Opportunities and challenges with ChatGPT in research

Writing a post about ChatGPT is always challenging, as the tool quickly changes and evolves, and the community is equally fast at finding new and better ways to use the tool.

At the time of writing this post, GPT 4o has been released, and I have played around with it only a little bit. Indeed, one of the challenges of mastering ChatGPT as a tool is that it takes a fair amount of time to really understand what the tool can do, find suitable ways in which we can leverage the tool at work, and then keep up to date as the tool continuously morphs.

Today, I will be sharing uses of ChatGPT (or any other generative AI tool of your liking) for research. I must say that I haven’t found yet a way to drastically alter the way I do research by using ChatGPT – and I think that is a good thing.

In fact, I strongly disagree with those researchers who advocate for using ChatGPT to do our literature review 100x faster. I think that taking the time to read and ponder the literature is an essential step for us as researchers. Yes, ChatGPT could read 300 papers in a few minutes and summarize all the information for us, but we will not develop a deep understanding of the literature ourselves. And developing this understanding is a crucial step in being able to generate new ideas, finding a gap in the literature, and being able to write a discussion.

With this concern and observation off my chest, here are a few ways in which I have tried using ChatGPT for my research:

If anything, ChatGPT is good with text. For a non-native English speaker as myself, ChatGPT can help. When the writing of a paragraph of a paper is difficult, I have tried copying and pasting this paragraph into ChatGPT and have asked it to explain it to me in plain English, as well as explain the main concepts discussed in the paragraph. This approach has helped my reading understanding, and allowed for a bit more interaction than looking things up in the dictionary.

While most publications are in English, I have used ChatGPT to help me have a look at papers in Portuguese and Chinese. I’ve asked it to translate passages, summarize passages, and I have also asked it questions such as: “what do the authors in this passage say about theory X or the use of technique Y?”. Being able to ask questions and interact with the text has been more helpful than throwing text into Google Translate.

I have used ChatGPT to read all the information from a call for funding proposals, and asked it to make a full list of keywords for me. I double-checked this information with my own reading and understanding of the call as well, and I was quite pleased with what ChatGPT had cooked up for me.

I don’t really want to use ChatGPT for doing writing and research (I do not want The Machine to do my job, after all), but I have uploaded the documentation of a call for proposals and chatted with the document to see which types of expenses could be covered and which ones not. I learned that I could expense specialized software, but not standard office software, and not the computer itself. I also confirmed it by doublechecking with the grant support office.

ChatGPT uses Python for its data analyses, so it is good at writing code. While I do not want ChatGPT to do all the programming for me, I have asked it to propose a way to write certain routines. I still prefer to outline a program myself, but then have asked it to write typical code for formatting plots or for suggesting a more efficient alternative to some clunky code that I wrote.

More than using ChatGPT for writing new code, I have used it for debugging code. As I have started to move away from Matlab and use Python instead, I am not that familiar with the way Python gives me errors. So, I have asked ChatGPT to read my code, the error warning, and to tell me what the error means, and how I should fix my code.

I have been playing with the data analysis options of ChatGPT by throwing large datasets in there and asking it to create interesting visualizations. For now, nothing really of my interest came out of this analysis, but it seems like something I could use to get new insights in a dataset. I would be hesitant to use the visualizations directly, but it could help get an initial idea on what to explore further.

I have used ChatGPT to let it provide critical comments on a draft of my writing. Similarly, I have used it to upload a concept note of a proposal together with the proposal call documents, and asked it to analyze if perhaps I am failing to address a certain aspect of the call, or if I should strengthen a certain aspect in the methodology. It took a bit of probing ChatGPT before it let go of its chirpy politeness and gave some thoughtful feedback on what to improve.

While I prefer Grammarly for checking my writing, I have asked at times to spellcheck my writing. Mostly, I feel that my typical Office built-in tools are good enough to spellcheck my writing, and I expect that these built-in tools will soon become much better as they will start to use AI as well.

Sometimes, when I try to shorten my writing, I end up writing very long sentences. At times, I find it hard to simplify my writing. Most of the times, I try to just chop the sentence up in different parts, but recently I have tried asking ChatGPT to simplify and shorten some long paragraph that I wrote and that I didn’t fully like. I also didn’t really like what ChatGPT gave me, but between two disliked versions I did manage to write something better.