From Virginia’s Newsletter:
I know two kinds of people: those who have been obsessively playing with and discussing ChatGPT and those who have at best a hazy notion that it exists. I’m in the obsessive group, as you already know if you read the Tennyson experiment I posted earlier.
For those in the hazy group, ChatGPT is a system that uses massive amounts of text to create a predictive model that enables it to mimic human writing. The shorthand is that it’s an AI chatbot, or autofill on steroids. You type in a request and it spits out an answer. This CNET column provides a solid backgrounder:
For example, you can ask it encyclopedia questions like, “Explaining Newton’s laws of motion.” You can tell it, “Write me a poem,” and when it does, say, “Now make it more exciting.” You ask it to write a computer program that’ll show you all the different ways you can arrange the letters of a word.
Here’s the catch: ChatGPT doesn’t exactly know anything. It’s an AI that’s trained to recognize patterns in vast swaths of text harvested from the internet, then further trained with human assistance to deliver more useful, better dialog. The answers you get may sound plausible and even authoritative, but they might well be entirely wrong, as OpenAI warns.
Even in its current, relatively primitive form ChatGPT portends both huge productivity increases and major disruptions in any enterprise in which writing matters. Instead of writing boilerplate corporate memos, managers will soon assign them to bots. The run-of-the-mill college grads who get paid to flood my mailbox with press releases and promotional emails should start thinking about careers as nail techs or phlebotomists—something in the physical world. Insight and beauty are still rare, but serviceable prose isn’t.
With the right prompts, ChatGPT can already mimic routine political journalism and quotidian high school and college essays. “What GPT can produce right now is better than the large majority of writing seen by your average teacher or professor,” writes Daniel Herman, a humanities teacher at Maybeck High School a small independent school in Berkeley, in The Atlantic.
I asked the program to write me a playful, sophisticated, emotional 600-word college-admissions essay about how my experience volunteering at my local SPCA had prepared me for the academic rigor of Stanford. Here’s an excerpt from its response:
In addition to cleaning, I also had the opportunity to interact with the animals. I was amazed at the transformation I saw in some of the pets who had been neglected or abused. With patience and care, they blossomed into playful and affectionate companions who were eager to give and receive love. I was also able to witness firsthand the process of selecting the right pet for the right family. Although it was bittersweet to see some animals leave the shelter, I knew that they were going to a loving home, and that was the best thing for them.
The application essay is vapid but convincing. The variety of word choice (“blossomed,” “bittersweet”) and sentence structure marks it above average. “Had the opportunity to” is a stylistic tell: Here’s a privileged person who’s been taught to sound grateful rather than to write succinctly. “I was also able to…” is the same. I’m sure admissions officers see thousands of such essays every year. If their value goes to zero thanks to automation, this reader, writer, and teacher won’t object.
While crashing the value of mediocrity, ChatGPT could increase the returns to excellence. (“Average is over,” as Tyler Cowen put it.) Think about what happened to graphic design. Many people used to make a living doing routine tasks, from laying out pages to selecting typefaces, that are now easily handled by software. Thanks to the graphic intelligence embedded in everyday tools, the standards for routine graphics, from websites and PowerPoint presentations to restaurant menus and wedding invitations, have increased.
. . . .
As I write, there are 28 student papers awaiting my grading attention. I doubt any used ChatGPT, partly because mentioning it in class produced mostly blank stares. (The most tuned-in student, however, said he’s started using it in place of Google.) Already, we’re getting confirmed reports of cheating on exams given on Canvas, the web-based system used by many colleges for assignments and grading. By next term, every class will have to take account of ChatGPT, either explicitly incorporating it as a starting point or going back to handwritten tests and essays.
The kind of mediocre writing that earns grade-inflated Bs is now replaceable by a bot. Maybe if those B-essay students started with AI-generated prose it would be easier to teach them to do better: to refine the ideas, dig down more on the facts, improve the writing style. Can ChatGPT be a time-saving tool, like a calculator or text search, rather than a threat?
Link to the rest at Virginia’s Newsletter
Ah, I would like to think the ChatGPT would be a teaching tool in the hands of the right student or teacher. It did do a serviceable job in Postrel’s example, and it’s an example of why I understand why artists are threatened by AI art: in the hands of a skilled user, you get “good enough results.” Good enough to have people ask why they should pay you to do what the machine can do, anyway.
And for the moment it would seem the above average writers will just shrug off the AI. But AI is only as good as its programming, and I’m really curious what ChatGPT would come up with if fed a steady diet of, e.g., Postrel and David McCullough, Agatha Christie and Terry Pratchett. Especially Pratchett! That’s when this topic will really become interesting. But I strongly suspect that an AI that can write or “think” on that level will remain science fictional.
I hope Postrel continues with her experiments. I’m in the middle of reading her book on the history of textiles. I miss her Deep Glamour blog, where her posts helpfully explained the difference between glamour and charisma. For me it helped explained why footage of charismatic cult leaders — the ones who get everyone to kill themselves — just leave me cold. Charisma is a “you have to be there” quality, whereas glamour works at a distance and is immortal, e.g., the everlasting glamour of Marilyn Monroe. Postrel has eclectic interests and is always interesting to read for that reason alone.
What happens when AI gets fed a steady diet of AI?
In other words, what happens when the vast majority publicly available text on the internet is created by AI, and then used to train AI?
Hmm, I wonder if ChatGPT uses poorly written buying/how to/other click bait sites and such for training?
I don’t know about that, but it apparently uses Archive of Our Own, so it’s being fed a steady diet of fanfic. The list of reasons to fume just keeps getting longer.
It would be interesting to see what it does if an author feeds it an unpublished story and tells it to edit it in the style of Tolkien, Heinlein, Christie, etc, depending on genre. The software can’t actually rewrite the story but it can rearrange it, maybe fix wordsmithing issues…
It may not obsolete writers but it might replace proofreaders and other pre-pub players. Useful for Indies.
More immediately useful: clearing out CAPTCHAS. 😉
“While crashing the value of mediocrity, ChatGPT could increase the returns to excellence.”
This is one of the better takes I have seen. ChatGPT can produce a grammatical paragraph of pablum. I have seen nothing to suggest that we will soon be seeing an AI that can produce good, factual, insightful writing.
“Instead of writing boilerplate corporate memos, managers will soon assign them to bots.”
I foresee much hilarity when those memos go out without any literate person reviewing them first.
I foresee much hilarity when those memos go out without any literate person reviewing them first.
Academic journals?