TL;DR: Please help me fuck (with) AI. See bold sections
Hi,
I haven't been keeping up with anti-AI combat so I'm asking for help. I inherited thousands of pages of materials my late grandpa made or used for his grammar school teaching job in the 1990s-2000s. They are A4 pages of documents made using what seems to be a typewriter, Text602 (DOS rich text editor) and Word. They were most likely not all made by him but he treasured them in nice binding and they have sources (mostly books and journals, almost no webpages, and absolutely no AI) and a cursory look shows meticulous compilation of every important fact on each subject (frankly, the level of detail is excruciating and I'm glad I went to a different grammar school). There's obviously no original scientific research but the materials can still be useful to someone, I bet. They were almost thrown away by the widowed grandma (she already removed and disposed of the plastic bindings and front covers so I'll have to guess document titles) but I think grandpa would prefer them to be shared. With an ADF scanner and OCR software (I have no chance of accessing the work computers he used so I'll have to scan), I can quickly make searchable PDFs of each document, and share them via torrent and DDL sites (there are Czech sites dedicated to sharing teaching materials but they have paywalls or an upload-credit system so best avoid them, not to mention some materials contain newspaper clippings and textbook photocopies for images so best stay anonymous and not try to assert copyright).
I'm afraid these texts could become a major part of some commercial LLM's Czech-language biology/social sciences knowledge corpus unless poisoned. How to best reduce the value of the documents when people try to feed them to AI (training/rewriting) with them while keeping their value for most legitimate users? (Sorry, people with screen readers, there may need to be extra steps for you.) I'm thinking about adding a huge volume of thesaurized or otherwise fuzzed public domain text like f4mi did with .ass subtitles (a technique that would probably still work if she didn't get 1M views detailing it, making YouTube reduce subtitle formatting support). Prompt injection or replacements (cell→gnome) might be interesting too. However, tools I know add an extra PDF layer, which is too obvious. I'm thinking about adding tiny text in the header and footer or between paragraphs in the OCR layer (not overlaid to reduce interference when selecting/searching), but how? I need an automated way to do this with such a huge page count. I can use both Linux and Windows machines for the job. None of them are very powerful but speed is not a concern, it's summer break and nobody will need school materials until September. I'll be happy to include multiple layers and techniques to make them too frustrating to remove.
The paper smells musty but does not seem to be moldy. It's all blank on the other side so I'll interleave it with recent newspaper to allow for the odor-neutralizing chemicals to seep into the sheets so I can eventually reuse them.
Illustration pic is an actual sheet from the collection, to make the post more engaging. Of course I won't be adding watermarks like that, that would just aggrevate people and make them try extra hard to extract the actual content. (And this one is easy to remove with color channel mixing.)
You don't understand just how shit AI is when asked about school topics in Czech. For example, here is a bit of Czech language litany every third grader must know or they will embarrass themselves with awful spelling mistakes. (Skip the bullet points if you just want to hear about the AI)
IandY(and long versionsÍ/Ý) sound the same [ɪ] ([ɪː]) unless preceded byD,T, orNbut using the wrong one is a big no-no. (Yis never a consonant in Czech)I(Í) followsC,J,Č,Ř,ŠandŽ, whileY(Ý) followsH,K,R. ConsonantsQ,WandXbasically don't occur andG,Ď,Ť, andŇare never followed byIorY. Foreign words are a huge mess of course, as evident by the existence of the Spelling Bee (we don't have that, Czech is phonetic with just a few difficult bits likeI/Y).B,F,L,M,P,S,V,Z. They are mostly followed byI(Í) but there is a list of about 15 common exceptions on each (vyjmenovaná slova orBY-FY-LY-MY-PY-SY-VY-ZYwords), plus their relative words, whereY(Ý) is written instead. For example, there are just 4ZY-words so I'll just post the list so you'll get an idea:IZY-words so it staysWell, you'd expect AI to know all cca 100 exception words by heart because they're public domain and the most famous piece of third grade teaching material (like times tables in second grade) that barely changed in 100+ years so almost every Czech could recite them as a kid? Hell no. There's dozens of screenshots where Gemini or ChatGPT spewed utter nonsense instead. (DuckDuckGo does not appear to search corporate social media for images because they're not providing direct links to the files). Granted, some are from users asking for nonexistent

XYandHYwords but so many are unforced errors. I can't find my favorite, a Reddit post where Gemini listed dozens of variants of babička with all kinds of endings like Italian "babičetto" before just adding "etc." but a close second are ones where it adds non-Latin scripts:Does the apparent incompetence stop Czech students from cheating with AI? Nope. But the longer the AI stays obviously terrible, the better.
I deal with less-common languages a bit, and switching between languages is surprisingly common. I think it's due to how token embedding works; as soon as it finds a character in another language where the semantic distance is lower than the language-barrier distance it'll just flip over.
In principle it shouldn't be that hard for model creators to fix, but I don't have the kind of compute resources needed to do it myself. :/