How Much Can You Trust AI?

The other day, I had a patron request for a newspaper article that seemed to link Charles “Pretty Boy” Floyd to a homicide in central Arkansas in 1913, when the future bank robber would have been nine years of age. My correspondent insisted that a CALS Encyclopedia of Arkansas entry cited this particular newspaper article, but I couldn’t find it. I even went to the trouble of searching through Newspapers.com and Chronicling America’s digitized newspaper collections, but to no avail.

When I asked my correspondent to give me the exact source from which she got this information, she replied that it was ChatGPT. This did not surprise me. OpenAI admitted last year that AI “hallucinations”—that is, outright lies delivered with utmost confidence—are mathematically inevitable. But more and more people are using these chatbots and accepting as gospel the answers to any questions they ask, and more and more of us in the library world are dealing with the effects of that credulity. Those who work in interlibrary loan, for example, are regularly dealing with requests for books that simply don’t exist, titles completely fabricated by whatever AI chatbot people are using.

Can any of these AI programs tell real information from fake? No, not really. For example, the Encyclopedia of Arkansas posts a yearly April Fools’ Day entry. These entries and their accompanying media remain preserved on the site, in a special April Fools’ Day section, with subheads that designate each entry as an April Fools joke, and though a few people fall for them each year, you would think that a computer program that literally scans each and every word of a piece would take the very obvious hints. But that is not what happens at all.

When I did a search using Google for the history of caber toss in Arkansas, it spat back at me an “AI Overview” that uncritically pulled from the EOA April Fools entry on caber toss:

Then I asked Google if any Arkansas poet might have influenced Edgar Allan Poe’s composition of “The Raven” (a reference to our foolish entry on “The Possum”). Here was Google’s answer:

Not only did Google’s AI pull from our joke entry, but it even got the contents of that entry wrong, describing the poem as Albert Pike’s (we wrote that he published it but did not write it), and claiming that in both “The Raven” and “The Possum,” an ominous creature perches near the entrance of the room, while actually, in the latter, said possum perches upon a jug of moonshine.

Google takes our little entries a lot more seriously than we do. But the problem goes beyond Google. Grokpedia, Elon Musk’s AI-driven competitor to Wikipedia, is also rife with misinformation, taking EOA joke entries as the real thing. For example, its entry on state legislator Madison K. Moran almost exclusively pulls from our joke entry on the Breeches Panic.

And it makes you wonder, if these high-powered exemplars of generative AI cannot tell that something is a joke, even when it is clearly marked as a joke, how can we trust them to deal with misinformation, propaganda, and all the other stuff flooding the web? Worse, as more and more people publish information rife with AI hallucinations, other AIs will harvest that information, eventually degrading the information ecosystem and producing model collapse (also called Habsburg AI).

These are fundamentally not reliable sources of information.

Of course, I understand why people are turning to some of these chatbots—the utility of Google has become so degraded of late that it can be impossible to find what you need. However, there exist other search engines. DuckDuckGo has become my standard and even as a NOAI version. There are also Ecosia, SearXNG, and Qwant.

And I understand that most people probably don’t really know that these AI models are glorified autocompletes, rather than any kind of actual machine intelligence. Thus do I recommend for you the following sources of information relating to AI:

• Better Offline. Headlined by Ed Zitron, this podcast regularly features two episodes a week, one of which is a monologue usually in the ten-minute range, the other an interview with one or more people. Although broadly tech-oriented, the podcast has in recent years focused much more on AI.
• Ed Zitron’s Where’s Your Ed At. The regular newsletter of Ed Zitron, with many longform pieces offering in-depth analysis of the claims offered by AI companies.
• Tech Policy Press. A nonprofit exploring very broadly the intersection of Big Tech and democracy, with a regular newsletter and podcast.
• 404 Media. A journalist-founded digital company exploring tech in all its aspects; they have done some groundbreaking pieces on AI.
• Paul Kedrosky. A research fellow at MIT’s Institute for the Digital Economy, as well as an investor and researcher in his own right. He regularly examines the economics of AI.
• DAIR Institute. The Distributed AI Research Institute seeks to implement governance frameworks for AI systems that eschew theft and exploitation.
• AI as Normal Technology by Sayash Kapoor. A Substack newsletter regularly breaking down the myths and hype of Artificial Intelligence.
• The AI Resist List. Activists calling out AI development “empires” and offering a different vision for a collective informational future.

The capacity of AI to sabotage our informational ecosystem is really only a small part of the larger concern about implementing AI systems across various sectors of our society. But it is perhaps the part that most concerns us here at the CALS Encyclopedia of Arkansas, as people dedicated to making a true account of state history available to the reading public. Granted, if anyone had taken me up on my challenge to produce their own “top” list for Arkansas, I’d be writing about that right now, rather than writing about the shortcomings of AI yet again. But here we are.

An artificial sweetener can be just as sweet as sugar, but no artificial intelligence actually demonstrates any capacity for intelligence. Any AI model that cannot tell the difference between history and a joke is itself a joke—and will soon be history.

Guy Lancaster is the editor of the CALS Encyclopedia of Arkansas.