I know there are tools that supposedly make it harder for AI to be trained on your images. But what about text or code? Is there any way to share things outside of a small group while minimizing the risk that some fucker will scrape them or otherwise use them for building an LLM?
Trying to prevent bits from being copied is always going to be futile.
In the long-ago times, websites that didn’t want crawlers to scrape them had a /robots.txt file that described what was and wasn’t allowed to be scraped.
But, the AI companies completely ignore that file. In fact, if anything, they may use it to find the “juicier” stuff to scrape. I know plenty of people who have decades of experience in running web servers who are going to extreme lengths (like banning entire countries, or huge IP ranges) to prevent what amounts to a DDOS on their servers from AI-company scrapers, because anything less extreme than that doesn’t work. For some of these people, more than 90% of the traffic they get is from AI scrapers. In fact, it seems like if you try to do something to discourage scrapers from AI companies, they think there’s something even juicier there and work even harder to get at it.
If they can find it, they’ll ingest it. And, these companies have spent literal billions on working around any attempt to prevent them from ingesting things, because they believe that whichever company manages to ingest the most content will come up with the best model.
If you have some kind of moral exception to your work ever ending up in an LLM, then you’re really out of luck. You better take up oil painting and never allow pictures of your art to be taken. The slop machines are hungry, and anything that becomes bits and ends up online will eventually be scraped.
As for making money as an artist, that’s another matter. It has always been hard, now it’s even harder. I don’t think you’ll make it easier for yourself by trying to prevent your stuff from being scraped. That just adds extra headaches you don’t need.
glue traps are a good solution to badly-behaved scrapers. just list a path in your robots.txt that slowly produces random text and links, all of which go to the same sort of page. if they respect robots.txt, they’re fine. if they don’t, they get stuck ingesting nonsense.
not a catch-all solution but a good tool.
AFAIK it was a glue trap / tarpit that one guy I know tried. Sure, it did result in the scraper getting a lot of useless data. OTOH, they dedicated massive resources to trying to get those non-existent documents, so his website was effectively DDOSed as a result. That’s why he’s now blocking entire countries.
yeah slowing down access and using a cheap generation method is important there.
Even that isn’t enough if they respond to that by just throwing more crawlers at your site.
yeah it’s not the only thing you want on there of course. i run a tarpit, fail2ban, and anubis.
Yeah, at one point he posted that fail2ban was only able to ban about 20 IPs/second, but he was getting 200 requests per second, and the bans weren’t having a meaningful impact on that (i.e. the requests were all coming from different IPs). From what I remember, his system load was over 1000.
oh wow. vps? they tend to live in busy ip blocks.
We can’t do anything until legislators act in the interests of their constituents and protect our rights with harsh penalties for companies infringing on our rights to our creative output. Unfortunately the bad guys have corrupted politics and political debate, controlling the media and influencers so it is challenging raising public awareness.
There is no technical solution that will resist attack. Create strong laws and put the fuckers in jail if they don’t comply. It’s the only sensible way.
Glaze (aka Nightshade) implants ‘poisoned’ data in your images that will ruin any data sets they’re put into (just learned to read, but leaving it up in case peeps don’t know about it).
Glaze and Nightshade are different methods. Glaze makes the AI incorrectly identify the art style and Nightshde makes the AI incorrectly tag the image content. The first one prevents the artist’s personal style to be learned by AI and the second one the specific image content (while also poisoning the model)
As a creator, I feel that my only recourse is to find an audience that will trust and respect me regardless of whether others steal my work. Foster that community. Sadly, there will always be plagiarists. But I hope communities (like on the fediverse) will become strong enough to be self-sufficient.
To share things with your friends, nothing better than P2P
(European options)
- https://www.sharrr.com/ (Swiss)
- https://pairdrop.net/ (Germany)
- https://hashtree.cc/ (Finland)
- https://syncthing.net/ (Sweden)
“Data wants to be free”
Sadly that’s true of data to train models as much as it is for piracy. It’s a double edged sword, otherwise you’re looking at implementing DRM.
For text maybe make your own Unicode scrambler - selecting readable letters from random sections of Unicode. Readable to a human but to an LLM they won’t be able to parse a word out of it. For code, maybe change the file extension and add metadata that is misleading? Or just don’t put your code online.
For the unicode scrambler: nah, as 30+ years of Captchas have proven, that’s the sort of thing that makes your stuff harder for humans to use/read forever while only temporarily inconveniencing the machines. Besides, pretty much anything you implement that might even kind of come close to working would pretty much destroy accessibility, meaning that handicapped people can’t read your stuff.
Also any invisible technique is broken by OCR, which is super easy to run nowadays. People will even do it to reuse the content and perhaps repost.
Anything that can be copied can eventually be leaked. However, if you are OK with just reducing the chance by 99%, a private repo/page with an “ask me for access” instruction keeps pretty much all current LLM scrapers out.
The short answer is that you can’t, really, anymore. You can put up barriers, but if your work is sufficiently “good” or popular, it will eventually leak and be ingested. Authentication (i.e. completely private websites) is the only real protection, but it has to be controlled by you or someone you trust, and in that case you’re severely limiting your reach/availability.
Encryption, just give the key to people in person.
Ask me and all my friends that ask me to whatsP And i say, no . . . lets simpleX. yeah nah. All the friends i only speak to in person. its fine though inertnrt is shite.
deleted by creator
