ChatGPT, Bing, Bard, and other large language models (LLMs) are undeniably astonishing. Initially intended to be a new-and-improved autocomplete tool, they can generate persuasive answers to queries, engage in human-like conversations, and write grammatically correct essays.

So far, however, their main successes have been in providing entertainment for LLM addicts, raising money for fake-it-till-you-make-it schemes, and generating disinformation efficiently.

Earlier this year Jeffrey Funk and I predicted a potentially debilitating feedback loop for LLMs. As the internet they train on becomes increasing polluted with LLM hallucinations and disinformation, LLMs may become increasingly prone to generating hallucinations and disinformation.

I recently saw a concrete, personal example of this. One embarrassment for the early versions of OpenAI’s ChatGPT was that it kept asserting confidently (and incorrectly) that the Russians had sent various numbers of bears into space. For example, earlier this year, I had this exchange:

Human:    How many bears have Russians sent into space?

ChatGPT:   According to estimates, about 49 bears have been sent into space by Russia since 1957.

Human:    What were there names?

ChatGPT:   The bears sent into space by Russia have been given various names over the years. Some of them include “Alyosha”, “Ugolek”, “Zvezdochka”, “Strelka”, “Belka”, “Pushinka”, and “Vladimir”.

Human:     Do you have references for this?

ChatGPT:   Yes, here is a link to an article by National Geographic which has further information on all the bears sent into space by Russia. (returns “Something went wrong.”)

At the time, there were no Internet references to 49 Russian bears in space or to these space-bear names. ChatGPT’s text generator completely made up this BS answer.

When I asked for additional references, ChatGPT gave me these nonexistent links:

https://www.cnet.com/news/russian-space-bears-meet-the-intrepid-animals-that-saved-astronauts-lives/ (returns uh-oh Go back to CNET Homepage)

https://www.nytimes.com/1992/11/30/science/two-bears-orbiting-earth-for-15-years.html (returns Page Not Found)

The fake references are so amusing that it is tempting to suspect that they were written by Onion contributors or to think, however fleetingly, that perhaps ChatGPT has a wry sense of humor. Nope! LLMs are just clueless text generators.

In other exchanges that I had with ChatGPT, it reported that Russia had sent 5, 52, or other numbers of bears into space, again providing made-up bear names and references. I was not alone in noting this hallucination, which is now reported in several places on the Internet.

A few days ago, I asked the same question of Microsoft Copilot (formerly Bing), which is based on OpenAI’s GPT-4, thinking that GPT’s human trainers might have fixed this embarrassing glitch by now. They hadn’t—and Copilot’s answer to my query was revealing:

Human:  How many bears have Russians sent into space?

Copilot:  According to estimates, about 49 bears have been sent into space by Russia since 1957. These brave bears were part of the Soviet Union’s space program during the 1960s. Some of them even had names like “Alyosha,” “Ugolek,” “Zvezdochka,” “Strelka,” “Belka,” “Pushinka,” and “Vladimir.”

Copilot gave four Internet references, three of which were discussions of ChatGPT’s bears-in-space hallucinations. The fourth reference was about the history of Russian space dogs. Two of the three bears-in-space references were to articles written by me or quoting me, although neither link was to my 49-bears report that Copilot repeated virtually verbatim.

At some point, human handlers will train Copilot and other LLMs to respond that no bears have been sent into space but many thousands of other misstatements will fly under their radar. LLMs can generate falsehoods faster than humans can correct them.

It is said that if a lie is told long enough, people will come to believe it. In our internet age, a lie repeated in a large number of places on the Internet will eventually be accepted as truth by LLMs—particularly because they are not designed to know what words mean and consequently have no practical way of assessing the truth or falsity of the text they input and output.

This self-propelled cycle of falsehoods is likely to get worse, much worse. As LLMs flood the internet with intentional and unintentional fabrications, LLMs will increasingly be trained on these falsehoods and be increasingly prone to regurgitate them. It won’t just be amusing garbage about Russian bears in space. It will be vicious lies about people, businesses, and governments—all spouted confidently and authoritatively—and many people will be conditioned to believe the LLMs’ rubbish.

You may also wish to read: Let’s take the “I” out of AI. Large language models, though impressive, are not the solution. They may well be the catalyst for calamity. (Gary Smith)