AI Slop and Human Play

Computers can win at chess and Go, but how about a rap battle?

Ideogram.ai

“AI slop” is AI-generated content—text, video, photos, ads—that is poorly executed, generally annoying, notably inauthentic, and overly abundant. I’ve started receiving a lot of emails from services wanting me to hire them to promote The Fabric of Civilization. Some of these emails seem to represent real people who are using AI to write. Others are unsigned and don’t appear to have a human (even a fake one) in the loop. They’re just playing the game of large numbers. All seem to be scraping Amazon and regurgitating book descriptions found online. I imagine many of the press releases clogging my email box are similarly composed, although it’s hard to tell the difference between lame press releases written by humans and lame press releases written by AI. Neither gives any thought to the exact interests of the recipient.

The proliferation of AI slop recalls the proliferation of spam, which at the turn of the century threatened to make email nearly useless until Bayesian spam filters came along. At first you had to install them separately. Now they’re built into your email system, along with increasingly sophisticated AI filtering. So spam is rarely a problem today. I suspect the solution to AI slop will be an AI agent who identifies and filters it for you.

Some AI critics see slop as evidence that human creativity will be drowned out by junk, others that human creativity will be swamped by human-quality AI competition. I’m enough of a snob to believe that much human “creativity” will indeed be replaceable by AI. But you don’t have a inalienable right to get paid for routine brain work.

I’m also optimistic about the ways in which generative AI can enhance and enable human creativity. I recently met the founders of Locunity, a media startup offering the local news that internet advertising has largely killed off. AI is key to their business model. The first step to establishing local coverage is feeding recorded public meetings into AI that can spin out stories on what the city council, the zoning board, or the board of supervisors did at their most recent meeting. Locunity is betting that enough people will have a direct financial interest in this information that they’ll be willing to pay for an easily digestible, well-organized newsletter. Once Locunity has enough area subscribers to its AI-only content, it will add a human reporter to do more enterprising stories.

At the Abundance conference, Daniel Golliher, whose Maximum New York is a must-read if you’re interested in cities (especially NYC), told me about how he used the AI tool V0 to create a “NYC Rent Guidelines Board Simulator,” using information on real New York buildings. It invites users to set increases on rent-stabilized and market-rate apartments and see what happens to building economics. In an email, Daniel explained:

• It shows a variety of different buildings in NYC with different shares of market versus rent-regulated units.

• The user gets to set the rate of increase on both rent regulated and market-rate apartments in the app. In real life, the Rent Guidelines Board only sets rent regulated unit increases. The RGB can generally only set one rate for all regulated units, despite the kind of building that they're in.

• This allows users to see how the "one size fits all" rent regulation system actually impacts different buildings differently, and relatively pushes costs onto market-rate units. So "freeze the rent" has real impacts one can't get away from, and it plays out differently for different buildings.

To create it all he needed was V0 and his own knowledge of how the rent regulation system works. “It (maybe) illustrates a point: building good tools like this requires having a knowledge base you can play with through the tools,” he writes. “Sometimes people feel blocked when trying to build these, but then aren't once you point them toward domains they know a lot about.” AI makes it possible for someone with a deep domain knowledge—in this case of the rent regulation system—and minimal if any software chops to create an impressive educational tool. Humans plus AI can do things that neither can accomplish alone.

Take “Bro-Botz” videos. Their creator Greg Beato caught my attention with a LinkedIn post in which he explained that he’d been fooling around with Suno AI for a while, writing songs, but decided to go for something more ambitious:

Partially I was inspired by the ongoing discourse over "AI slop" and the presumption that any use of generative AI renders content suspect, invalid, unworthy of attention. And partially I just figured it would be a fun diversion and a way to explore this concept for a character, or rather, characters, that I've had for a while now….

Who are they? A pair of state-of-the-art humanoid robots who aren't actually all that interested in leveling up to superintelligence or even AGI. Envious of their counterparts from Boston Robotics, Tesla, and Unitree Robotics, ChadGPT and Joe Brogan, aka the Bro-Botz, mostly just want to create dance videos that go viral on YouTube and rack up social media clout.

