LatheOperator

I can't believe I'm saying it but DON'T TOUCH GRASS

Is that why you're so hot or vice versa?

The thing is, if I don't add a "good enough for searching" OCR layer, someone else will - AI scraper or legitimate user. That's a cheap automatic operation. I'll better do it myself and poison it in a way that won't interfere with searching, like replacing some "are" with "aren't", such common words are rarely searched for. If there's a chance AI will fall for the metadata and invisible layer contents, that will decrease the requirements for visible poisoning, which is necessary but annoying.

Some of the text is indeed justified so I could do the multi-column trick that seems like the best compromise. The gap can be as narrow as one space. Or larger if I can write a script to detect lines and connect the columns with gibberish. A human can use zoom or window positioning to view one column at a time. I don't and will never have access to files the printouts are from (some are presumably in .602 format, others probably .doc), others are handwritten or typewritten, some have images glued on top or hand-traced; and Czech OCR is only about 99.5% reliable so as easier as it would make the endeavor, I can't be sure to preserve everything if I try to convert them into editable documents as an in-between step.

Woah, that's fucked up. But LLMs have very different performance by language so I think Czech responses are largely based on Czech-language data (maybe with training dataset augmentation with auto-translated works, judging by the literally translated technical terms). And I don't think this would happen in my country.

Yes but high school biology knowledge won't be all that helpful in surveillance tools. I'm targeting chatbots students might want to use to cheat. If an essay is less coherent or truthful, teachers will be able to more confidently punish the student or at least apply some extra scrutiny like a random oral exam on the topic (yes, those are still done here). Trust me that free AIs will become more limited once investor money dries up and paid tools will get price hikes, so if students see that the expensive model can't produce a Czech text that fools their biology teacher, they might just stop paying, restoring some of their information literacy and reducing corporate revenue a bit. I agree that the surveillance battle is real but that's not fought on this front. The major chokepoint for surveillance is government accountability (very low in today's US with ICE agents etc.), slightly less knowledgeable chatbots don't hinder it very much.

I think the AI boom will be over before it makes sense to scrape this, especially if I succeed at making it so that only a human fluent Czech speaker (or maybe a very bespoke program debugged by a Czech speaker) can see through the obfuscation (we have minimum wage so that would be costly). We don't have AGI and fully autonomous agent workforces as promised, and that's not gonna change if the remaining <20%, hardest-to-clean digital human-made data gets fed to them. Not to mention some was acquired illegally already, with pending lawsuits. The investors, including governments (thankfully not mine) will eventually realize that LLMs are simply not delivering nearly as much as they cost (including externalities unless the government is shit and passes them to people living near datacenters, laid-off programmers etc.) and never will, although that might take a while.

Yes, I will use both licence terms and "Made by bad AI" in metadata to discourage scraping but it's tempting to also poison the Czech-language biology knowledge base of bots who use the materials anyway. A good interleaving text that won't get filtered as off-topic might be multi-step roundabout machine translation of the original using shitty local tools.

Lbraries are cool but don't spread information very far, especially when "for local lending only, no copying". Those would almost never get used and probably thrown away as soon as the libraries realized the difficult copyright situation (not all are by my grandpa, many are unclear due to missing cover pages). Even when incompatible with screen readers (which someone criticized me for), a poisoned PDF is way more usable.

Anyway, the poisoning, outside the invisible layers and "transcript", will be visibly baked into the image too but obvious to humans (typewriter vs rendered text that inverts some statements with "not", "never" etc. at the ends of lines or between words in tiny font and scatters lines between paragraphs saying the text is unreliable, by random bad AI models, or just "clear AI tells" like "Sure, I can do that! 😉 Here's your summary:✨"). I don't think anyone will develop an OCR font filter for a few thousand pages in Czech to illegally defeat my DRM enforcing the included licence (unfortunately DRM is incompatible with CC so I can't use the "CC-BY-NC-SA" shorthand, not that scrapers, my enemies, respect it anyway): there is some protection in the obscurity.

Screen readers and LLMs both need plain text. Sorry, I'm not giving it to them. There will be an invisible OCR layer for searchability (and to discourage scrapers from re-doing OCR instead), like in many scanned PDFs, but poisoned. There's not many blind teachers anyway.

Yet, this is more accessible than what someone else suggested: making photocopies and donating them to libraries "for local lending only". Those would almost never get used and probably thrown away as soon as the libraries realized the difficult copyright situation (not all are by my grandpa, many are unclear due to missing cover pages)

That's expensive and a bit impractical to access, isn't it?

I want the poisoning to work on file level because it's inevitable and welcome for the documents to be shared between teachers and students for free on all kinds of existing platforms. I don't own any domains, anyway, and it might be best to scatter the documents around to make them harder to blacklist.

Edit: Google has both search crawlers and AI scraping bots. Even if both are separate, easy to filter or even abiding to robots.txt, the company has indicated that opting out of or hindering scraping will impact search ranking. Of course I'd use a throwaway gibberish $1 domain and couldn't care less about search ranking, but the power of their opaque, corporate algorithm is immense and maybe would spread to DNS blocking (they control 8.8.8.8). I don't want to play a cat-and-mouse game (and expect users of the docs to play along).

Good point. However, distillation (cannibalizing better LLMs) is a frequent technique so better convince the scraper it's not good even for that. That's why I plan to visibly digitally stamp OpenAI GPT 2.0 says: or This Deepseek response has been rated inaccurate: above some paragraphs of the typewritten text (of course in dozens of variations, maybe even different languages, to undermine search-and-replace). Humans will know it's fake (especially if I add a disclaimer) but scrapers, including ones that re-render and OCR the PDF themselves to get rid of misleading metadata and invisible layers, will most likely rate the text low in value. Of course ethically (and arguably legally) trained commercial AI would reject any text if it is released CC-BY-NC-SA 4.0 but I can't use that because scrapers for AI ignore licences in practice, and CC specifically forbids taking technical measures to devalue the text for some users.