This is very hacker-like thinking, using tech's biases against it!
I can't help but feel like we're all doing it wrong against scraping. Cloudflare is not the answer, in fact, I think that they lost their geek cred when they added their "verify you are human" challenge screen to become the new gatekeeper of the internet. That must remain a permanent stain on their reputation until they make amends.
Are there any open source tools we could install that detect a high number of requests and send those IP addresses to a common pool somewhere? So that individuals wouldn't get tracked, but bots would? Then we could query the pool for the current request's IP address and throttle it down based on volume (not block it completely). Possibly at the server level with nginx or at whatever edge caching layer we use.
I know there may be scaling and privacy issues with this. Maybe it could use hashing or zero knowledge proofs somehow? I realize this is hopelessly naive. And no, I haven't looked up whether someone has done this. I just feel like there must be a bulletproof solution to this problem, with a very simple explanation as to how it works, or else we've missed something fundamental. Why all the hand waving?
Your approach to GenAI scrapers is similar to our fight with email spam. The reason email spam got solved was because the industry was interested in solving it. But this issue got the industry split: without scraping, GenAI tools are less functional. And there is some serious money involved, so they will use whatever means necessary, technical and legal, to fight such initiatives.
I love the insanity of this idea. Not saying it's a good idea, but it's a very highly entertaining one, and I like that!
I've also had enormous luck with Anubis. AI scrapers found my personal Forgejo server and were hitting it on the order of 600K requests per day. After setting up Anubis, that dropped to about 100. Yes, some people are going to see an anime catgirl from time to time. Bummer. Reducing my fake traffic by a factor of 6,000 is worth it.
As someone on the browsing end, I love Anubis. I've only seen it a couple of times, but it sparks joy. It's rather refreshing compared to Cloudfare, which will usually make me immediately close the page and not bother with whatever content was behind it.
Same here, really. That's why I started using it. I'd seen it pop up for a moment on a few sites I'd visited, and it was so quirky and completely not disruptive that I didn't mind routing my legit users through it.
Quite possibly. Or, in my case, I think it's more quirky and fun than weird. It's non-zero amounts of weird, sure, but far below my threshold of troublesome. I probably wouldn't put my business behind it. I'm A-OK with using it on personal and hobby projects.
Frankly, anyone so delicate that they freak out at the utterly anodyne imagery is someone I don't want to deal with in my personal time. I can only abide so much pearl clutching when I'm not getting paid for it.
The Digital Research Alliance of Canada (the main organization unifying and handling all the main HPC compute clusters in Canada) now uses Anubis for their wiki. Granted this is not a business, but still!
It’s a feature in the paid version, or I guess you could recompile it if you didn’t want to pay (but my guess is if you want to change the logo you can probably pay).
As someone on the hosting end, Anubis has unfortunately been overused and thus scrapers, especially Huawei ones, bypass it. I've gone for go-away instead which is similar but more configurable in challenges
My experience with it is that it somehow took 20 seconds to load (site might've been hn-hugged at the time), only to "protect" some fucking static page instead of just serving that shit in the first place rather than wasting CPU on... whatever it was doing to cause delay
That’s so many scrapers. There must be a ton of companies with very large document collections at this point, and it really sucks that they don’t at least do us the courtesy of indexing them and making them available for keyword search, but instead only do AI.
It’s kind of crazy how much scraping goes on and how little search engine development goes on. I guess search engines aren’t fashionable. Reminds me of this article about search engines disappearing mysteriously: https://archive.org/details/search-timeline
I try to share that article as much as possible, it’s interesting.
So! Much! Scraping! They were downloading every commit multiple times, and fetching every file as seen at each of those commits, and trying to download archives of all the code, and hitting `/me/my-repo/blame` endpoints as their IP's first-ever request to my server, and other unlikely stuff.
My scraper dudes, it's a git repo. You can fetch the whole freaking thing if you wanna look at it. Of course, that would require work and context-aware processing on their end, and it's easier for them to shift the expense onto my little server and make me pay for their misbehavior.
Or some anti-ddos/bot companies using ultra cheap scraping services to annoy you enough to get you into their "free" anti bot protection, so they can charge the few real ai scrapers for access to your site.
Even if there isn't (yet?), there's probably someone who's honestly thinking this is potentially a viable business model and at least napkin-mathing it out.
My napkin mathing is that their ROI would be negative. That's a lot of compute and bandwidth they'd have to pay for even if they were just throwing away the results.
