AI for students and PhD candidates: working with verifiable sources

You have a stack of PDFs to read, a thesis to structure and a dataset to analyze, all due soon. This article walks through three specific tasks you can finish today with AI tools, without cheating and without exposing sensitive data. Each task starts from a file provided with the article and gives you a prompt to copy, an expected result and the human check to do afterward.

The rule that runs through the whole article: AI helps you search, extract, check and rephrase. It doesn’t write your intellectual contribution and it doesn’t invent sources. You remain the author, and you disclose your use according to the rules of your course, your university and the journal you’re targeting.

What you need

• An account on a consumer AI app: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) or Le Chat (Mistral). The three tasks below work with a free account, with stricter usage limits.
• The sample files for this article: three abstracts of fictional papers, a course sheet and a small data file. They let you reproduce each task without handling copyrighted documents or real data. The downloadable files are in French; you can copy the English versions shown on this page instead.
• For task 3, a way to run a script: Python or R installed on your computer, or the environment provided by your lab.

Check one thing before you start: your institution’s AI use policy and each instructor’s guidelines. They take precedence over any general advice.

Two rules to know before you paste anything

Disclose your use. The major scientific publishers and the International Committee of Medical Journal Editors (ICMJE, January 2026 update) ask for the same thing: an AI can’t be an author of a paper, and any substantial use must be disclosed at submission.

Elsevier requires a statement placed just before the references that names the tool and how it was used. Basic spelling and grammar correction doesn’t need to be disclosed. On the university side, UCLouvain’s note to students (approved on February 17, 2026) rests on transparency, authenticity and a ban on misleading the grader; it allows language assistance features without mentioning them, but prohibits any uncited verbatim reuse.

In France, there is no single national form: policies are adopted institution by institution (Paris 1 Panthéon-Sorbonne and the Sorbonne Law School, among others). Read yours.

Protect the data. Never paste real data from participants, patients or interviews into a consumer chatbot. On the free and personal plans of ChatGPT and Claude, what you type can be used to train future models until you turn that option off, and turning it off isn’t enough to make a research protocol compliant.

For real sensitive data, go through an Education or Enterprise account negotiated by your university (with a data processing agreement), a self-hosted survey tool such as LimeSurvey, and your ethics committee. In this article, all the files are synthetic.

Working in a private project

A project (Projects in ChatGPT and in Claude) is a private workspace: you put your files and a standing instruction in it, and the assistant draws on that content for all your conversations inside the project.

You keep the same framework from one session to the next without rewriting it. Projects exist on every subscription tier, including free (Claude limits free accounts to five projects; check the current limit in your account).

This article relies on these private projects and on source notebooks. In the exercises below, the project answers from the files you upload and the messages you send; it doesn’t act on its own in other software unless you’ve turned on an authorized connector. This difference from a real automated system comes up again further down.

Build a source-grounded reading matrix (task 1)

This is the simplest task: copy and paste is all it takes. The goal is to turn a small set of papers into a comparison table where each row points back to its source.

App: Google’s NotebookLM (see the official page at notebooklm.google.com), or a ChatGPT or Claude project. NotebookLM is designed to answer from the sources you upload and to cite them, which helps you trace each claim back to where it came from.

That doesn’t establish accuracy, so check every citation in the source. A free notebook currently accepts up to 50 sources (PDF, DOCX, TXT, CSV and pasted text, among others); check the current limit in your account.

Sample file: resumes_articles_fictifs.txt (three made-up abstracts), provided with this article. In real work, replace them with your own PDFs. Content to copy:

[Source 1] Dumont, C. & Vaskey, L. (2021). Spaced review and four-week retention
in first-year students. Randomized controlled trial, 120 first-year undergraduates
split between spaced review (6 short sessions) and massed review (1 long session).
Measure: free recall at 4 weeks. The "spaced" group's recall is higher by
18 percentage points. Stated limitations: a single university, no follow-up beyond
4 weeks, 11% attrition.

[Source 2] Ferreira, A., Blum, S. & Ndiaye, K. (2023). Active recall versus rereading: a
meta-analysis of 34 studies. Overall effect size g = 0.52 in favor of active recall.
High heterogeneity. Stated limitations: publication bias not formally assessed,
definitions of "active recall" vary across studies.

[Source 3] Novak, P. & Petit, M. (2022). Flashcard app versus paper notes
in nursing students. Quasi-experimental study, 88 students, no
randomization. No significant difference in exam grade, but higher
self-reported confidence in the app group. Stated limitations: no
randomization, self-reported confidence measure, duration limited to one semester.

