Bolds the start of each word so you can scan the text faster.
Theme
Language
App
Writing a thesis
~ min read
📋30-second summary
A thesis isn’t a long essay: it’s months of construction work, with a supervisor following you and a committee that will press you on every page. If AI writes it, you fall apart at the defense.
Same mechanics as Drafting professional documents: outline first, filling second. Specific to the thesis: the outline is born with the supervisor, and every chapter is a separate chat.
AI as a second reader, not as an author: you paste the paragraph, it flags gaps, objections, weak transitions. No “write me the next paragraph”.
Use AI to simulate the supervisor before the real submission. It reduces the rounds of correction, it doesn’t replace them.
Many universities ask you to declare LLM use. Write it honestly in the preface and check your university’s policy.
A thesis isn’t a longer version of an essay. It’s months of
construction work, with a method (chapters that stand up on their own
and connect to each other), and an ongoing conversation with a
supervisor who follows it at every stage. It’s also the first piece of
academic work where the Writing on behalf of someone else: the text is signed by X but written by Y, usually for a fee. Common for autobiographies of public figures; in academia, it's considered fraud. pitch lands the hardest:
a hundred pages are a lot, AI writes them fast, it looks like a
reasonable shortcut.
It isn’t. If AI writes it, you’re the first one who loses. At the
defense your supervisor notices (the style, the passages you can’t
defend, the citations that don’t ring a bell), and the committee will
ask you things you can’t answer. A thesis isn’t something you
delegate: you build it with AI as a second reader, not as an
author. Everything that follows starts here.
The underlying mechanics are the same as in Drafting professional
documents: you don’t ask
AI to write the document from a blank page, you start from an outline
and fill it in section by section. On a thesis this pattern matters
even more, because a blank page stretched over a hundred pages
multiplies the damage: if AI writes two thousand words on top of the
wrong outline, you’ve thrown away two days of work.
Two things change compared with the professional document.
The outline is born with the supervisor, not with AI alone. AI can
propose a generic table of contents (“introduction, state of the art,
method, results, discussion”), and it’s useful as a starting point.
But it’s the supervisor who validates that the structure holds up the
academic argument of your field: that the The opening section of a thesis or paper that reconstructs what existing literature has already said on the topic, and where your contribution fits in. Not a list of authors, a reasoned discussion of the field. isn’t a
list of authors with no thread, that the method fits the research
question, that the conclusions answer the initial hypotheses. Bring
the outline to your supervisor at the first meeting, not after the
third chapter is finished. The validated outline is the constraint
everything else starts from.
Every chapter is a separate dialogue. A thesis is too long to fit
in a single chat: AI loses precision as the context stretches, and
each chapter has its own bibliography, its own subject, its own
problems. Open a dedicated chat for chapter 1, one for chapter 2, one
for chapter 3. When you come back to chapter 2 after two weeks, reopen
that chat: AI has in front of it what you’ve already discussed on
that chapter, and doesn’t get confused with the rest of the thesis. In
the chat’s initial instructions paste the overall outline of the
thesis and where the chapter sits in it, so AI has the broader context
without carrying it all in the chat. It’s 200-300 words, it doesn’t
bloat the prompt.
The central pattern of this lesson. You write a paragraph. You paste
it into AI with a specific question, usually something like:
“This paragraph defends the thesis I want to argue, which is X.
Where is it weak? What objections would a strict examiner raise?”
AI doesn’t add content: it points out the gaps. It tells you where
your argument skips a step, where a transition between paragraphs
doesn’t hold, what the three most obvious objections a critical reader
would raise are. Then you go back to the text, rewrite it yourself,
and re-check.
This way of using AI feels like a fellow doctoral student reading a
draft for you at the bar: they don’t rewrite it, they tell you “I’m
not following here”, “this passage is weak”, “you didn’t close the
loop”. Except you get the colleague twice a year, and AI whenever you
want.
What you can ask without problems:
“Make this sentence clearer without changing the meaning.”
“Does this transition between paragraph A and paragraph B work, or
is there a logical jump?”
“Does this argument hold? What are the three biggest gaps you
see?”
“Give me three objections a strict critic could raise against this
paragraph.”
“I’m trying to say X. Does it come through? If not, where do I lose
the reader?”
What NOT to ask, ever:
“Write the next paragraph.”
