Which tool to choose
Here it's not so much the tool that matters as the role you give it.
- For structuring and revising: a capable general-purpose assistant, to which you assign the role of advisor or critical reader.
- For working on your own sources without it inventing anything: NotebookLM, which answers only on the loaded documents.
- For checking language and style after writing: any assistant, asking for precise corrections and not rewrites.
How to do it
The rule that holds it all together: the AI touches everything except the sentences of the final text.
- Make the map before writing. The operational syntax:
I'm writing a paper on [thesis]. Here's my outline: [paste]. Point out to me: where the reasoning has a logical jump, which objection an examiner would raise, which point is weak or poorly supported by the sources. Don't rewrite anything.
- Write the first paragraph of each section yourself, by hand. It's the moment when you really think: if you skip it, the paper loses your head.
- Use the AI as a mirror, not as a pen. The operational syntax:
Read this paragraph that I wrote myself. Tell me whether the argument is clear and whether a step is missing to reach the conclusion. Don't rewrite it: point out to me what to improve, I'll correct it myself.
- Have it interrogate you on your thesis. Ask the AI to play the examiner and attack your conclusions: the cracks it finds now are the ones you don't want to discover at the discussion.
- Final form check. Only at the end, ask for a review of typos, repetitions and convoluted sentences, without distorting the text: "flag the errors and the unclear sentences, but leave my words."
A concrete example
Davide is writing his bachelor's thesis on the circular economy. He has the outline but fears the gaps. He submits it to the AI in the role of advisor: the AI notes that between the chapter on costs and the one on benefits the step about who pays for the transition is missing. Davide adds a section, which at the discussion will be the most appreciated. He writes each chapter on his own. At the end he has the typos checked. The real advisor tells him you can feel his voice: that's precisely what using it as support, not as a substitute, allowed him to keep.
When it does NOT work (and how to fix it)
If the AI starts rewriting your paragraphs even when you don't ask it to
It tends to "improve" by replacing your voice with its own, flat one. Remedy: add to every request "don't rewrite, only point out." If it does it anyway, ignore the rewritten text and use only the suggestions, rephrasing in your own words.
If you notice the text no longer sounds like yours
A sign that you've delegated too much. Remedy: reread aloud. The sentences you'd never say are the AI's: rewrite them the way you'd speak. The oral discussion unmasks in a few seconds a paper not written by the person presenting it.
If you don't know where support ends and substitution begins
The practical line: the AI can work on what you've already thought (criticizing, organizing, correcting), not produce the thinking in your place (writing arguments, drawing conclusions). If you wouldn't be able to re-explain a paragraph aloud, you've delegated it too much.
A tip from someone who actually uses it
Save the conversations with the AI together with the paper. If your university or your advisor asks for transparency about AI use, you have the proof of how you used it: as a consultant, not as an author. Many university regulations now distinguish between lawful use and plagiarism, and being able to show the process keeps you safe.
Frequently asked questions
How much AI use is allowed in a paper?
It depends on your course's regulations, and they must be read before starting. The general rule that holds everywhere: the AI as a support tool (research, revision, structure) is almost always allowed if declared; having it write the content and presenting it as your own is plagiarism. When in doubt, ask the advisor.
Will the advisor notice if I used AI to write?
Often yes, and not because of anti-AI software, which makes mistakes. They notice at the discussion: someone who didn't write a text struggles to defend it, to explain the choices, to answer the objections. You can only defend a paper if you thought it through yourself.
If the AI writes better than me, why shouldn't I use it for the text?
Because the paper doesn't measure how beautiful the text is, but how well you can reason and defend a thesis. A text polished by the AI leaves you speechless at the question "why did you write this?". Writing badly but writing your own is recoverable; not being able to answer at the discussion isn't.