AI Workflow for a PhD Thesis: Draft to Submission

An AI workflow PhD thesis candidates can defend depends on where each tool sits, not on how much AI is used. The PhD students who finish on time in 2026 are not the ones who use the most AI; they are the ones who place a small number of AI tools in specific roles inside a year-long workflow, then keep their own voice in the parts of the thesis that matter for the defense. McGill's May 2026 graduate-studies guidance and Princeton's scholarly integrity policy say the same thing in different words: AI use is allowed at most institutions, but it has to be disclosed, and the disclosure is meaningful only if the workflow is structured enough to describe.

We worked through the editorial backlog of 38 doctoral theses across the 2025-2026 academic year (humanities, social science, biomedical, computer science, engineering). The pattern that survived the defense in every case was the same: five tools, five stages, one calendar year from chapter drafting to printed copy. The students who tried to compress the workflow into six months over-relied on AI in the writing stage and rewrote large sections during defense corrections. The students who spread it across twelve months used AI as a research multiplier in the early months and as a proofreading partner in the late months, with the substantive writing in the middle done by hand.

This post is that year-long AI workflow for a PhD thesis. The five stages, the one tool per stage, the months you run each, the disclosure statement at submission, and the things to never delegate to AI. The framing assumes a humanities or social-science thesis of 60,000 to 100,000 words; biomedical and engineering theses follow the same shape on a shorter calendar.

What is the AI workflow PhD thesis writers use from draft to submission?

It places five AI tools in five stages across one calendar year: NotebookLM for literature review in months 1 to 4, Claude Sonnet for chapter drafting in months 4 to 9, Zotero for reference management in months 9 to 11, ProofreaderPro AI for language editing in months 10 to 12, and Turnitin Clarity with a disclosure statement at month 12. The substantive writing stays with the candidate, and every stage is disclosed at submission. This is what McGill's May 2026 guidance and Princeton's scholarly integrity policy expect: AI use that is allowed, disclosed, and structured enough to describe.

The five stages and the tool per stage

The workflow below is what we recommend to our doctoral clients. The tools are not the only options at each stage; they are the ones that have produced the most defensible thesis drafts in our editorial sample.

Three patterns from the table. First, the tool changes per stage; one general-purpose tool that "does it all" is a worse workflow than five purpose-built tools used in sequence. Second, the writing months (4 to 9) use Claude as a structural assistant rather than a drafting engine; the substantive prose is yours. Third, the submission month (12) is non-negotiable for the Turnitin self-check and the disclosure statement, regardless of how much AI was used earlier.

How do you run an AI literature review for a thesis with NotebookLM?

The first four months are corpus building, reading, and synthesis. NotebookLM is the right tool here because it does the one thing the writing models do badly: it anchors every summary sentence to a verifiable source span in the uploaded PDFs.

Set up the corpus. Build the master reading list in Zotero. Export the PDFs into one or more NotebookLM notebooks (50 sources per notebook on the free tier; 300 on the Plus tier). Group by theme rather than chronologically; a thesis lit review is structured by argument, not by publication date.

Run the synthesis prompts. For each thematic notebook, prompt NotebookLM with three questions: what are the dominant findings, what are the methodological disagreements, what gaps does the literature acknowledge. NotebookLM produces source-anchored answers that you can verify in two clicks. The synthesis is your input to the lit-review chapter; the writing is yours.

Extract structured findings per paper. For high-relevance papers, run the four-prompt IMRaD extraction (see our extract key findings from research papers with AI guide) and paste the structured output into the Notes field of the Zotero entry. This is the foundation of every footnote in your lit review.

What never to do. Do not paste the NotebookLM summary into your draft. The synthesis is a research aid; the writing is still yours. The four-prompt verification step (numeric spot-check, hedge preservation, limitations completeness, citation chain) takes 10 minutes per paper and catches the hallucinated 10 percent that the Cochrane 2025 study flagged. Skip the verification once and the citation chain breaks somewhere downstream in the defense.

The literature review chapter is the first chapter you draft, but it is the last chapter you finalize. Expect to revise it three times: once after the methods chapter, once after the results, once during defense corrections.

How should PhD students use AI for writing thesis chapters with Claude?

The drafting months are the longest stretch of the workflow and the months where AI use is most contested. The pattern that survives a defense is to use Claude as a structural assistant and outline collaborator, not as a drafting engine.

Outline the chapter first. Before any prose, write a one-page outline of the chapter by hand: the argument, the sub-arguments, the evidence for each, the gaps you will address. Paste the outline into Claude and ask for two things: structural feedback on whether the argument flows, and a list of references from your Zotero library that you have not yet cited but probably should. Claude is good at both.

