A second-year PhD student uses a generative AI assistant for almost everything: summarising papers, suggesting paragraph structures, writing analysis code, drafting emails. His productivity rises noticeably. At his progress review, a committee member asks: “What do you see as the biggest limitation of the three foundational studies you rely on?” He falters. He has “read” all three through summaries but has never wrestled with any of their methods sections himself. Nobody accuses him of cheating. The problem is subtler: he has skipped exactly the work that turns a student into a researcher.
This article is not a call to avoid AI. It helps you distinguish tasks worth delegating to free up time from tasks worth doing yourself because the struggle is the learning.
The paradox of a very convenient tool
Learning research has long shown that “desirable difficulties” such as recalling, explaining and finding your own errors make knowledge last, even though they feel slow and tiring at the time. Generative AI assistants remove precisely those difficulties. The immediate output improves, but the underlying capability may not grow. For experts this is fine: they delegate to focus on harder problems. For learners it is a trap.
A delegate-or-do matrix
| Skills you need to own long term | Supporting skills | |
|---|---|---|
| Already mastered | Delegate partly, but do it yourself now and then to stay sharp | Delegate freely; just check the output |
| Still learning | Do it yourself; use AI as a tutor to question you, not a substitute | Delegate, but understand the output before using it |
A concrete example: formatting a reference list in APA style is a supporting skill, so delegate it. Writing a literature review is a core skill of the profession; if you are still learning it, write it yourself, even if the first draft is poor.
Five core skills most at risk
1. Reading a paper deeply
A summary tells you what a paper concludes; only deep reading tells you how far to trust it. The questions examiners, reviewers and colleagues ask almost always sit at that second level.
2. Writing as thinking
Writing an argument paragraph is how you discover what you do not yet understand. Accepting a fluent paragraph from a machine skips that discovery. Many supervisors notice that a clumsy self-written draft shows them how a student thinks; a polished machine draft does not.
3. Statistical reasoning
AI writes analysis code quickly and usually with correct syntax. But which model to choose, which assumptions to check and how to interpret coefficients are matters of judgement. If you cannot explain every line of the code to your supervisor, you do not yet own the analysis.
4. Finding and evaluating literature
Knowing a field means knowing who matters, which debates are live and which journals to trust. That knowledge builds from hundreds of times following citations yourself, not from lists handed to you.
5. Talking about your research without notes
At a viva, a job interview or an unexpected conference question, no assistant stands beside you. This skill only comes from explaining your work regularly.
Signs of over-reliance
- You cannot start a paragraph without asking the AI first.
- You cannot recall the methods of the papers you cite most.
- You can run an analysis but cannot explain why you chose that model.
- When your supervisor asks “what do you mean here?”, you have to reread the sentence to find out.
- You feel anxious about working without tools, in an exam or interview for example.
The feeling of “I get it” after reading a clear summary is deceptive. Learners tend to overrate their understanding right after a fluent explanation and only find the gaps when they have to explain it themselves. So the best test is always to say it or write it again without looking at the screen.
If two or three of these sound familiar, it is time to adjust, not by banning AI but by reclaiming the practice you have been skipping.
Set up AI-free zones for yourself
- First drafts: write the first draft of every important section without AI. Use AI at the revision stage, as a critical reader.
- One slow read a week: pick an important paper, read it in full and write five lines on its strengths and weaknesses before looking at any summary.
- Explain the code: for any analysis where AI helped write the code, annotate every block in your own words.
- Practise speaking: once a month, give a five-minute progress update to labmates without slides.
For supervisors
Outright bans rarely work and are hard to enforce. A more practical approach is to design checkpoints where students must show their own capability: oral discussions of papers they have just read, step-by-step explanations of analyses, and a look at draft history rather than only the final version. And tell students explicitly which tasks you encourage them to use AI for and which you want them to do themselves, with the reasons. Students follow guidance far better when they understand why.
When delegating is safe
A simple test: if the tool disappeared tomorrow, could you still do this task, just more slowly? If yes, you are using AI as a tool. If no, you are using it as a crutch, and it is worth spending time strengthening your own legs.
Something to do this week: list the five things you most often ask AI to do, mark which fall in the “core skill, still learning” cell of the matrix, and pick one to do yourself for the next two weeks. You may be surprised by how much the slowness teaches you.
Câu hỏi thường gặp
Can heavy AI use weaken research skills?
It can, if you hand over tasks you are still learning, such as deep reading, argument writing or choosing an analysis. Delegating mastered or supporting tasks is usually harmless.
Which research tasks should I do myself rather than with AI?
Deep reading of key papers, first drafts of arguments, choosing and interpreting statistical analyses, finding and evaluating literature, and practising talking about your research.
How do I know if I rely on AI too much?
Warning signs include being unable to start writing without AI, being unable to explain your own analysis, not remembering the methods of papers you cite often, and anxiety about working without tools.
How should supervisors manage students’ AI use?
Rather than banning it, design checkpoints such as oral discussions, step-by-step explanations of analyses and review of draft history, and be explicit about which tasks AI is encouraged for and why.
Is there a quick test for whether delegating to AI is fine?
Ask whether you could still do the task, only more slowly, if the tool vanished tomorrow. If yes, AI is a tool; if not, that is a skill you need to practise.
