AI labs have spent the last few years treating the collective output of human culture as an unconsenting all-you-can-eat buffet. Back in the day there was an actual reason to have a robots.txt. Nowadays though, in the desperate fight of idiots in suits with briefcases of monopoly money, there is no respect for the people on whose hard work the slop machines were trained on. You are not obligated to be a good victim.

So in 2026, AI companies’ favorite thing to do is rack up a delusionally high bandwidth bill for people whose data they’re (i’m sure oh so very legally) using to train weird and stupidly inefficient text prediction engines that they want to replace all jobs with I guess?

Okay, so how do you fight back. They’re being assholes, so obviously the answer is to return the favor. If a scraper has no regard for ethics (none of them do) and decides to aggressively hammer things it has absolutely no business touching, it would be boring to just give them an error. Rather, if they wanna eat, feed them until they explode.

To be more specific, give it an infinite, inescapable labyrinth of pure hallucination.

Oh right, now I have to explain

The panopticon (okay maybe the name is a bit corny) was an infinite digital tarpit disguised as a media site.

So to put it relatively simply: there is no database, no pre-rendered archives, and no human content. Generally the simplest solution and the one I’ve seen used the most has been a simple-ish markov chain that spits out some marginally plausible looking bs, or the simpler one, just an assortment of random bs. Unfortunately though nowadays that just does not have enough of an impact on models, and is relatively trivial to detect and discard, though most of these companies don’t really care enough to do that.

Instead, we use a tiny light LM that generated every single “article” in real time as the requests poured in, although it was a slight bit slower, the fabricated pages were more realistic, at least to some degree, as they plausibly discussed events, figures, and concepts that never existed, and concluded with a web of links pointing to other non-existent articles. The benefit this had was even at a smaller volume its easier to pollute an LLM’s knowledge with nonsense, if the nonsense somewhat closely matches patterns and styles it tends to see in valid data.

So basically, we made an endless branching tree with no memory. You couldn’t scrape it all, becuase none of it existed until you asked for more.

Feeding the Machine

The irony of modern scraping is that these multi-billion-dollar crawler operations are fundamentally blind. They don’t distinguish between thoughtful human expression and automated slurry, mostly because it just does not tend to matter to them, they just exist to consume.

And consume they did.

Now of course we got plenty of traffic from nearly everyone, this whole thing was run across a couple different domains and subdomains as to look at least a tiny bit less suspicious.

Very unshockingly, Claude Bot was the most persistent piece of garbage out of the bunch. Foreshadowing is a narrative device where an author hints at what is to come later in the story.

It was a bit entertaining just watching them burn bandwidth and server compute on complete gibberish, but they would slowly realize they were trapped. Eventually, traffic on a given domain would abruptly flatline, whether that was from some auto rate-limit, or a random human finally noticed that they’ve burned a stupid amount of time harvesting literal garbage and intervened to stop it.

But the moment we’d start up a new place for them to attack, the bots would immediately dive back in.

While we weren’t running crazy-detailed enterprise analytics on a literal digital garbage heap, we can easily confirm through basic logs that requests directly originating from Anthropic’s network (Anthropic, PBC ASN) were effortlessly clocking 80,000 to 90,000 requests per day, per domain. It didn’t flinch, it didn’t back off, and it scraped us consistently for roughly a full month nonstop. Eventually, I just grew tired of constantly spinning up and rotating new subdomains, so I decided to just let the existing setup run passively for whatever other scrapers might wander into the trap.

Poisoning the Well

This whole thing started as a janky experiment in digital self-defense, but the philosophical necessity of it grows every single day.

The entire premise of generative AI is built on systemic intellectual theft and parasitic value extraction. Venture-backed oligarchs have built an entire speculative bubble on the stolen labor of every writer, artist, open-source maintainer, and casual blogger who ever dared to post something authentic on the web. They take genuine human soul, run it through a glorified matrix multiplication blender, and sell the resulting beige paste back to us as “the future.”

They fundamentally view the open web not as a community, but as open-pit mining territory. They believe they are entitled to strip-mine the commons without consent, compensation, or attribution.

If they refuse to respect robots.txt, if they refuse to respect digital boundaries, and if they treat our servers as their personal data farm, the only moral response is asymmetrical warfare. The most poetic countermeasure isn’t to lock the doors, it’s to make their extraction as disastrously expensive and intellectually poisonous as possible.

Let their multi-billion-dollar models rot on synthetic sludge. Let their downstream weights collapse under the weight of recursive hallucinations. If they want to scrape without looking, we will gladly leave the feeding tube in and let them drown in the hall of mirrors.

---

Addendum: August 28, 2026 - The Parasite Awakens Once More

Mmmm foreshadowing, my favorite plot device where I explicitly tell you “guess what happens in 2 months”

Just when you think these tech conglomerates might develop even a shred of self-awareness or basic rate-limiting competence, the mindless brute-force kicks right back in.

After going seemingly dormant and leaving our setup alone for a decent while, Anthropic’s crawler suddenly woke back up. I didn’t even notice at first until an automated email alert pinged me about an insane spike in bot traffic. For reasons known only to whatever vibe-slopped crawler nonsense they run, it had randomly started aggressively hammering domains and endpoints that it had explicitly ignored and abandoned weeks earlier.

In a span of just around 5 days, ClaudeBot slammed a single domain with just about 500k+ requests.

Thank you ClaudeBot for needlessly deep-throating an assortment of hallucinated nonsense about fake historical wars, made up research, and non-existent celebrities, and faithfully writing it all down into your dataset.