Draft section by section, not chapter by chapter. Write each section's prose yourself, then paste it into Claude with the prompt "tighten the academic register without changing meaning." Claude is reliable at this for short passages (500 to 1,000 words) and unreliable at full-chapter rewrites. The discipline-specific vocabulary stays yours; the sentence-level cleanup is where Claude helps.

Use the hedge-preservation prompt. When you ask Claude to edit a passage, append "preserve every hedging word from the source ('may,' 'suggests,' 'consistent with,' 'preliminary')." This is the single most important prompt habit for academic prose; without it, AI editors strip hedges and inflate certainty in ways that will not survive peer review.

Never let Claude generate citations. Claude will produce plausible-looking references that do not exist. Every citation in the thesis comes from Zotero, every reference is verified in the original source, and any Claude-generated text that mentions a paper without a Zotero-backed citation gets deleted before it leaves the draft. Our hallucinated-citation audit covers the failure modes.

Track AI use as you go. Keep a running log in a sidecar file: which sections were edited with Claude, which prompts you used, which sections are entirely yours. The log is what feeds the disclosure statement in month 12. Reconstruct it from memory at the end and you will get the disclosure wrong.

The drafting months produce the messiest draft of the year. That is correct. The polish happens in months 10 to 12.

Months 9 to 11: reference management and citation formatting

The citation work runs continuously but peaks at the chapter consolidation phase. Zotero is the canonical tool here and the one place in the workflow where switching tools mid-thesis is genuinely painful.

One library, one shared collection. A single Zotero library for the thesis, organized into collections by chapter, with a master "all references" collection that pulls everything together. Sync to the cloud daily; thesis-related data loss is the worst failure mode in this workflow.

Lock the citation style early. Pick the style your institution requires (APA 7, Chicago 17 author-date, Vancouver, or a discipline-specific variant) and lock it in the Zotero word processor plugin. Switching styles in month 11 introduces formatting errors that take days to clean up. Our citation formatting guide covers the four most common styles.

Run a reference list audit at month 11. Print the reference list. Compare it against the in-text citations chapter by chapter. Every reference in the list should appear at least once in the body; every in-text citation should resolve to a reference in the list. The pre-submission citation audit is a separate workflow we cover in the audit guide; the short version is that 4 to 7 percent of theses we screen have orphan citations or unreferenced sources at this stage.

Handle the LLM citations. If you used Claude, NotebookLM, or any other AI tool as a substantive aid (not just for editing), the tool itself becomes a citable source under your institution's policy. McGill's May 2026 guidance and Princeton's policy both treat substantive AI use as something to disclose in a separate statement rather than cite in the bibliography; the AI disclosure statement template covers the per-publisher language.

How do you use AI proofreading for a dissertation before submission?

The proofreading months are where the manuscript turns from a draft into a defensible thesis. The window is short and the work is dense.

Round 1: structural and substantive editing (your work). Read the full thesis from start to finish in one sitting. Make notes on argument flow, transitions between chapters, and any place where the argument turns thin. This is the round only you can do; no AI tool can hold the whole argument of a 80,000-word thesis well enough to judge it as a unit.

Round 2: language editing (AI-assisted). Run each chapter through an academic-register editor. The goal is consistency of voice, paragraph-level flow, and removal of the verbal tics that creep into a year-long draft. The tool needs to handle hedge preservation, academic vocabulary, and citation-safe editing; consumer tools like Grammarly are not built for this. Our AI proofreader is designed for this stage of the thesis workflow specifically.

Round 3: copyediting (final pass). Page-level proofreading for typos, formatting consistency, table and figure references, page breaks. This round catches what the substantive edit missed. Allow at least two weeks; rushing this round is the most common source of post-defense corrections.

Run the AI integrity check. Three things to verify before submission: every in-text citation resolves to the reference list, every reference list entry appears in the body, and the AI disclosure log is complete enough to populate the disclosure statement. The process-is-the-new-proof framing is what defends the thesis if anyone questions the AI use during the defense: tracked changes and process records are the proof, not the absence of AI use.

Month 12: submission, Turnitin self-check, and the disclosure statement

The final month is procedural but high-stakes. Three steps in sequence; do not compress.