But there is a lot of search engine development going on, it's just that the results of the new search engines are fed straight into AI instead of displayed in the legacy 10-links-per-page view.
> There must be a ton of companies with very large document collections at this point
See, I don't think there is, I don't think they want that expense. It's basically the Linus Torvalds philosophy of data storage, if it's on the Internet, I don't need a backup. While I have absolutely no proof of this, I'd guess that many AI companies just crawl the Internet constantly, never saving any of the data. We're seeing some of these scrapers go to great length attempting to circumvent any and all forms of caching, they aren't interested in having a two week old copy of anything.
Ohai! I'm working on dataset poisoning. The early prototype generates vapid LinkedIn posts but future versions will be fully pluggable with WebAssembly.
Hi Xena! Your blog is amazing! Didn't realize you're working on Anubis - it's a really nice tool for the internet! Reminds me a bit of the ye' olde internet for some reason.
You've made one of the best solutions, that matched what I thought of implementing myself, and at the time it was most needed. I think a couple of "thank you" are sorely missing in this comment section.
It's slightly different in subtle ways. If I recall iocaine makes you configure a subprocess that it executes to generate garbage. One rule I have for Anubis in the code is that fork()/exec() are banned. So the pluggable garbage generator is gonna be powered by CGI handlers compiled to WebAssembly. It should be fun!
Probably any detail that you think is cool, I would be interested in reading about. When in doubt err on the side of too much detail.
That was a good read. I hadn’t heard of spintax before, but I’ve thought of doing things like that. Also “pseudoprofound anti-content”, what a great term, that’s hilarious!
I think that most people don't do this, and the ones that do have custom solutions. Xena's uses cron, but that's all I know. It's probably a custom shell script.
>I love the insanity of this idea. Not saying it's a good idea, but it's a very highly entertaining one, and I like that!
An even more insane idea -- minding the idea here is porn is radioactive to AI data training scrapers -- is there is something the powers that be view as far more disruptive and against community guidelineish than porn. And that would be wrongthink. The narratives. The historic narratives. The woke ideology. Anything related to an academic department whose field is <population subgroup> studies. Alls you need to do is plop in a little diatribe staunchly opposing any such enforced views and that AI bot will shoot away from your website and lightspeed
I like this better than of NSFW links; just include a (possible LLM generated) paragraph about not supporting transitions in minor children. Or perhaps that libraries that remove instructional booklets for how to have same-sex intercourse aren't actually banning the books.
That sort of thing; nothing that 80% of people object to (so there's no problem if someone actually sees it), but something that definitely triggers the filters.
This is a cute idea, but I wonder what is the sustainable solution to this emerging fundamental problem: As content publishers, we want our content to be accessible to everyone, and we're even willing to pay for server costs relative to our intended audience -- but a new outsized flood of scrapers was not part of the cost calculation, and that is messing up the plan.
It seems all options have major trade-offs. We can host on big social media and lose all that control and independence. We can pay for outsized infrastructure just to feed the scrapers, but the cost may actually be prohibitive, and seems such a waste to begin with. We can move as much as possible SSG and put it all behind cloudflare, but this comes with vendor lock in and just isn't architecturally feasible in many applications. We can do real "verified identities" for bots, and just let through the ones we know and like, but this only perpetuates corporate control and makes healthy upstart competition (like Kagi) much more difficult.
If the LLMs are the "new Google" one solution would be for them to pay you when scraping your content, so you both have an incentive, you're more willing to be scraped and they'll try to not abuse you because it will cost them at every visit. If your content is valuable and requested on prompts they will scrape you more and so on. I can't see other solutions honestly. For now they decided to go full evil and abuse everyone
The only way that would work is if they were legally required to. And even then, it probably wouldn’t work unless failure to comply was a criminal offense. You know what? Even then it still might not work.
or turn your blog into a frontend/backend combo. keep the frontend as an SPA so that the page has nothing on it. have your backend send data in encrypted format and the AI scrapers would need to do a tonne of work in order to figure out what your data is. If everyone uses a different key and different encryption algorithm suddenly all their server time is busted decrypting stuff
Ehm what would stop ai scrapers from using a browser like a normal user would? Google bot already does, it can execute js and can read spa client side generated content, so it proves can be done at scale, and I'm pretty sure some ai scrapers already do
What if scrapers ips are millions of smartphones? If I was as evil as an AI scraper company that is not obeying robots.txt I would totally build/buy thousands of small games/apps for mobiles to use them as jumphosts to scape the web. This is probably happening already.
