AI is creating the Play-Doh internet
By Mark Hurst • February 10, 2023

Everyone is trying to figure out ChatGPT right now, and more generally the entirety of the AI-generative tools out there that promise to deliver text, images, and music at the click of a button. Automated creativity! Frictionless research! Infinite media! These and other promises are circulating, to the delight of some (salivating investors) and the dread of the rest (who are likely face unemployment if the claims come true).

I already wrote one piece on the drawbacks of ChatGPT (Jan 13, 2023), focusing on the chat-bot’s design of giving the answer, which may or may not be true, rather than guiding the user to multiple sources for further study. I stand by everything in the column – including my Super Bowl prediction – but today I want to expand my argument a bit to encompass all of the AI generators.

We need to improve our understanding of this new thing, this combination of huge data sets (due to surveillance and automated web scraping) and pattern-matching algorithms, creating these automated outputs. It’s such a new phenomenon that we haven’t yet figured out how to describe it.

Alan Kay has this famous quote that if you want to teach something, you have to relate it to something you already know. So below, I’d like to suggest my favorite analogy to help explain the AI future we’re headed into.

AI as Play-Doh extruder

Play-Doh is a helpful way to think about AI generators. You may have used a toy extruder: you put a bunch of Play-Doh inside and then squeeeeeze, and out it comes in gloppy, mushed-up tubes. It’s fun as a toy, but maybe not a great approach for building the future of human civilization.

plastic Play-Doh extruder toy with threads of blue Play-Doh extending outward

ChatGPT is doing this to text content on the internet. The folks at OpenAI (now owned in large part by Microsoft) hoovered up huge bodies of text online like Wikipedia, as well as books, articles, and other sources. When you ask ChatGPT a question, it squeeeeezes all of that text together and plops out an answer, word by word, using its pattern-matching algorithm. Who knows what sources contributed to the answer, or what the algorithm was doing, or even whether the answer is accurate: it’s all just an extruded glop of text.

Could comedy writing work the same way? An AI generator of Seinfeld-like content, called “Nothing Forever,” was recently streamed 24/7 on Amazon’s Twitch platform (before being pulled off for spitting out offensive jokes). I started this week’s Techtonic episode with a few minutes of “Nothing Forever”: you can listen to it here; the episode page has links to more info. My point is that text, squeezed out into the vague shape of ChatGPT answers, or Seinfeld routines, is about as authentic as a pile of extruded Play-Doh pretending to be a sculpture by Rodin.

And it’s not just text. Music is getting the treatment, too. A couple of years ago on Techtonic I covered the OpenAI Jukebox, which hoovered up a giant body of actual music, then squeeeeezed it together to glop out some extruded music-like product. It’s uncanny valley for the ears: for example, here’s an example of Elvis Presley extruded product, from my interview with Ryan Walsh on the June 1, 2020 Techtonic.

Music extrusion has gotten more sophisticated since then. Over at Google (where they’re also working on a ChatGPT competitor called Bard), there’s something called MusicLM that generates music-like content: working from a text prompt, it draws from a library of actual music and squeeeezes it out, plopping out notes and robot-vocals to make something vaguely resembling a song. For example, here is a slow-tempo reggae-like music product that I played on Techtonic this week. As I point out, the vocals are nonsensical – though much clearer than the fuzzy Elvis “vocals” from OpenAI Jukebox in 2020. The technology is advancing.

As AI generators continue to gain traction, we risk finding ourselves in a Play-Doh internet – full of extruded content of varying quality, squeezed together from unknown sources, while the companies responsible – the same Big Tech beasts as always – claim no liability for any errors their extruders might have plopped out. This worries me enough that, as I mentioned briefly above, I devoted my Techtonic episode this week to “our AI generated future”: see the playlist, listen to the entire show, or download the podcast.

Finally, the idea of squeezing content came up in a related sense this week: Ted Chiang, writing in the New Yorker, compared the AI generators to image compression – the algorithms that compress (squeeze!) a JPEG image to take up less space, at the cost of some of the image’s clarity. From Chiang in the New Yorker (Feb 9, 2023):

Think of ChatGPT as a blurry jpeg of all the text on the Web. It retains much of the information on the Web, in the same way that a jpeg retains much of the information of a higher-resolution image, but, if you’re looking for an exact sequence of bits, you won’t find it; all you will ever get is an approximation. But, because the approximation is presented in the form of grammatical text, which ChatGPT excels at creating, it’s usually acceptable. . . .

[But] What use is there in having something that rephrases the Web? If we were losing our access to the Internet forever and had to store a copy on a private server with limited space, a large-language model like ChatGPT might be a good solution, assuming that it could be kept from fabricating. But we aren’t losing our access to the Internet. So just how much use is a blurry jpeg, when you still have the original?

Chiang makes an important point. We already have a way to write, to edit, to compose, to perform. We have each other: talented people who can create, and interested people who can interact with and share those works. The internet can be useful to all of us, creators and audiences alike, without turning into a Play Doh extruder.

Rephrasing Chiang, here’s my question: Why squeeze the life out of the internet, when we have everything we need already?

Simpsons image: 'Remove the stone of crypto. Attach the stone of generative AI'

P.S. As a bonus, I’m including the quote that Ross Douthat finished his newsletter with today. As I said up top, everyone is trying to figure out ChatGPT now. Walter Kirn offers some insight in Goodbye to the Future (Compact magazine, Feb 8, 2023):

What chatbots do is scrape the web, the library of texts already written, and learn from it how to add to the collection, which causes them to start scraping their own work in ever enlarging quantities, along with the texts produced by future humans. Both sets of documents will then degenerate. For as the adoption of A.I. relieves people of their verbal and mental powers and pushes them toward an echoing conformity, much as the mass adoption of map apps have abolished their senses of direction, the human writings from which the A.I. draws will decline in originality and quality along with their derivatives. Enmeshed, dependent, mutually enslaved, machine and man will unite their special weaknesses — lack of feeling and lack of sense — and spawn a thing of perfect lunacy, like the child of a psychopath and an idiot.

I can hear the objections to this dire scenario of a million gung-ho programmers as well as the ambitious A.I. itself, but I, a creative writer, am wed to it. I think dramatically first and scientifically second, such is my art. My ancient and possibly endangered art is imagining worst cases and playing them out to their bitter, tragic ends, as Sophocles did when he posited a king who unwittingly killed his father, married his mother, and then launched an inquiry into the matter after vowing to slay the perpetrator. See? See what writers were capable of then?

Now we have ‘Ant-Man.’ And worse, ‘Ant-Man’ sequels, enhanced by C.G.I.

Until next time,


Mark Hurst, founder, Creative Good – see official announcement and join as a member
Read my non-toxic tech reviews at Good Reports
Listen to my podcast/radio show:
Subscribe to my email newsletter
Sign up for my to-do list with privacy built in, Good Todo
On Mastodon:

- – -