Steps:

1. Open NotebookLM and create a notebook, or create a project in ChatGPT or Claude.
2. Add the sources: upload the sample file, or paste the three abstracts as separate sources.
3. Paste the instruction below.
4. Get the table and check each cell against the cited source.

Prompt to copy:

You are a literature review assistant. Use ONLY the sources
uploaded to this notebook. Don't add any outside knowledge and don't invent
any reference. Build a table with one row per source and these columns:
Source (number and authors, year), Question or objective, Study design,
Sample, Main result, Stated limitations. After the table, state in
one sentence per source where the information comes from (file name). If a piece of information
is missing from a source, write "not stated" instead of guessing it.

Expected result (from the fictional abstracts):

Human check: open each source and confirm the figure and the limitation in its row. A matrix that doesn’t point back to the source is worthless for a literature review.

Common error and fix: the assistant fills an empty cell with a plausible value that isn’t in the text. If you see a figure that doesn’t appear in the source, delete it and run it again with the instruction “write not stated if the information isn’t in the source.” Whatever the tool, check each cell against the cited source: an assistant that shows a citation can still point to the wrong passage.

Create a comprehension and review exercise (task 2)

The goal is to turn a course sheet into review questions, then quiz yourself, with the sheet as the only reference.

App: NotebookLM (designed to answer from the uploaded sources and cite them) or a project. Then check each answer against the sheet.

Sample file: fiche_cours_imrad.md (a short, self-contained sheet on the IMRaD structure and experimental design), provided with this article. Content to copy:

The IMRaD structure. An experimental scientific paper often follows the IMRaD structure:
Introduction, Methods, Results, Discussion. The Introduction states the research question
and the hypothesis. The Methods section describes how the study was carried out, precisely
enough to be reproduced. The Results present the observed data, without interpreting them.
The Discussion interprets the results, compares them with the literature and states the limitations.

Experimental design. The independent variable is the one the researcher manipulates. The
dependent variable is the one the researcher measures. A control group is the point of
comparison: it doesn't receive the intervention being tested. Randomization assigns
participants to groups at random, which reduces selection bias. Internal
validity indicates the extent to which the observed effect is really due to the independent variable;
external validity, the extent to which the results generalize to other contexts.

Points to watch. A correlation between two variables doesn't establish cause and
effect. Sample size affects the precision of estimates. A study's limitations
must be stated in the Discussion.

Steps:

1. Upload the sheet as the only source in a notebook, or in a project.
2. Paste the instruction below to generate the questions.
3. Switch to quiz mode: answer from memory first, then compare.
4. Check that each answer appears word for word in the sheet.

Prompt to copy:

Using ONLY the uploaded sheet, write 8 review questions and their answers.
Number them. Each answer must rest on a specific sentence in the sheet;
don't add anything that isn't in it. Then switch to quiz mode: ask me
the questions one at a time, wait for my answer, then say whether it's correct,
quoting the sentence from the sheet that supports it. If one of my questions goes beyond the
sheet, answer "outside the material provided."

Expected result (excerpt):

• Q1. What does the acronym IMRaD stand for? Answer: Introduction, Methods, Results, Discussion.
• Q2. Which section presents the data without interpreting it? Answer: the Results.
• Q3. What is the difference between the independent variable and the dependent variable? Answer: the independent variable is manipulated by the researcher, and the dependent variable is measured.
• Q4. What is a control group for? Answer: it is the point of comparison and doesn’t receive the intervention being tested.

Then, in quiz mode: “Question 4 of 8: what is a control group for?”

Human check: for each answer, find the original sentence in the sheet. If the assistant supports an answer with a sentence that doesn’t exist, throw the question out. This exercise tests your understanding; it doesn’t replace reading the course.

Common error and fix: the assistant adds an outside detail to an answer (a statistic, an example that isn’t in the sheet). Restate the “only the sheet” constraint and ask for the supporting sentence. On a very short sheet, a notebook may cite the whole document instead of a specific passage, which is normal.

Files to download (in French): Python script (zip), quiz data, fictional abstracts and course sheet (zip).

Check a data analysis script on a fictional dataset (task 3)

The goal is to ask the AI for an analysis script, run it yourself, then reconcile the numbers with a calculation you can check. The AI writes useful code, but you’re the one accountable for the results.

App: ChatGPT or Claude to generate the code; Python or R to run it.