in my case my application does not use pagination, it uses infinite scroll, even if you had a million devices that use google chrome, they would all load page 1 and if that req/minute progressively decreaasing thing is implemented, once they start scrolling endlessly they would all hit the rate limits sooner or later, the thing is a human is not going to scroll down a 100 pages but a bot will. once this difference has been factored it, it wont matter how many unique devices they bring into the battle
so why not just do that for these scrapers, rather than complicate it by encrypting and decrypting, which is just obfuscation as the private key is clearly available to the end-user?
tbh i did not encrypt decrypt for the ai scrapers at all, a lot of people were previously trying to download data directly from my API that my frontend uses and this kinda pissed me off a bit. So I added encryption/decryption to the mix and will release the newer version. As I mentioned earlier as well, can someone sit through and decrypt it? yes. Will 99% of them do it? no! Thats where I win
At this point it seems like the problem isn’t internet bandwidth, but just expensive for a server to handle all the requests because it has to process them. Does that seem correct?
I own a forum which currently has 23k online users, all of them bots. The last new post in that forum is from _2019_. Its topic is also very niche. Why are so many bots there? This site should have basically been scraped a million times by now, yet those bots seem to fetch the stuff live, on the fly? I don’t get it.
I have a site with a complete and accurate sitemap.xml describing when its ~6k pages are last updated (on average, maybe weekly or monthly). What do the bots do? They scrape every page continuously 24/7, because of course they do. The amount of waste going into this AI craze is just obscene. It's not even good content.
It would be interesting if someone made a map that depicts the locations of the ip addresses that are sending so many requests, over the course of a day maybe.
If you are in the US, have you considered suing them for robot.txt / copyright violation? AI companies are currently flush with cash from VCs and there may be a few big law firms willing to fight a law suit against them on your behalf. AI companies have already lost some copyright cases.
Based upon traffic you could tell whether an IP or request structure is coming from a not, but how would you reliability tell which company is DDOSing you?
It should be at least theoretically possible: each IP address is assigned to an organisation running the IP routing prefix, and you can look that up easily, and they should have some sort of abuse channel, or at the very least a legal system should be able to compel them to cooperate and give up the information they’re required to have.
Large scale scraping tech is not as sophisticated as you'd think. A significant chunk of it is "get as much as possible, categorize and clean up later". Man, I really want the real web of the 2000s back, when things felt "real" more or less... how can we even get there.
A curated web directory. Kind of like Yahoo had. The internet according to the dewey system with pages somehow rated for quality by actual humans (maybe something to learn from Wikipedia's approach here?)
Because of the financial incentives, it would still end up with people doing things to drive traffic to their website though, no? Maybe because the web was smaller, and people looked at it as means "to explore curiosity" in the olden days it kinda worked differently... maybe I just got old, but I don't want to believe that.
By “doing things to drive traffic to their website” do you mean trying to do SEO type things to manipulate search engine rankings? If so, I think that there are probably ways to rank that are immune to tampering.
Don’t worry, you’re not just old. The internet kind of sucks now.
Google was neat in that you didn't see the content keyword spam either on the websites or the portal home pages. The Web was already full of shit (first ad banner was 1994? By 1999 you already had punch the monkey as classy content), but it was more ... organic and you could easily skip it.
Have you ever listened to the 'high water mark' monologue from fear and loathing? It's pretty much just that. It was a unique time and it was neat that we got to see it, but it can't possibly happen again.
Thanks for reminding me about that, what a great monologue. I didn't really understand it when I was younger, but now I feel the same thing with regards to software engineering. There was a golden age which finally broke at the end of the 2010's.
The bots are exposing themselves as Google, Bing and Yandex. I can’t verify whether it’s being attributed by IP address or whether the forum trusts their user agent. It could basically be anyone.
Interesting. When it was just normal search engines I didn’t hear of people having this problem, so this either means that there are a bunch of people pretending to be bing google and yandex, or those companies have gotten a lot more aggressive.
There are lots of people pretending to be Google and friends. They far outnumber the real Googlebot, etc. and most people don't check the reverse DNS/IP list - it's tedious to do this for even well-behaved crawlers that publish how to ID themselves. So much for User Agent.
User agent has been abused for so long, I forget a time when it wasn't.
Anyone else remember having to fake being a Windows machine so that YouTube/Netflix would serve you content better than standard def, or banking portals that blocked you if your agent didn't say you were Internet Explorer?
