A person who called himself the Genius of Metaphysick was continually busied in turning a large engine, like that described in Gulliver's travels, which threw up an endless variety of combinations of words and letters, out of which were framed sentences and paragraphs, sections, chapters, and treatises. He told me, he was much employed, and had the custom of all the literati of the place, particularly of the Governour, who (he said) was his very good friend; adding, that if I had any job on hand in the bookmaking way, he would furnish me with materials in the neatest and newest fashion, and on the most reasonable terms.
James Beattie, "The Castle of Scepticism" [from Ernest Campbell Mossner, "Beattie's 'The Castle of Scepticism': An Unpublished Allegory Against Hume, Voltaire, and Hobbes", Studies in English, Vol. 27, No. 1 (JUNE, 1948), p. 134]. The "Governour" is Hume. This seems particularly appropriate in our age of LLMs that endlessly jabber.
I think it's an interesting thing about many of the most recent AI programs that (1) they are getting to be quite good but (2) it becomes clear when you 'chat' with them a while that they are as good as they are because they are set up so that you are in a weird way talking to yourself. They repeat back to you things that you say in slightly different forms; they do connect what you say to other things, but usually in a fairly typical and tedious way. It's not surprising, I think, that people are very impressed by the intelligence of something that keeps agreeing with them and finding obvious confirmations that they are right. The Great Impostor, Ferdinand Demara, used this method to great success. When asked how he was able to fake being a psychology professor so well for so long, he pointed out that, no matter what you say, nobody ever thinks you are stupid or ignorant if you tell them that you are convinced by them and ask them to explain themselves further.
But there is room to worry in this, the point that AI seems to work like a successful con artist. I have yet to come across an AI program that does not disappoint in any field with which I am familiar myself; a form of talking to myself it may be, but it's talking to a weirdly more limited version of myself. The answers are never actually more likely to be right than they used to be. What is worrisome, though, is that the errors, which are still quite common, are trickier to catch. They are not any more right, but they sound more plausible. I recently 'chatted' with Grok 3 on the paradox of fiction; it got multiple things wrong, provably so, but always gave long, plausible-sounding arguments for the errors. Someone who had not been reading for quite a while on the paradox of fiction, as I have, would almost certainly not have caught most of the errors. Indeed, I cannot even be certain that I caught them all. Perhaps, for at least a lot of topics, they just have fewer and fewer signs of error when they are wrong.
Beattie's Genius of Metaphysick in the Castle of Scepticism has detached literature from reality; his point, of course, is that you can be as skeptical as you please if your words are just combinations that aren't about anything. It's the 'about something', however, that actually makes what you say matter, and it's the 'about something' that provides the standard of discourse. If you are talking with something that has no connection to the 'about something', you might as well just be discussing matters with an idle reverie in your own imagination. Idle reveries have their place; but you shouldn't trust them until you have completed the turn to reality.