So here they are. Claude and I wrote the lyrics. Suno AI and I created the music. ChatGPT and I generated the base images that we fed to VEO to generate video clips in 8-second takes. Then I used Adobe Premiere to compose the complete video. The results were pretty sloppy, but also, I would humbly assert, somewhat entertaining.

One of the most interesting things about the process was that I also relied on ChatGPT and Gemini to help me decipher Photoshop and especially Premiere (which I'd never used before). By the end of a week of editing, I knew so much more than when I'd started, and will use that knowledge to approach my next effort with something a little more akin to an actual workflow that will ideally lead to more character design continuity, better lip syncing, more consistent quality control. Overall, in other words, a higher quality of slop!

I wanted to know more about exactly how much was Greg and how much was Claude et al. As I suspected, the AI tools were exactly that: help for realizing Greg’s vision. He did the serious creative lifting. He wrote the lyrics with occasional help finding robot-related words (e.g., actuator) and confirming technical details. But jokes, cultural savvy, and clever word play are still human territory.

As we talked, we came to an interesting conclusion. We know that AIs can win at chess and even Go. As hard as they are, these are games with clearly articulated, mathematically precise rules dictating a finite number of possible moves and board positions. But a rap battle is harder, and AIs may never be able to master the form. It’s not just that there are many more words than squares on a board but that rappers constantly alter and play with the words and their meanings. “It’s funny,” Greg remarked, “It’s funny, because in theory language is their forte, and it is—but so far at least, not novel language in any reliably compelling way.”

AIs can’t play. But they can help us play.

Recommended links:

What History Can Teach Us About Copyright, AI, and ‘Market Floods’: “If AI is only used to create an abundance of content truly devoid of human creativity—what some call “slop”—then old-fashioned human authorship will become comparatively rare and thus more valuable, not less. If, in contrast, artists and authors use AI to generate creative works that add new cultural ideas, then that would seem to be very much in line with copyright’s core purpose; these new works may legitimately compete with existing ones, but that does not necessarily justify protectionism for old-fashioned techniques of authorship.”

George Eliot Archive: An amazing resource that includes access to everything George Eliot published—fiction, poetry, translations, and nonfiction—as well as commentaries by her contemporaries, many of her letters, images, and more. One noteworthy feature: “Our major born-digital project, the George Eliot Text Explorer, has been in development since 2022 and represents the first machine-readable, open-access version of Eliot’s complete published works. By converting PDF documents into TEI-encoded XML files, we’ve created a foundation for advanced text analysis tools, including the AI Analysis of Eliot's fiction (Chen & Rilett, 2023) and the Text Explorer application (Cui, Sun, and Rilett, 2023-24), which enables instant keyword and phrase searches. The project is further enhanced through integration with Voyant Tools, an open-source text visualization platform.”

The Grift Artist: An ode to “donald boat,” who “replies to famous people on X demanding that they buy him things—expensive video game equipment, books, energy drinks—and they do.’

Building in the world of flesh needs regulatory transparency: Ruxandra Teslo on how making some pretty basic information public could dramatically improve the prospects for new drugs, especially from smaller companies.

A CTD (Common Technical Document) is the international format for drug approval submissions. It contains everything: the chemistry, manufacturing, and controls (CMC) data that so often delay biotechs and cause hundreds of millions in lost revenue, complete clinical protocols and datasets, and the detailed back-and-forth between companies and FDA reviewers.

I recently saw this for the first time, thanks to Brian Finrow of LumenBio, who generously shared with me the CTD he had acquired for $25,000 from a failed biotech. Thousands of pages of what look like “boring details” are in fact the practical playbook of how drugs succeed—or fail—through the FDA.

For small innovators, this kind of information is invaluable. Large pharmaceutical incumbents already have private archives of CTDs and decades of institutional know-how. That gives them a structural advantage. This is confirmed by empirical results from the economics of innovation. When it comes to the biopharmaceutical industry, regulatory complexity favors large incumbents.

The post surfaces the kind of in-the-weeds knowledge that can lead to serious reform but, as Ruxandra notes, the FDA can’t do much on its own.