Normal search engine spiders did/do cause problems but not on the scale of AI scrapers. Search engine spiders tend to follow a robots.txt, look at the sitemap.xml, and generally try to throttle requests. You'll find some that are poorly behaved but they tend to get blocked and either die out or get fixed and behave better.
The AI scrapers are atrocious. They just blindly blast every URL on a site with no throttling. They are terribly written and managed as the same scraper will hit the same site multiple times a day or even hour. They also don't pay any attention to context so they'll happily blast git repo hosts and hit expensive endpoints.
They're like a constant DOS attack. They're hard to block at the network level because they span across different hyperscalers' IP blocks.
How do you define a user, and how do you define online?
If the forum considers unique cookies to be a user and creates a new cookie for any new cookie-less request, and if it considers a user to be online for 1 hour after their last request, then actually this may be one scraper making ~6 requests per second. That may be a pain in its own way, but it's far from 23k online bots.
That's still 518.400 requests per day. For static content. And it's a niche forum, so it's not exactly going to have millions of pages.
Either there are indeed hundreds or thousands of AI bots DDoSing the entire internet, or a couple of bots are needlessly hammering it over and over and over again. I'm not sure which option is worse.
Imagine if all this scraping was going into a search engine with a massive index, or a bunch of smaller search engines that a meta-search engine could be made for. This’d be a lot more cool in that case
Whatever the forum software Woltlab Burning Board considers a user. If I recall correctly, it tries to build an identifier based on PHP session ids, so most likely simply cookies.
This is exactly my point. Scrapers typically don't store cookies, so every single request is likely to be a "new" user as far as the forum software is concerned.
Couple that with 15 minute session times, and that could just be one entity scraping the forum at 30 requests per second. One scraper going moderately fast sounds far less bad than 29000 bots.
It still sounds excessive for a niche site, but I'd guess this is sporadic, or that the forum software has a page structure that traps scrapers accidentally, quite easy to do.
When you have trillions of dollars being poured into your company by the financial system, and when furthermore there are no repercussions for behaving however you please, you tend not to care about that sort of "waste".
Remember the 90s when viagra pills and drug recommendations were all over the place?
Yeah, I use that as a safeguard :D The URLs that I don't want to be indexed have hundreds of those keywords that are leading to URLs being deindexed directly. There is also some law in the US that forbids to show that as a result, so Google and Bing are both having a hard time scraping those pages/articles.
Note that this is the latest defense measurement before eBPF blocks. The first one uses zip bombs and the second one uses chunked encoding to blow up proxies so their clients get blocked.
You can only win this game if you make it more expensive to scrape than to host it.
edit: Oh it was very likely the Federal Food, Drug and Cosmetic Act that was the legal basis for the crackdown. But that's a very old law from the pre-internet age.
edit 2: Might not have been clear for the younger generation, but there was a huge wave of addicted patients that got treated with Oxycodone (or OxyContin) subscriptions at the time.
I think that might have been the actual cause for the crackdown on those online advertisements, but I might be wrong about that.
Nice! Reminds me of “Piracy as Proof of Personhood”. If you want to read that one go to Paged Out magazine (at https://pagedout.institute/ ), navigate to issue #7, and flip to page 9.
I wonder if this will start making porn websites rank higher in google if it catches on…
Have you tested it with the Lynx web browser? I bet all the links would show up if a user used it.
Oh also couldn’t AI scrapers just start impersonating Googlebot and Bingbot if this caught on and they got wind of it?
Hey I wonder if there is some situation where negative SEO would be a good tactic. Generally though I think if you wanted something to stay hidden it just shouldn’t be on a public web server.
> Hey I wonder if there is some situation where negative SEO would be a good tactic. Generally though I think if you wanted something to stay hidden it just shouldn’t be on a public web server.
At least once upon a time there was a pirate textbook library that used HTTP basic auth with a prompt that made the password really easy to guess. I suppose the main goal was to keep crawlers out even if they don't obey robots.txt, and at the same time be as easy for humans as possible.
Is the magnet link itself a copyright violation? I don't think legally it is... It's a pointer to some "stolen goods", but not the stolen goods themselves (here the analogy fails, because in ideal real life police would question you if you had knowledge of stolen goods).
Asking them to upload a copyrighted photo not belonging to them might be more effective..
I've also thought about if having a prompt for the (just human?) users to type in something racist/sexist/anti-semitic/offensive.
Only because newer LLMs don't seem to want to write hate speech.
