I’m an investigative journalist focusing on technology, so I often hear people say that they haven’t found A.I. useful at all. That hasn’t been true for me. 

A lot of the work I do essentially amounts to querying databases, and A.I. can be very useful for that: like a Command-F you can use on anything, or a spreadsheet you can ask questions. But it’s recognizably still a computer — it will do what you tell it to do, and exactly what you tell it to do. This means that if you want to use it to manipulate a database, you have to construct the database and very carefully design the instructions for how you want the system to interact with it; conceptually, this is like writing in a programming language, just easier. But a lot of what you gain in ease, you lose in precision. This means that for me, A.I. is generally most useful as a tool for rough first passes on large amounts of data, which then have to be manually checked to make sure there weren’t mistakes as a result of my having given the wrong instructions. 

A.I. has greater utility for my more technical colleagues, largely because they have a more sophisticated understanding of what it can do and how to instruct it than I do. But while it’s an incredible tool, A.I. is not magic, and what they’re doing isn’t different in kind from my work. Part of what worries me is that politicians and executives who don’t, I think, understand it all that well see it as an end, not a means. At least in my area, it’s really best suited to brute-force tasks humans can do but that you wouldn’t assign them to because it would take too many of them, or to tasks where you don’t need perfect accuracy, and good enough is fine. This is exactly the opposite of the use cases tech executives commonly imagine for consumer applications, where a simple prompt can do something like produce a presentation, or perhaps raise your children for you so you have more time to work on presentations.

In my personal time, I largely use A.I. for doing web searches for things like recipes, where I’ve started to find search engines useless because sites are so overdesigned for robots. If I’m in the middle of making something and just want a rough ballpark for a proportion in a sauce I’m not likely to find in one of my cookbooks, I’ll use a chatbot rather than Google to save time. Because I’m not giving it careful instructions, though, the results aren’t reliable — A.I. is fine for a weeknight dinner, but not what I’d use to learn how to make something for my future mother-in-law.

I’ve also tried using A.I. as a reading companion, whether to find out what book a scene I only half remember is in or to find something in a long and complicated book. I would like to be able to ask ChatGPT something like, “Can you find me the scenes where the narrator makes an ass of himself in front of Mme de Guermantes?” but it never works; even though In Search of Lost Time is public domain and most assuredly in the bot’s training data, it always comes up with a lot of flatly wrong results delivered in a smugly patronizing tone with unwanted literary analysis attached. It’s not performing the search, but probabilistically generating text I’m likely to want to read. I do know that if I wanted to prompt the bot correctly, I’d feed in the text for it to index and then write a description tightly controlling the parameters for how to interact with that text and respond to my queries. That, though, would be work, and I don’t want to do that, because I’m reading to relax.