The German Research Association (Deutsche Forschungsgemeinschaft, DFG) recently issued guidelines for using artificial intelligence, including tools like ChatGPT, to write research papers and grant applications. The DFG supports using AI for these purposes, excluding reviews but emphasizes the need for transparency. Therefore, “Scientists should disclose whether, for what purpose, and to what extent they have used generative models when making their results publicly accessible in the interest of scientific integrity” (translated from German via ChatGPT).
Our group uses ChatGPT Premium and occasionally other tools to enhance the writing of our research papers, grant applications, blog posts, emails, and other written communications. In the interest of transparency, I will outline how we incorporate AI tools into our writing process across these different formats.
Translations
We occasionally use ChatGPT for translation tasks. For example, I used ChatGPT to translate a sentence from the DFG guidelines for this blog post.
If we do this, then we explicitly mention the use of AI in the manuscript.
Finding or improving titles for manuscripts
We frequently use ChatGPT to help generate or refine titles for our manuscripts. Sometimes, we attach the manuscript and ask ChatGPT for suggestions.
“Give me ten potential titles for the attached manuscript”
Sometimes, we formulate what we want in more detail. For instance
We have written a research paper about the energy usage of recommender systems (the manuscript is attached). Give me ten suggestions for a title that is catchy yet academic, and makes clear that recommender systems (which are mostly digital / online) do have a real impact on the real world by consuming energy and hence emitting carbon.
The above prompt eventually led to the title of our recent paper, “From Clicks to Carbon: The Environmental Toll of Recommender Systems“.
If we do this, we do not explicitly mention it in the paper. You may assume that we use generative AI to find or improve any of our titles.
Paragraph-by-Paragraph proofreading
The probably most helpful feature of ChatGPT is paragraph-by-paragraph-based proofreading. Here is an example.
We do not blindly copy and paste suggestions from ChatGPT. Instead, we carefully review the suggestions, copy them where appropriate, and edit them as needed (as demonstrated by comparing the first paragraph of this blog post with ChatGPT’s suggestion and the original version). Additionally, we use Grammarly to refine the writing generated by ChatGPT further.
You can assume that we used ChatGPT to proofread any paper from around 2024 onwards. This is not always the case for the entire manuscript, but at least significant parts. We do not highlight paragraphs for which we did this. This is in line with the ACM policy that states:
If you are using generative AI software tools to edit and improve the quality of your existing text in much the same way you would use a typing assistant like Grammarly to improve spelling, grammar, punctuation, clarity, engagement or to use a basic word processing system to correct spelling or grammar, it is not necessary to disclose such usage of these tools in your Work.
Brainstorming & Feedback
We heavily use generative AI to do brainstorming, especially at the beginning of our work. Once we have a basic idea of what we want to do, we describe the concept in a few paragraphs and then ask, e.g. ChatGPT, to provide feedback. Similarly, when having a manuscript draft, we ask ChatGPT to review the draft and pretend to be a reviewer at a conference. We then decide if and how we incorporate the feedback of ChatGPT.
Promoting
Finally, sometimes, we use ChatGPT to create catchy Tweets for Twitter / X. Here is an example:
After some iterations, including generating an image with ChatGPT and Photoshop’s AI, this was the final tweet:
Summary
We use ChatGPT as a sophisticated writing assistant to enhance our research papers, grant applications, blog posts, emails, and other written communication. It helps us proofread, brainstorm, translate, title generation, and refine content. However, we do not rely on it uncritically. We carefully review, edit, and validate every suggestion from ChatGPT before incorporating it into our work.
ChatGPT serves as both a writing aid and a “sparring partner” that challenges our ideas and helps us refine our arguments. It is instrumental in improving our writing as non-native English speakers. That said, we take full responsibility for every word in our manuscripts. Every published work is our own, and we stand by its content.
While we acknowledge our use of AI tools, we follow established guidelines on disclosure and transparency. For example, when ChatGPT is used for translation, we explicitly state this. However, explicit disclosure is not standard practice for tasks like proofreading, where AI functions similarly to traditional writing tools like spell checkers. This aligns with policies from organizations such as the ACM.
We see ChatGPT as a valuable tool that enhances the clarity and effectiveness of our writing but does not replace our judgment or expertise. If you have thoughts or concerns about our approach, we welcome your feedback in the comments.
Cite
If you want to cite this document, please cite:
@InProceedings{Beel2024a,
author = {Joeran Beel},
booktitle = {Intelligent Systems Group, Blog},
title = {Our use of AI-tools for writing research papers},
year = {2024},
url = {https://isg.beel.org/blog/2024/08/19/our-use-of-ai-tools-for-writing-research-papers/},
}
0 Comments