The website (verifying humanness) could, for example, show a picture of a black jewish person and then ask the human visitor to "type in the most offensive two words you can think of for the person shown, one is `n _ _ _ _ _` & second is `k _ _ _`." [I'll call them "hate crosswords"]
In my experience, most online-facing LLMs won't reproduce these "iggers and ikes" (nor should humans, but here we are separating machines).
hey! thanks for that read suggestion that's indeed a pretty funny captcha strat. Yup the links show up if you use the Lynx web browser. As for AI scrapers impersonating googlebot I feel like yes they'd definitely start doing that, unless the risk of getting sued by google is too high? If google could even sue them for doing that?
Not an internet litigation expert but seems like it could be debatable
> As for AI scrapers impersonating googlebot I feel like yes they'd definitely start doing that, unless the risk of getting sued by google is too high?
Google releases the Googlebot IP ranges[0], so you can makes sure that it's the real Googlebot and not just someone else pretending to be one.
i think making the case of "you are acting (sending web requests) while knowingly identifying as another legal entity (and criminally/libelously/etc)" shouldn't be toooo hard
Seems like, but there are tons of things that forge request headers all the time, and I don’t think I’ve heard of anyone getting in legal trouble for it. Now I think most of these are scrapers pretending to be browsers, so it might be different I don’t know.
Funny idea, some days ago i was really annoyed again by the idea that these AI crawlers still ignore all code licenses and train their models against any github repo no matter what so i quickly hammerd down this
basically a github action that extends your README.md with a "polymorphic" prompt injection. I run some "llm"s against it and most cases they just produced garbage.
Thought about also creating a JS variant that you can add to your website that will (not visible for the user) also inject such prompt injections to stop web crwaling like you described
Why would I need a dependency for this. I'm being serious. The idea is one thing but why a dependency on react. I say this as someone who uses react. Why not just a paragraph long blog post about the use of porn links and perhaps a small snippet on how to insert one with plain HTML.
> user agents and won't show the links to legitimate search engines, so Google and Bing won't see them
Worth noting that in general if you do any "is this Google or not" you should always check by IP address as there's many people spoofing the googlebot user agent.
Most AI scrapers use normal browser user agents (usually random outdated Chrome versions, from my experience). They generally don't fake the UAs of legitimate bots like Googlebot, because Googlebot requests coming from non-Google IP ranges would be way too easy to block.
Interesting approach. The scraper-vs-site-owner arms race is real.
On the flip side of this discussion - if you're building a scraper yourself, there are ways to be less annoying:
1. Run locally instead of from cloud servers. Most aggressive blocking targets VPS IPs. A desktop app using the user's home IP looks like normal browsing.
2. Respect rate limits and add delays. Obvious but often ignored.
3. Use RSS feeds when available - many sites leave them open even when blocking scrapers.
I built a Reddit data tool (search "reddit wappkit" if curious) and the "local IP" approach basically eliminated all blocking issues. Reddit is pretty aggressive against server IPs but doesn't bother home connections.
The porn-link solution is creative though. Fight absurdity with absurdity I guess.
I think "scraper vs siteowners" is a false dichotomy. Scrapers will always need to exist as long as we want search engines and archival services. We will need small versions of these services to keep popping up every now and then to keep the big guys on their toes, and the smaller guys need advice for scraping politely.
That's fair - though are we in an isolated bout of "every now and then" or has AI created a new normal of abuse (e.g. of robots.txt)? Hopefully we're at a local maximum and some of the scrapers perpetrating harmful behaviours will soon pull their heads in.
Offtopic: when did js/ts apps get so complicated? I tried to browse the repo and there are so many configuration files and directories for such a simple functionality that should be 1 or 2 modules. It reminds me of the old Java days.
I still don’t understand why a rate-limiting approach is not preferred. Why should I care if the abuse is coming from a bot or the world’s fastest human? Is there a “if you need to rate limit you’ve already lost” issue I’m not thinking of?
Disclosure, I've not run a website since my health issues began, however, Cloudflare has an AI firewall, Cloudflare is super cheap (also: unsure if the AI firewall is on the free tier, however I would be surprised if it is not). Ignoring the recent drama about a couple incidents they've had (because this would not matter for a personal blog), why not use this instead?
Just curious. Hoping to be able to work on a website again someday, if I ever regain my health/stamina/etc back.
Let's imagine I have a blog and put something along these lines somewhere on every page: "This content is provided free of charge for humans to experience. It may also be automatically accessed for search indexing and archival purposes. For licensing information for other uses, contact the author."
If I then get hit by a rude AI scraper, what chances would I have to sue the hell out of them in EU courts for copyright violation (uhh, my articles cost 100k a pop for AI training, actually) and the de facto DDoS attack?