“Expand this paragraph to 500 words.”
“Write the chapter’s conclusion.”
“Rephrase this paragraph so it sounds like I wrote it.”
The rule that separates the two lists is simple. The first list leaves
the words to you: AI reacts to what you’ve written, then you go back
and rewrite. The second one puts the words in its mouth, and you
become an editor who approves. “Make it clearer” touches vocabulary
and flow, leaves the structure yours. “Rephrase it as if I’d written
it” rewrites the structure and asks you to sign it: that’s the same
thing as “write it for me”. The examiner at the defense can feel it.
Not because there’s a magic detector (AI-generated-text detectors are
unreliable), but because you can’t answer the questions: the passages
you didn’t build don’t ring familiar when you have to defend them out
loud.
Every supervisor has unwritten expectations: citation style (Chicago,
APA, footnotes), the angle of the work (historical-comparative,
analytical, experimental), strictness on the hypotheses (“strong” if
they want a precise thesis to defend, “soft” if a well-done survey is
enough). These expectations aren’t in any thesis-writing handbook:
they sit in your supervisor’s head, and if you don’t ask you’re left
guessing.
Before you start, ask directly:
“What kind of thesis are you expecting? More concise or more
extended? What angle: historical, analytical, experimental? What do
you expect me to push on the most?”
This thirty-minute conversation saves you three months of revisions in
directions your supervisor doesn’t care about.
Once you have a draft chapter, before you take it to the
supervisor: use AI as a supervisor simulator. Typical prompt:
“Pretend to be my supervisor, a professor of [subject] with a
[angle] angle, who pushes on [what the supervisor values]. I’m
handing you chapter 2 of my thesis. What would you ask me to
revise?”
AI doesn’t replace the real submission, and its simulation won’t be
perfect. But it catches 60-70% of the most obvious corrections (unclear
passages, missing citations, wobbly arguments), and you arrive at the
supervisor with a chapter that’s already been combed through. The
rounds of correction drop from four to two. What’s left are the real
corrections, the ones the supervisor gives you because they know your
thesis and your field: the value of those meetings goes up, not down.
Master’s student in art history. Writing her thesis on Mannerism in
Florence, chapter 3, paragraph on the role of Medici patronage. She’s
written a first version, wants to know if it holds up.
👤
You
This is a paragraph from chapter 3 of my thesis. The thesis I want to argue is that Medici patronage isn't the only cause of the birth of Mannerism in Florence: it's an important factor but overrated by traditional scholarship.
Paragraph: "In the 1520s, Medici patronage undoubtedly gave impetus to the new figurative experiments we now call Mannerism, funding artists like Pontormo and Rosso Fiorentino. However, attributing the birth of Mannerism to Medici patronage alone means ignoring the autonomous evolution of pictorial language, which in Florence was already in transformation under the pressure of the encounter with Rome and with Michelangelo's manner."
Does this paragraph defend my thesis? Where is it weak? Three objections from a strict examiner.This is a paragraph from chapter 3 of my thesis. The thesis I want to argue is that Medici patronage isn't the only cause of the birth of Mannerism in Florence: it's an important factor but overrated by traditional scholarship.
Paragraph: "In the 1520s, Medici patronage undoubtedly gave impetus to the new figurative experiments we now call Mannerism, funding artists like Pontormo and Rosso Fiorentino. However, attributing the birth of Mannerism to Medici patronage alone means ignoring the autonomous evolution of pictorial language, which in Florence was already in transformation under the pressure of the encounter with Rome and with Michelangelo's manner."
Does this paragraph defend my thesis? Where is it weak? Three objections from a strict examiner.
🤖
AI
The paragraph sets up the thesis but only half-defends it. It argues that Medici patronage isn't the only cause, but doesn't clarify by how much it's overrated, nor against which concrete alternatives. A strict examiner would raise at least three objections.
Objection 1: the competing patronage is missing. You cite "the encounter with Rome" as an autonomous factor, but you don't name the parallel Roman patronage, which is the main argument against your thesis. If Mannerism is born in Rome too, under Leo X and Clement VII (who are Medici, but act as popes and not as lords of Florence), an examiner asks you: are you saying that Medici-Florentine patronage is overrated, or that Medici patronage in general is? The distinction changes the thesis significantly.
