as a physicist and computational scientist. By Collin Wittenstein.
At the end of my mathematics bachelor thesis, I included a whole chapter documenting my usage of AI tools during my work on the thesis. As of November 2024, a slightly adjusted and shortened version of this documentation now serves as an official example for declaring AI use in theses at the Institute of Mathematics at the Johannes Gutenberg University of Mainz.
Since I don't want to write such detailed chapters for every academic work but still want to maintain transparency regarding my AI use, I created this Code of Conduct as a reference document that I can link to in my academic projects.
- Large Language Models: Claude, ChatGPT
- Code Assistance: GitHub Copilot
I use LLMs as supplementary explanation tools to help clarify and understand difficult concepts. I treat them as knowledgeable colleagues who would never admit that they are wrong or don't know something. All explanations are verified and processed - I trace through arguments myself and form my own understanding rather than accepting information at face value. Sometimes it is easier and faster to verify a solution than to come up with it yourself.
Example: When trying to understand a paper about the temporal double slit experiment, at one position I was confused by an asymmetry which I thought shouldn't be there. Claude helped explain this asymmetry by referring to an explanation in the same paper, which came about 5 pages later, which I probably wouldn't have found this fast.
As someone with dyslexia and English as a second language, I use AI to:
- Check grammar and spelling mistakes
- Improve clarity in my written expression
- Suggest alternative phrasings for ideas I've already developed
- Provide feedback on draft content
I retain editorial control, incorporating only suggestions that preserve my intended meaning and voice.
I keep GitHub Copilot active during coding sessions primarily as an enhanced autocomplete tool (mostly for Julia, but also for LaTeX)
- AI helps with syntax and routine coding patterns
- I thoroughly verify and understand all AI-suggested code
- I do not do "vibe coding" (accepting code without understanding) in my academic work
Example: While the typical use of Copilot for coding tasks should be clear, I also use it when writing LaTeX—particularly for equations. For instance, after typing out one expression, Copilot can often suggest the LaTeX code for a related or rearranged version. This isn’t mathematically hard, but it saves time—especially with long or annotated variable names. For more involved transformations, however, Copilot often fails and I fall back on manual input.
- Tasks where verification would be impossible or too time-consuming
- Complex implementations where accuracy is critical but verification is challenging
- Solutions to problems where I want to develop the skills myself
Example: For my Bachelor's thesis in mathematics, I needed to translate a quite long algebraic expression from Mathematica into Julia code. As it wouldn't have been possible to check this translation, I did not use LLMs for this problem.
This code of conduct was last updated in May 2025 with AI assistance for writing and editing, following the same principles outlined within. It represents my personal commitment to ethical AI use in academic contexts. I welcome feedback and discussion regarding both my principles and how else to efficiently and supplementally use AI in academia.
Signed: Collin Wittenstein