This is post number 6224 on my blog. I have no idea whether my previous posts have been part of the vast amount of internet content used to train ChatGPT. Probably not. But if it were, ChatGPT could easily reply to a request like "write a blog post about World of Warcraft in the style of Tobold". And the result would look pretty convincing. Now if you read some of my blog posts, you probably have some opinion on my intelligence, whether that is a positive or a negative opinion. But that opinion is something that has formed in your brain, based on the very limited information you have about me from my blog posts. The same leap to judgement would happen if you read a ChatGPT blog post: If it sounded reasonable and was well argued, you would attribute some "intelligence" to its author, ChatGPT. Thus we talk about the progress of artificial intelligence, because we can see outputs that resemble the output of intelligent people. But that is all fake.

I have written enough blog posts about MMORPGs and World of Warcraft for an artificial intelligence to be able to recognize patterns in that wealth of data, and to be able to reproduce something superficially similar. If you used the same data and asked ChatGPT to write a blog post about the NFL, the result would look a lot more like garbage, because I never wrote about that subject, and so there isn't much data to go on. ChatGPT would still have a lot of information about how I use the English language, and how I tend to write, but with no data on the subject, the "write a blog post about the NFL in the style of Tobold" request would probably yield something extremely vague, lacking logical coherence about the subject matter.

You also couldn't ask ChatGPT to write a "review in the style of Tobold" about a newly released game I haven't talked about yet. Unless you feed it with both my blog, and a bunch of reviews from other people about the game, in which case ChatGPT could project my writing style on the information about the game it got from elsewhere. But if the other reviewers liked the fast action sequences of the game in question, ChatGPT might well repeat that in fake Tobold's review, and regular readers of mine might spot that to be inconsistent with my preference for slower games.

The surprising thing is that the limited ability of ChatGPT to write a text based on pattern recognition and repeating what other people have said can very well be all that is needed for certain applications. A teacher giving students the homework to write an essay about some subject doesn't expect the result to be brilliant and original. You can easily let ChatGPT write that for you, and just edit it in case there are some mistakes a human wouldn't have made. I don't know if you ever googled for some sort of "game guide" information about some game, but if you do, you would find that there are already a lot of sites out there which seem to have copied and pasted that sort of information. It wouldn't be hard for ChatGPT to replace some of the people writing for that sort of publication. Even scientific publications (of which I have written a bunch) tend to start with a more general introduction to the subject area, e.g. climate change, and that general part could easily be written by ChatGPT.

A lot of articles about ChatGPT have been written this year in which the author prodded ChatGPT with questions until ChatGPT answered with something that sounded creepy, or emotional, or self-aware. You can get ChatGPT to devise a plan for robots to take over the world. And that makes for a nice, click-bait headline about "artificial intelligence plans to take over the world!". But that only looks scary because we tend to attribute thinking and intention to what other people say. ChatGPT producing the plan for robots to overthrow the rule of humanity doesn't know what it is writing, nor does it have any intention behind that. It just assembled a text, based on user prompts and pulling from existing texts in its database.

There are areas of human content creation, e.g. on social media sites like YouTube, where humans deliberately adjust their style to the algorithm the site uses to recommend content. An artificial intelligence application that was specifically trained on YouTube recommendations would be quite good at writing a script for a YouTube video that would have a lot of success. The same artificial intelligence would be unable to write of script for a YouTube video that was original and highly intelligent. Because artificial intelligence can't even recognize real intelligence; it can however recognize patterns of success and reproduce them. That says more about our patterns of consumption of online content than it says about the ability of machines to think.

- posted by Tobold @ 9:28 AM Permanent Link