Sample file: quiz_revision.csv, twelve fictional rows (ID, group, review hours, score out of 20). The hours are weekly, not cumulative. One value in the hours column was entered as 250, a deliberate error: a week only has 168 hours, so 250 is physically impossible. The file content, to copy as is:

participant_id,groupe,heures_revision,score_sur_20
P01,A,5,12
P02,A,8,14
P03,A,6,13
P04,A,7,15
P05,A,4,11
P06,A,6,10
P07,B,10,16
P08,B,12,18
P09,B,9,15
P10,B,11,17
P11,B,250,19
P12,B,8,16

Steps:

1. Describe your columns and paste the instruction below. Don’t upload real participant data; here, the file is synthetic.
2. Copy the script it produces and run it on the file.
3. Read the output, then recalculate one value by hand.
4. Deal with the outlier together with your supervisor before making any decision.

Prompt to copy:

Here is a CSV file named quiz_revision.csv with the columns: participant_id,
groupe (A or B), heures_revision (weekly review hours), score_sur_20 (score out of 20).
Write a standalone Python script that: 1) loads the file; 2) flags missing
values and physically impossible values (review hours
above 168, which is more than the hours in a week) without deleting any
row; 3) gives, per group, the number of observations, the mean and the standard deviation
of the score; 4) runs an independent-samples Student's t-test comparing
the scores of groups A and B, stating that it assumes equal variances. If a
score is missing, flag it and stop the inferential calculation instead of filtering out the
row. Don't invent any data, don't delete any row, don't replace any missing
value. Add a comment explaining each step.

To reproduce the exact output, put the reference script analyse_quiz.py (provided with this article) and quiz_revision.csv in the same folder, then run python3 analyse_quiz.py. The last line (p-value) uses SciPy. Without SciPy, the script still prints the means, standard deviations and t, then tells you where to read the p-value; don’t install any package without permission.

Expected result, on a machine with SciPy installed (the reference script prints its output in French):

Lignes: 12
Valeurs manquantes -> heures: 0, score: 0
heures_revision min=4 max=250 (une valeur > 168 h depasse les heures d'une semaine, donc impossible: a verifier a la main)
Groupe A: n=6 moyenne=12.5000 ecart-type(echantillon)=1.8708
Groupe B: n=6 moyenne=16.8333 ecart-type(echantillon)=1.4720
Test t (variances egales): t=-4.4590, ddl=10, difference des moyennes=-4.3333
scipy ttest_ind: t=-4.4590, p=0.001218

In order, these lines give the row count, the missing values, the hours range with a warning that a value above 168 is impossible and must be checked by hand, the mean and sample standard deviation for each group, and the equal-variance t-test with its degrees of freedom and the difference in means. Without SciPy, the first six lines are identical and the last one is replaced by a note telling you to read the p-value in a t table with 10 degrees of freedom.

Human check: recalculate group A’s mean by hand. The six scores are 12, 14, 13, 15, 11 and 10, so 75 divided by 6 equals 12.5. The script shows 12.5, which is consistent.

The value of 250 hours shows up in the maximum, beyond the 168 hours in a week: it’s up to you to decide, with your supervisor, whether it’s a typo to correct or a row to exclude, with a written justification.

This t-test assumes equal variances between the two groups, an assumption you need to check. Since the dataset is fictional, the p-value (0.0012) establishes no real effect, and it doesn’t say that method B “causes” better results either.

Common error and fix: some models delete the row with 250 on their own, silently filter out a missing value, or invent a p-value without running the code. Reject all three.

Require the script to flag the anomaly without touching it and to stop the calculation if a score is missing instead of dropping the row, and only keep a p-value if it comes from your own run. Falsifying data or cleaning it blindly is the most serious misconduct in research.

Reusable project instructions

Set the standing instruction once per project. It applies to all your conversations in that project and saves you from repeating the framework. Example for a research paper:

Context: I'm preparing a university paper. Answer in English, in a plain
academic register. Rely only on the files in this project and on the
sources I provide; never cite a reference you haven't seen in these
files. When you extract a piece of data, state which file it comes from. Don't write
analysis or discussion paragraphs for me: suggest outlines,
questions and rephrasings, and flag what still needs checking. If a
piece of information is missing, write "not stated" instead of guessing it.

A reminder: this instruction lives in your private project and only frames your conversations. You don’t have to publish it or share your files.

Choosing the right tool for the task

Tell five things apart. The model is the engine (for example OpenAI’s GPT families, Anthropic’s Claude, Google’s Gemini, Mistral). The consumer app is the interface (ChatGPT, Claude, Gemini, Le Chat).

