Today’s guest post is from Dieter Zinnbauer of the Copenhagen Business School’s Sustainability Center:

Jim Anderson over at the World Bank blog and Matthew Stephenson on this blog kicked of an interesting discussion about how the new era of artificial intelligence—particularly the natural language chat-bots like OpenAI’s revolutionary ChatGPT—will affect the anticorruption field. As Matthew suggested, the ability of ChatGPT to generate plausible-sounding (if a bit bland) summaries and speeches on corruption-related topics should inspire all of us real humans to aim to do more creative and original—and less bot-like—writing and speaking on anticorruption topics. And both Jim and Matthew suggested that in this field, as in many others, ChatGPT can also be a valuable aid for researchers and advocates, performing in seconds research and drafting work that might take a human being several hours.

Yet while ChatGPT may be able to assist in some tasks, we shouldn’t get too excited about it just yet, especially when it comes to research. Some of its limits as a research tool are already well known and widely discussed. But I wanted to call attention to another problem, based on a couple of recent experiences I had trying to use ChatGPT as a research aid.

First off, I’ve been working on a paper about works and what doesn’t work in anticorruption trainings. I asked ChatGPT to do a write-up on this. What it produced was pretty good—not really at the level of the more imaginative and comprehensive discussion that a human expert could provide, and that I tried to offer in my paper, but still something quite acceptable. But what really got me excited when I followed up with ChatGPT by asking it to supplement its overview with references to recent academic papers, including working papers. When I did so, the bot pointed me to an article published in a well-known academic journal that, from the title and ChatGPT’s summary, looked directly on-point, and that I had completely overlooked. Here’s a verbatim reproduction of the citation and summary that ChatGPT provided:

The Effect of Anti-corruption Training on Ethical Decision Making: A Randomized Controlled Trial” by Sjoerd Beugelsdijk, Joost Herman, and Jasper Vrugt. Published in the Journal of Business Ethics, Vol. 146, No. 3 (Apr., 2017), pp. 463-478.

This article presents the results of a randomized controlled trial investigating the effect of anti-corruption training on ethical decision making. The authors find that the training had a positive impact on ethical decision making and reduced corruption-related behaviors.

Wow, this was really impressive. This is one of the main journals in my area of expertise and I had completely overlooked what sounds like a super-relevant piece of work. All kudos to ChatGPT for suggesting an important contribution that could really inform and improve my work. There was only one problem:

The reference turned out to be completely made up.

The Journal of Business Ethics has an article about conflict minerals in the volume and page range that ChatGPT indicated, and there is no trace of the article that ChatGPT had referenced anywhere at any time in the journal. The authors are real, but one of them works on earth engineering, one on humanitarian action, and the third in international business. None of them lists a related publication and they seemed to have never published together. One author confirmed by email that he had no clue about this. Google scholar cannot find any article of that name either.

I also had ChatGPT to provide me with a list of references to articles that employed the concept of “ambient accountability,” a term that I coined and wrote about a few years back. Yet again, ChatGPT gave me some really cool recent references, causing me to become temporarily excited about the continued popularity of my idea. Alas, it turned out—to my chagrin—that several of the references that ChatGPT provided were again completely made up.

I doubt this problem is limited to research on corruption and anticorruption, but because many people in this field are getting excited about the potential of ChatGPT and similar tools to aid them in their research—as demonstrated by Jim and Matthew’s posts, and the interest they generated—I wanted to sound a note of caution based on my experience. Of course, it’s still the early days of this new technology. The learning process will be ongoing, both for the machines and the people who use them. But as of now, it seems that when ChatGPT doesn’t know the answer to a question, it sometimes—alas, like more than a few human researchers—will resort to just making things up.