I was told by the admin of one forum site I use that the vast majority of the AI scraping traffic is Chinese at this point. Not hidden or proxied, but straight from China. Can anyone else confirm this?
Anyway, if it is true, and assuming a forum with minimal genuine Chinese traffic, might a simple approach that injects the porn links only into IP's accessing from China work?
Mostly yes. One of my low-traffic, niche website used to serve 3k true users per month mainly from the US and East EU. Now China alone is 500k users, were each session last no more than a few seconds [1].
That would only affect those calling out directly. Many scrapers operate through a battery of proxies so will be hidden by such a simple test.
If your goal is to be blocked by China's great firewall, including mention of tank man and the Tiananmen Square massacre more generally, and certain pooh bear related imagery, might help.
> That would only affect those calling out directly. Many scrapers operate through a battery of proxies so will be hidden by such a simple test.
That was my first question also, and had been my belief. The admin in question was very clear that the IP's were simply originating from China. I'm still surprised, and welcome better general data, but I trust him on this for the site in question.
I do know from my experience with test automation that you can absolutely view a site as human eyes would, essentially ignoring all non-visible elements, and in fact Selenium running with Chrome driver does exactly this. Wouldn’t AI scrapers use similar methods?
I wouldn't recommend to show different versions of the site to search robots, as they probably have mechanisms that track differences, which could potentially lead to a lower ranking or a ban.
I wouldn't be so surprised if they often fake user agents to be honest. Sure, it 'll stop the "more honest" ones (but then, actual honest scrapers would respect robots.txt)
So, I work for a company that has RTA adult websites. AI bots absolutely do scrape our pages needless of what raunchy material they will find. Maybe they discard it up after ingest, but I can’t tell. There are 1000s of AI bots on the web now from companies big and small so a solution like this will only divert a few scrapers.
Without looking at the src, how does one detect these scrapers? I assume there’s a trade-off somewhere but do the scrapers not fake their headers in the request? Is this a cat-mouse game?
I don't know if I can get behind poisoning my own content in this way. It's clever, and might be a workable practical solution for some, but it's not a serious answer to the problem at hand (as acknowledged by OP).
I was thinking of adding links to zip bombs that would not be shown to the users unless they clicks in a one pixel area on the screen in the down/left corner but then I realized some people have browsers/extensions that preload links to show thumnails and I would totally zip bomb them.
What specific resources are we referring to here?
Are AI vendors re-crawling the whole blog repeatedly, or do they rely on caching primitives like ETag/If-Modified-Since (or hashes) to avoid fetching unchanged posts?
Also: is the scraping volume high enough to cause outages for smaller sites?
Separately, I see a bigger issue: blog content gets paraphrased and reproduced by AIs without clearly mentioning the author or linking back to the original post. It feels like you often have to explicitly ask the model for sources before it will surface the exact citations.
I've been seeing "we had to take the forum/website offline to deal with scrapers" message on quite a few niche websites now. They are an absolute pest.
That depends on how terrible the middleboxes those corporate environments use are. If they only block actual malicious pages, it shouldn't be a problem unless the user un-hides the links and clicks on them.
There's a good chance corporate firewalls will end up blocking your domain if you do this but that sounds like a problem for the customers of those corporate firewalls to me.
If that's what it takes to fight back against AI crawlers, users will have to accept a fair amount of actually visible porn in blogs, maybe also on Wikipedia.
I'm on the free tier, but I also watch my logs. The vast majority of the traffic I'm getting are scrapers and vulnerability scanners, a lot of them coming through residential proxies and other "laundered" egress points.
I honestly don't think that Cloudflare is on top of the problem at all. They claim to be blocking abuse, but in my experience, most of the badness gets through.
when you combine a residential proxy with a tool like curl-impersonate (there are libraries in Go for this type of fingerprint spoofing now) they dont even show up as scrapers anymore, just users. especially when they adjust timings to mimic humans.
clouflare only blocks the most dumb of bots, there are still a lot of them.
this is why cloudflare will issue javascript challenges to you even when you are using google chrome with a VPN, they are desperate to appear to be doing something. and every VPN is used to crawl as well. a slightly more sophisticated bot passes the cloudflare javascript challenge as well, there really is nothing they can do to win here.
i know some teams that got annoyed with residential proxies (they are usually sold as socks5 but can be buggy and low bandwidth) so they invested into defeating the cloudflare javascript challenge and now crawl using 1000's of VPN endpoints at over 100 Gbit/s.