The subscription sets your usage limits. The API is for building a model into your own program. The connector (the MCP protocol, or platforms such as Zapier, Make and n8n) links the AI to a piece of software.

To choose quickly: a short, closed task (correcting pasted text, summarizing a paragraph, generating title ideas) runs very well on a fast, inexpensive model. A reasoning task on long documents (comparing fifteen papers, structuring a thesis, reviewing an analysis script) calls for a reasoning model with a large context window.

To turn a chapter into slides, Microsoft 365 Copilot works directly in Word and PowerPoint, where many universities already give you a license. Exact model names change almost every month: check the version available in your account and see our reference, Choosing your AI model and organizing work across several agents, instead of relying on a fixed list.

For language review, checkers such as Antidote or DeepL Write handle spelling, grammar and style; UCLouvain’s note lists this kind of language assistance among the uses that don’t need to be disclosed. That doesn’t exempt you from checking your course’s guidelines.

A coordinator and subagents, done by hand or as a real system

You’ll read everywhere that AI now works in teams: a “coordinator” breaks the task down and hands it to specialized “subagents.” Two warnings will save you from a costly confusion.

Giving three roles in a single prompt doesn’t create three independent agents: it’s one AI playing three characters. Copying a result from one chat window to another is manual delegation: you do every handoff yourself.

Here’s an exercise you can do by hand, in an hour, to understand the principle:

1. Open four chat windows. You are the coordinator.
2. “Explorer” window: list the relevant sources for your question, from the fictional abstracts provided, not from a list the AI makes up.
3. “Reader” window: paste one of these abstracts and ask it to extract the method, sample and results.
4. “Evaluator” window: paste the same fictional abstracts again, then ask it to assess the stated limitations of each study. A fresh chat can’t assess limitations without the text in front of it.
5. “Writer” window: paste the validated extracts, with the source file name for each, and ask for a sourced summary that you then correct.

At every step, you transfer the text and you check it. A real automated system chains these steps without you copying and pasting: it needs access to databases and files through connectors, permissions, technical implementation and a human approval step before anything is sent.

That goes beyond simply signing up for a subscription: these connections have to be configured and authorized, often with some development. The diagram below sums up the difference.

Coordinator and subagents: the manual exercise compared with an automated system

1. Manual exercise
doable today, with no tooling

You = coordinator

you copy and paste between the windows ↓

Explorerfind the sources

Readerextract method, sample, results

Evaluatorassess how solid it is

Writerwrite the sourced summary

Each role is a separate chat window. You’re the one who moves the text from one to the other. This is manual delegation, done by hand.

2. Connected orchestration
requires tooling

Coordinator (AI)

automatic delegation ↓

Research subagent

Extraction subagent

Summary subagent

↓

Connectors, MCP or API to bibliographic databases, files, calendar

↓

Human approval before anything is sent

Access, permissions and technical implementation, beyond a simple subscription.

Tasks never to hand over to AI

• Writing your contribution: analysis, discussion, methodology, conclusion. These parts carry your scientific responsibility.
• Inventing references, quotations or data. Check every source in its original document.
• Fabricating data or result images. An AI-generated figure that invents a result is prohibited; a chart you produce from your own data and your own code remains legitimate. Check the policy of the journal you’re targeting: Elsevier, for example, prohibits using AI to create or alter results in images, while allowing basic adjustments such as brightness.
• Using it to hide your use of it. Transparency is the rule; disclose according to your course, your university and the journal.

Checklist for your next session

• I’ve read my institution’s AI policy and the instructor’s guidelines.
• In a consumer tool, I only use synthetic (fictional) files.
• Every extracted piece of data is checked in its original source.
• I’ve run every script myself and recalculated one value by hand.
• I’ve noted where and how I used AI, ready to disclose it.
• I cite no reference I haven’t read.

Support for teams that need it

You can do these three tasks on your own, with a consumer subscription. If your lab, department or practice wants to go further, with a configured assistant, agreed data rules, connectors authorized case by case and an approval step before anything is sent, that’s the work MACCUS does.

The assistant works through Telegram (text, voice notes, documents) in French, English and Dutch; the files and memory are hosted in France, with part of the processing done by providers based in the United States.

It isn’t an all-EU solution or a guarantee of compliance, and it replaces neither scientific judgment nor the disclosure of your use. To discuss it based on how your own organization works: discover the MACCUS assistant.

To start with a single task and understand the logic from end to end, read the guide Using AI in your profession. If you freelance alongside your studies, the guide AI for consultants and freelancers applies the same methods to client work.