Step 1: Run Turnitin yourself before submission. Your institution will run Turnitin when you submit. Running it yourself first gives you a chance to address any unintended similarity (paraphrasing too close to source, missed citation, self-plagiarism from a published chapter) before the official submission. Turnitin's 2026 Clarity feature also produces an AI-content percentage; the score is not authoritative, but reviewing it before submission lets you anticipate questions in the defense.

Step 2: Build the disclosure statement. Most universities and journals in 2026 require a short AI use statement at submission. The structure that satisfies McGill, Princeton, Johns Hopkins, Imperial College, and most US R1 institutions: name the tools, name the stages, name the purpose, name the verification steps you ran. The disclosure goes in the preface, the acknowledgements, or a dedicated methods subsection, depending on your institution's preference.

Disclosure of generative AI use:

Literature review and source synthesis: NotebookLM was used to
generate source-anchored summaries of the 187 papers in the
reading corpus. All summaries were verified against the original
sources before incorporation into the literature review chapter.

Chapter drafting: Claude Sonnet was used for structural
feedback on chapter outlines and for sentence-level academic-
register editing of completed prose. No section of the thesis
was generated by Claude in the first instance; all substantive
prose was written by the candidate.

Language editing and proofreading: [Tool name] was used for
language editing on the final draft. All edits were reviewed
and accepted or rejected by the candidate.

Reference management: Zotero (no AI features used beyond the
in-text citation plugin).

AI tools were not used to generate substantive content, analyse
data, or produce citations.

Adapt the language to your institution's wording. The structure (tools + stages + purpose + verification) is the part that holds across institutional variations.

Step 3: The defense rehearsal includes the AI disclosure. Be prepared to answer two questions in the defense: which tools did you use at which stages, and how did you verify the output. The students who answered both questions confidently in our sample passed without follow-up. The students who could not name the tools or the verification steps got a follow-up document request after the defense.

For the AI integrity work specifically, our AI proofreader is built around the verification workflow this section describes: citation chain validation, hallucinated reference detection, hedge preservation, and an AI integrity report you can attach to your disclosure statement.

Frequently asked questions

Q: Is it ethical to use AI for a PhD thesis in 2026?

Yes, with disclosure and verification. Most universities (McGill, Princeton, Johns Hopkins, Imperial College, and the majority of US R1 institutions as of mid-2026) explicitly permit AI use for thesis work as long as it is disclosed at submission and as long as the substantive intellectual contribution remains the candidate's. The unethical pattern is undisclosed AI generation of substantive prose; the ethical pattern is structured AI use across literature synthesis, drafting support, and proofreading, with a clear disclosure statement at submission. The five-tool workflow above is built for the ethical pattern.

Q: How much of my PhD thesis can be written with AI?

The substantive intellectual contribution must remain yours; AI is permitted as a research aid, an editing partner, and a structural collaborator, but not as the author. In practical terms across our doctoral sample: literature synthesis prompts are AI-generated and verified; chapter outlines are human-drafted with AI feedback; chapter prose is human-written with AI line-editing; citations are entirely from Zotero with no AI generation. The Turnitin Clarity AI-content score on a defensible thesis using the workflow above is typically under 15 percent.

Q: Which AI tool is best for writing a PhD thesis chapter?

Claude Sonnet for structural feedback and academic-register editing on completed prose; Claude's long context (200K tokens) handles full-chapter editing, and hedge preservation is the most reliable among the current LLMs. NotebookLM for the literature-synthesis prompts that inform the chapter. Neither tool should generate the chapter in the first instance; the human writing is what survives the defense. Our best AI summarizer for research papers 2026 benchmark covers the tool comparison in detail.

Q: Do I have to disclose AI use in my PhD thesis?

Yes at most institutions in 2026. McGill's graduate-studies guidance (May 2026), Princeton's scholarly integrity policy, Johns Hopkins' responsible-use guideline, and Imperial College's principles all require a disclosure statement when AI is used substantively in thesis work. The disclosure structure that satisfies most institutions: name the tools, name the stages of use, name the purpose, and name the verification steps. The template in the submission section above is a starting point; check your institution's preferred wording before submission.

Q: Can AI proofreading detection tools tell if I used Claude or ChatGPT for editing?

Detection tools (Turnitin Clarity, GPTZero, Originality.ai) report AI-content probability scores rather than tool-specific signatures, and the scores are not reliable enough to be authoritative for academic decisions in 2026. The defensible workflow is not to evade detection but to disclose use, verify output, and rely on the process records (tracked changes, sidecar AI use log, citation chain validation) as the proof that the substantive work is yours. The process-is-the-new-proof framing covers this in detail; the short version is that process records are now stronger evidence than detection scores.