evade the machine

(Repost, because the first time apparently angered the robots.) Once more unto the breach, dear friends (sigh)—this time, we'll look at the ways AI keeps eating the internet (and itself 🐍).

-

It's a well known fact that Big Tech scrapes material to train their LLMs—but what does scraping actually look like? What exactly goes into it?

Even before their use of the controversial “Books 1” and “Books 2” datasets was exposed, Big Tech’s data appetites extended deep into the open web. In a 2023 investigation, the Washington Post analyzed Google's C4 dataset, which was used to train their early LLM models. The Post found content from over 15 million websites, including heaps of original writing: news articles, academic writing, personal and public blogs, and (of course) fanfic were all represented.

Much of this writing came from well-known and well-loved hosting sites: Tumblr, AO3, scribd, b-ok.org, Patreon, Kickstarter, medium, WordPress, Blogspot, LiveJournal, etc.

The C4 dataset itself was sourced from a large database belonging to Common Crawl, a nonprofit who's whole raison d'etre is to scrape the internet bare. Nearly every name in Big Tech has dipped into Common Crawl's enormous database to train their AI models (OpenAI, Meta, and so on—all of whom "donate" to the non-profit for privileged access.)

Common Crawl is definitely a bit… problematique. They've been known to conceal contents in their archives; and shoot down simple requests for attribution of scraped content: "It's not our job," said Rich Skrenta, its executive director, in conversation with The Atlantic.

And once the data's scraped into the archive, it's nearly impossible to remove. Skrenta said: "you can't delete anything from [the Common Crawl archive]. (Which is why AO3 received a big no when it requested for all scraped data to be removed in 2023.)

If that weren't awful enough, the Common Crawl's databases are actively making the internet (even) worse, more dangerous, more racist, more generally fucked up. In a 2023 paper analyzing datasets used to train image-generators, researchers found that models trained on one of the Common Crawl datasets were "five times more likely […] to label black male faces as criminals." Another study in the same year discovered that a different Common Crawl data set "contained thousands of images of verified and suspected [CSAM]." (Karen Hao, Empire of AI, 137)

Since 2023’s blowback, the behemoths have gotten hyper-secretive about the datasets they use to train their AI. We don't know what OpenAI’s GPT-4o was trained on, much less Sora or GPT-5.

We also know that in 2024, LLM trainers hit a data-wall. Meaning, the models exhausted all the available data on the internet, pushing them to find new, and non-public, sources of data.

Given the secrecy of it all, we should all be, hmm........ a lil skeptical of any claims of transparency.

And then, of course—there's Google. When you’ve spent decades cataloguing the internet (and when you effectively own the archive), training your own model—Gemini—comes with certain… conveniences.