Most discussions about AI ask whether students will stop thinking. That is the wrong question. The better question is why so much thinking was never captured in the first place.
Eight months into a project, a supervisor asks a fair question: why were those three patients excluded? You remember deciding. You remember it was a careful decision, made for a reason. Then you spend an evening paging through two notebooks, an email thread, and a folder called final_v3, and the reason is in none of them.
Every PhD already runs on external memory
Papers, notebooks, Zotero libraries, highlighted PDFs, supervisor meetings: nobody holds a doctoral project in their head, and nobody ever did. What none of these tools was built to hold is the reasoning behind a decision: why this cut-off, why that method, why the paper that contradicted the plan did not change it. That is what slips away.
For decades this weakness stayed comfortable, because everyone reconstructed on demand. A committee asked, you paused, you rebuilt the argument from fragments and confidence, and mostly it worked.
AI makes the weakness visible
A language model can generate a plausible answer faster than you can reconstruct your own past thinking. When the fluent answer arrives in seconds and the real reason lives in a notebook you cannot find, the fluent answer wins by default. The model has outrun your memory of your own reasoning.
So the most exposed researchers right now are the ones whose reasoning was never written down anywhere it could be retrieved.
Organise around what you know
Research is usually organised around documents: a folder for papers, a folder for drafts, notes filed by topic. A thesis is built from claims, and every claim has a status you can name.
What do you currently know? Why do you believe it? What would change your mind?
Organise the work around those three questions and the reasoning stops depending on memory. A decision made in March is still explainable in November, because the note that holds the claim also holds why you believed it and what its source was. An AI suggestion can then be weighed against a recorded reason instead of a reconstructed one.
The future researcher will not be the one who writes without AI. It will be the one who can always explain where every important claim came from.
Heidi