Objection 2: "undoubtedly" is too big a concession. You open the paragraph saying that Medici patronage "undoubtedly gave impetus" to the Mannerist experiments. Then halfway through the paragraph you say that traditional scholarship overrates it. The examiner hears the contradiction: if the impetus is undoubted, what are you basing your critique of scholarship on? You need to quantify what kind of impetus you mean (specific commissions, money, access to ancient models) and show that the artists could have produced that language with a different patron too. Without this, "overrated" stays asserted but not defended.
Objection 3: the "autonomous evolution of pictorial language" is too vague. Key phrase of the paragraph, but it's an assertion, not an argument. Which autonomous evolution? Encounter with Michelangelo, yes, but also with whom? Pontormo looks at Dürer and northern prints, Rosso Fiorentino dialogues with Parmigianino: these are dynamics autonomous from the patron. If you name them, the paragraph holds. If "autonomous evolution" stays generic, the examiner suspects you're covering up a gap.
Reinforcement suggestion, not a rewrite: add an explicit comparison with Clement VII's Roman patronage (one or two names, one or two dated commissions) and name at least two influences other than patronage that weigh on Pontormo's or Rosso's language. From there the paragraph rewrites itself.The paragraph sets up the thesis but only half-defends it. It argues that Medici patronage isn't the only cause, but doesn't clarify by how much it's overrated, nor against which concrete alternatives. A strict examiner would raise at least three objections.
Objection 1: the competing patronage is missing. You cite "the encounter with Rome" as an autonomous factor, but you don't name the parallel Roman patronage, which is the main argument against your thesis. If Mannerism is born in Rome too, under Leo X and Clement VII (who are Medici, but act as popes and not as lords of Florence), an examiner asks you: are you saying that Medici-Florentine patronage is overrated, or that Medici patronage in general is? The distinction changes the thesis significantly.
Objection 2: "undoubtedly" is too big a concession. You open the paragraph saying that Medici patronage "undoubtedly gave impetus" to the Mannerist experiments. Then halfway through the paragraph you say that traditional scholarship overrates it. The examiner hears the contradiction: if the impetus is undoubted, what are you basing your critique of scholarship on? You need to quantify what kind of impetus you mean (specific commissions, money, access to ancient models) and show that the artists could have produced that language with a different patron too. Without this, "overrated" stays asserted but not defended.
Objection 3: the "autonomous evolution of pictorial language" is too vague. Key phrase of the paragraph, but it's an assertion, not an argument. Which autonomous evolution? Encounter with Michelangelo, yes, but also with whom? Pontormo looks at Dürer and northern prints, Rosso Fiorentino dialogues with Parmigianino: these are dynamics autonomous from the patron. If you name them, the paragraph holds. If "autonomous evolution" stays generic, the examiner suspects you're covering up a gap.
Reinforcement suggestion, not a rewrite: add an explicit comparison with Clement VII's Roman patronage (one or two names, one or two dated commissions) and name at least two influences other than patronage that weigh on Pontormo's or Rosso's language. From there the paragraph rewrites itself.
Three things to notice. AI didn’t write a new paragraph: it showed
where yours is weak, with three precise objections that have names and
references. The objections aren’t “do better”: they’re specific things
the student can check against the sources (Clement VII’s patronage,
Pontormo’s influences). And the final suggestion is reinforcement,
not replacement: it tells you what to add to close the gap, leaving
the words to you.
Don’t have AI write the paragraph and then “rephrase” it. It’s
the most common shortcut and the most dangerous: the logical structure
stays AI’s, and at the defense the passages don’t ring familiar. The
examiner hears the difference between an argument you built yourself
and one you dressed up in your own words afterwards.
Don’t use AI to generate the bibliography. Invented citations
are the fastest way to sink a thesis: plausible names, plausible
journals, plausible years, and nothing real. Sources have to be looked
up in academic databases and verified one by one. On how to handle
bibliography and citations safely with AI, see Sources and
citations, the next
lesson.
Don’t stack eight chapters without showing one to the supervisor.
AI doesn’t know what your supervisor really wants until you ask.
Showing chapter 1 after the first week, and having one corrected every
two or three weeks, is the rhythm that keeps you aligned. Arriving with the
whole thesis in hand a month before the defense is the recipe for
rewriting it twice.