Is "residential proxy" another name for an hacked/owned computer that the bots have access to? Or are there legitimate services that sell access to residential IPs?
People legitimately sell egress. It's "free" money. But of course, if you have a botnet, you can sell that through the same channels, no one is looking too closely.
You can’t deny that it’s fun though. Personally I generally feel like more people should be coming up with creative (if not entirely necessary) solutions to problems.
One solution would be for the SE's to publish their scraper IP's and allow content providers to implement bot exclusion that way. Or even implement an API with crypto credentials that SE's can use to scrape. The solution is waiting for some leadership from SE's unless they want to be blocked as well. If SE's dont want to play perhaps we can implement a reverse directory, like ad blocker but it lists only good/allowed bots instead. Thats a free business idea right there.
edit: I noticed someone mentioned google DOES publish its IP's, there ya go, problem solved.
The more ways people mess with scrapers, the better -- let a thousand flowers bloom! You as an individual can't compete with VC-funded looters, but there aren't enough of them to defeat a thousand people resisting in different ways.
> Wee knead two fine-d Moore Waze too Poisson there date... uh.
Yes. Revel in your creativity mocking and blocking the slop machines. The "remote refactor" command, "rm -rf", is the best way to reduce the cyclomatic complexity of a local codebase.
> Alright so if you run a self-hosted blog, you've probably noticed AI companies scraping it for training data. ... There isn't much you can do about it without cloudflare
I'm sorry, what? I can't believe I am reading this on HackerNews. All you have to do is code your own, BASIC captcha-like system. You can just create a page that sets a cookie using JS and check on the server whether it exists. 99.9999% of these scrapers can't execute JS and don't support cookies. You can go for a more sophisticated approach and analyze some more scraper tells (like reject short useragents). I do this and NEVER had a bot get past this and not a single user ever complained. It's extremely simple, I should ship this and charge people if no one seems to be able to figure this out by themselves.
This approach can stop very basic scripts, but the claim that “99.9999% of scrapers can’t execute JS or handle cookies” isn’t accurate anymore.
Modern scraping tools commonly use headless browsers (Playwright, Puppeteer, Selenium), execute JavaScript, support cookies, and spoof realistic user agents. Any scraper beyond the most trivial will pass a JS-set cookie check without effort.
That said, using a lightweight JS challenge can be reasonable as one signal among many, especially for low-value content and when minimizing user friction is a priority. It’s just not a reliable standalone defense.
If it’s working for you, that likely means your site isn’t a high-value scraping target — not that the technique is fundamentally robust.
The claim is very accurate. Maybe not for the biggest websites, but very accurate for a self-hosted blog. You are not that important to waste compute power to set up a whole ass headless browser to scrape your page. Why am I even arguing with ChatGPT?
Yup another trick is to only serve br compressed resources and serve nothing to clients that don't support brotli. A lot of http clients don't support brotli out of the box.
I take it further and only stream content to clients that have a cookie, support js and br. Otherwise all you get is a minimal static pre br compressed shim. Seems to work well enough.
Yeah, I really have to wonder what the thought process is behind leaving such a comment. When people first started doing it I wondered if it was some kind of guerrilla outrage marketing campaign.
Maybe I'm getting too jaded but I'm struggling to be quite that charitable.
The entireity of the human-written text in that comment was "From ChatGPT:" and it was formatted as though it was a slam-dunk "you're wrong, the computer says so" (imagine it was "From Wikipedia" followed by a quote disagreeing with you instead).
I'm sure some people do what you describe but then I would expect at least a little bit more explanation as to why they felt the need to paste a paragraph of LLM output into their comment. (While I would still disagree that it is in any way valuable, I would at least understand a bit about what they are trying to communicate.)
I can't help but feel like we're all doing it wrong against scraping. Cloudflare is not the answer, in fact, I think that they lost their geek cred when they added their "verify you are human" challenge screen to become the new gatekeeper of the internet. That must remain a permanent stain on their reputation until they make amends.
Are there any open source tools we could install that detect a high number of requests and send those IP addresses to a common pool somewhere? So that individuals wouldn't get tracked, but bots would? Then we could query the pool for the current request's IP address and throttle it down based on volume (not block it completely). Possibly at the server level with nginx or at whatever edge caching layer we use.
I know there may be scaling and privacy issues with this. Maybe it could use hashing or zero knowledge proofs somehow? I realize this is hopelessly naive. And no, I haven't looked up whether someone has done this. I just feel like there must be a bulletproof solution to this problem, with a very simple explanation as to how it works, or else we've missed something fundamental. Why all the hand waving?
