Responsible Research Practice and Use of LLMs

Researchers are responsible for the quality, accuracy, and integrity of the work they produce. This includes work created with the help of LLMs, coding assistants, or other automated tools.

Use tools in ways that help you work effectively while continuing to develop the skills you need as an independent researcher.

Develop the skills you need as a researcher

A PhD gives you time to develop skills that require sustained practice: critical reading, synthesis, technical reasoning, argumentation, scientific writing, debugging, research judgment, and the ability to structure complex ideas. These are precisely the skills that make a PhD valuable in this field.

Make sure your use of tools leaves you enough opportunity to develop these skills yourself. When deciding, for instance, whether to use an LLM for a task, consider whether doing the task yourself would help you develop expertise that you still need to acquire.

Responsible use of LLMs and AI tools

Take responsibility for your output

Anything you send to me or to collaborators is treated as your own work, regardless of the tools used to produce it.

Before sharing text, slides, code, analyses, or rebuttals:

  • read and check the material carefully;
  • verify factual claims, references, calculations, proofs, and generated code;
  • make sure you understand every substantive claim and technical choice;
  • check that ideas follow in a clear and natural order;
  • introduce concepts and notation before using them;
  • make sure each paragraph, section, or slide has a clear purpose;
  • test and understand any code you use.

Do not send lightly reviewed generated material and leave collaborators to identify its problems.

Use LLMs in ways that support your development

LLMs can be useful for brainstorming, editing, coding assistance, and routine tasks. Use them in ways that preserve your own judgment and understanding.

You should be able to explain and defend the arguments, equations, experiments, code, figures, slides, rebuttals, and written claims that you produce.

Writing, structuring an argument, deciding what information a reader needs and in what order, synthesizing the literature, debugging, and responding to criticism are part of research training. Make sure you continue to practise these skills yourself.

Be aware of cognitive offloading

LLMs can be extremely powerful, and I use them myself. They can also produce code, arguments, or explanations that look convincing while containing serious errors. Recognizing these errors can require substantial expertise in both the scientific problem and the tools being used.

Heavy reliance on LLMs can therefore become self-reinforcing: the more intellectual work you delegate, the less practice you get developing the expertise needed to evaluate the output, and the harder it becomes to notice when the tool is wrong.

This is one form of cognitive offloading: transferring part of the cognitive work to an external tool. Some offloading is useful and can make research much more efficient, but excessive reliance can reduce engagement, motivation, and opportunities for learning.

During a PhD, efficiency matters, but so does developing the judgment and expertise needed to work independently and use these tools effectively.

Revise existing work carefully

Be particularly careful when using an LLM to modify an existing draft. Large rewrites can remove important details, change correct arguments, introduce inconsistencies, or undo choices that were made deliberately.

Read the complete result and inspect the changes carefully. Preserve correct arguments, notation, references, terminology, and parts of the narrative that have already been developed and discussed.

Avoid workflows where you pass feedback to an LLM and return its output without evaluating the changes yourself. Decide how to address the feedback and make sure the resulting version remains coherent as a whole.

For code, you should be able to understand, test, maintain, and debug what you use. For papers, slides, and rebuttals, you should understand the argument and make the substantive decisions yourself.

Confidentiality and publication policies

Follow the relevant institutional, confidentiality, and publication policies when using AI tools. Do not provide confidential material, unpublished work belonging to others, sensitive data, or restricted code to external services unless this is permitted.

Further reading

Experimental and computational practice

Failed experiments are part of research. Investigate them carefully before giving up on an idea or project.

Before concluding that an approach does not work, check the implementation, data pipeline, baselines, optimization, hyperparameters, evaluation procedure, sanity checks, and plausible failure modes. Bugs can be subtle, and a negative result is meaningful only when there is reasonable confidence that the experiment tested what it was intended to test.

If a project encounters difficulties, investigate them systematically and persist through ordinary technical problems. The advisor may help diagnose difficult issues, but researchers remain responsible for the technical execution of their own projects.

Resources: NeurIPS Paper Checklist and Good Enough Practices in Scientific Computing.

Research integrity

Report results honestly and accurately. Keep appropriate records, give proper attribution, make research reproducible where practical, and be open about errors and limitations.

Fabrication, falsification, plagiarism, deliberate misrepresentation of results, or concealment of material problems with an experiment or analysis are unacceptable.

You remain responsible for the accuracy and integrity of material produced with the help of AI tools, including references, factual claims, proofs, experimental results, and code.

Resource: French Office for Research Integrity (OFIS).