconsistent terminology, missing caveats, or sections that sound certain without data support. It can also suggest counterarguments for a literature review.

This is not a replacement for systematic verification. For empirical results, the numbers must come from the dataset, code, and analysis logs. For background claims, the sources must be primary or authoritative. The useful workflow is: write the claim, ask AI to test it, then verify the answer manually. The researcher decides whether the claim survives.

Prepare for the defence by testing arguments

The fourth use is preparing for the defence. Examiners usually test contribution, method, limitations, and alternative explanations. AI can help by generating likely questions from a chapter summary, converting methods into plain-language explanations, or creating mock defence prompts. It can also help organise answers into a short structure: claim, evidence, limitation, next step.

The risk is that AI may invent questions that fit a generic pattern rather than the actual thesis. Use it to practise, not to predict. A supervisor, advisor, or peer should check whether the questions reflect the real argument and whether the answers stay within the boundaries of the evidence.

Use the four methods if you need structure, skip them if you need certainty

These four methods are most useful when the problem is mechanical but time-consuming: locating missing sources, tidying code, testing whether a claim is overstated, or rehearsing defence questions. They are less useful when the core issue is scientific judgement, ethical review, or whether the thesis argument is actually defensible. In those cases, AI can suggest options, but the decision still belongs to the candidate and their supervisor.

For readers, the practical takeaway is to give AI a narrow task, keep verification human, and document where it was used. That approach preserves academic control while reducing the time spent on repetitive cleanup.