Fabricating Reality with ChatGPT

Last month, a non-profit group called OpenAI made available an artificial intelligence program called ChatGPT that has educators worried that students will use it to write fake essays. ChatGPT relies on information it can find on the World Wide Web up through the end of 2021 and is able to converse on a wide range of subjects.

When it first came out, I was doing research on the history of steam locomotive technology for my other blog, Streamliner Memories. I asked it some questions that I already knew the answers to and it made some reasonably intelligent replies.

More recently, I’ve been looking at the economics of ocean liners, as the Canadian Pacific Railway was a major operator of ocean liners in both the Atlantic and Pacific markets. The history of technologies such as railroads and steamships is well documented but the economics are not. Canadian Pacific spent (in today’s dollars) hundreds of million of dollars buying steamships that sometimes made only 120 or so voyages across the oceans before being scrapped. How could this pay off?

So I asked ChatGPT, “Is there any literature or research on the economics of ocean liners in the late 19th and early 20th centuries?”

“There is a significant amount of literature and research” on the subject, it responded. “One example of a book that discusses the economics of ocean liners is The Financing, Building and Operation of Ocean Liners by J. E. Wilton-Jones.”

That sounded like exactly what I was looking for. But when I looked for the book on Bookfinder.com (which has the inventories of hundreds of on-line booksellers), Worldcat.org (which has the catalogs of thousands of libraries), Google, DuckDuckGo, Project Gutenberg, and archive.org, none of them had ever heard of it.

I went back to ChatGPT and asked who published the book and when. “According to my knowledge, The Financing, Building and Operation of Ocean Liners by J.E. Wilton-Jones was published by the Institution of Mechanical Engineers in London in 1920.” That led me to contact Adrian Clement, the librarian of the Institution of Mechanical Engineers.

“I’m afraid I can find no record of this book in our collection,” he replied. He added that he checked the catalogs of the Lloyd’s Register Foundation, the U.K.’s National Bibliographic Database, and WorldCat, and was unable to find the book in any of those places either.

Returning to ChatGPT, I asked where it got its information about this imaginary book. It “is a well-known book in the field of maritime history, and I was programmed with knowledge about it and other similar works,” it replied.

I also asked for other references on the economics of ocean liners. ChatGPT was happy to help.

“Sure! Here are a few sources on the economics of ocean liners in the late 19th and early 20th centuries:
“1. The Economics of Ocean Liner Shipping, 1820-1939 by Eric Grove is an in-depth look at the economic factors that influenced the development of the ocean liner industry during this period.
“2. The Golden Age of Ocean Liners by John Maxtone-Graham is a comprehensive history of the ocean liner industry from the late 19th to the early 20th centuries, with a focus on the technological, social, and economic factors that shaped the industry.
“3. The Ocean Travel and Cruise Ship Industry by David Kirkpatrick and David Shapland is a more recent book that provides an overview of the history and economics of the ocean liner industry, including a discussion of the key economic factors that have shaped the industry over time.”

Eric Grove was a naval historian who wrote many books about ships, but as far as I can tell none were about ocean liners. John Maxtone-Graham wrote a book titled, Crossing and Cruising: From the Golden Era of Ocean Liners to the Luxury Cruise Ships of Today. Someone else, Lee Server, wrote the book titled The Golden Age of Ocean Liners. Finally, there are authors named David Kirkpatrick and David Shapland, but I can’t find any books they wrote together and neither wrote about ocean liners. The closest is a book by Kirkland about aquatic turtles.

About this time, I learned that ChatGPT was insisting to someone that Hillary Clinton was the 45th president of the United States and that Donald Trump had never been elected. It also fabricated parts of someone’s resume.

So I asked ChatGPT to tell me about myself. “Randal O’Toole is a policy analyst and author who focuses on land use and transportation issues,” it responded. “O’Toole is known for his criticism of light rail and other forms of mass transit, and has argued that the construction of such systems is often wasteful and inefficient.” So far so good. “He has also written about the dangers of suburban sprawl and the need for more compact, walkable communities.” Wait a minute!

OpenAI warns that ChatGPT “may occasionally generate incorrect information.” But this is more than occasional. While implying that I believe there is a need for more compact, walkable communities might be dismissed as merely incorrect, completely fabricate imaginary books seems much more insidious.

I’m not one to worry that Skynet is going to try to take over the world and exterminate humanity. But I find it hard to believe that OpenAI would deliberately program ChatGPT to fabricate books and other references. Until its creators can give a reasonable explanation of why it is doing that and how they will fix it, I will have much deeper worries about AI than that students will try to use it for writing essays.