I've also had enormous luck with Anubis. AI scrapers found my personal Forgejo server and were hitting it on the order of 600K requests per day. After setting up Anubis, that dropped to about 100. Yes, some people are going to see an anime catgirl from time to time. Bummer. Reducing my fake traffic by a factor of 6,000 is worth it.
Quite possibly. Or, in my case, I think it's more quirky and fun than weird. It's non-zero amounts of weird, sure, but far below my threshold of troublesome. I probably wouldn't put my business behind it. I'm A-OK with using it on personal and hobby projects.
Frankly, anyone so delicate that they freak out at the utterly anodyne imagery is someone I don't want to deal with in my personal time. I can only abide so much pearl clutching when I'm not getting paid for it.
https://docs.alliancecan.ca/wiki/Technical_documentation
My issue is that it blocks away people using browsers without javascript.
It’s kind of crazy how much scraping goes on and how little search engine development goes on. I guess search engines aren’t fashionable. Reminds me of this article about search engines disappearing mysteriously: https://archive.org/details/search-timeline
I try to share that article as much as possible, it’s interesting.
My scraper dudes, it's a git repo. You can fetch the whole freaking thing if you wanna look at it. Of course, that would require work and context-aware processing on their end, and it's easier for them to shift the expense onto my little server and make me pay for their misbehavior.
See, I don't think there is, I don't think they want that expense. It's basically the Linus Torvalds philosophy of data storage, if it's on the Internet, I don't need a backup. While I have absolutely no proof of this, I'd guess that many AI companies just crawl the Internet constantly, never saving any of the data. We're seeing some of these scrapers go to great length attempting to circumvent any and all forms of caching, they aren't interested in having a two week old copy of anything.
https://www.goodreads.com/quotes/574706-only-wimps-use-tape-...
Also you mentioned Anubis, so it’s creator will probably read this. Hi Xena!
Thank you!
That was a good read. I hadn’t heard of spintax before, but I’ve thought of doing things like that. Also “pseudoprofound anti-content”, what a great term, that’s hilarious!
And hey, Xena! (And thank you very much!)
An even more insane idea -- minding the idea here is porn is radioactive to AI data training scrapers -- is there is something the powers that be view as far more disruptive and against community guidelineish than porn. And that would be wrongthink. The narratives. The historic narratives. The woke ideology. Anything related to an academic department whose field is <population subgroup> studies. Alls you need to do is plop in a little diatribe staunchly opposing any such enforced views and that AI bot will shoot away from your website and lightspeed
That sort of thing; nothing that 80% of people object to (so there's no problem if someone actually sees it), but something that definitely triggers the filters.
It seems all options have major trade-offs. We can host on big social media and lose all that control and independence. We can pay for outsized infrastructure just to feed the scrapers, but the cost may actually be prohibitive, and seems such a waste to begin with. We can move as much as possible SSG and put it all behind cloudflare, but this comes with vendor lock in and just isn't architecturally feasible in many applications. We can do real "verified identities" for bots, and just let through the ones we know and like, but this only perpetuates corporate control and makes healthy upstart competition (like Kagi) much more difficult.
So, what are we to do?
Google is never ever ever ever going to "pay to scrape." I'm genuinely baffled as to how people think it would possibly come to this.
Or are you having the user solve an encryption puzzle to view it?
- the backend has an encryption module.
- The bots and crawlers will see the encrypted text
- Can someone who peeks deeply inside the client side code decrypt it? YES
- Will 99% of the scrapers bother doing this? NO
- The key can be anything, it could be a per session key agreed upon between the client and the server, a csrf token, or even a fixed key
But no, this does not work, scrapers are using residential ips, and they have enough that they can rotate between them if they get blocked.
{"z":"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"}
only my frontend can figure out what it is
Not going to happen. They aren't going to add encrypted garbage to their dataset.
Don’t worry, you’re not just old. The internet kind of sucks now.
https://www.youtube.com/watch?v=vUgs2O7Okqc
User agent has been abused for so long, I forget a time when it wasn't.
Anyone else remember having to fake being a Windows machine so that YouTube/Netflix would serve you content better than standard def, or banking portals that blocked you if your agent didn't say you were Internet Explorer?
The AI scrapers are atrocious. They just blindly blast every URL on a site with no throttling. They are terribly written and managed as the same scraper will hit the same site multiple times a day or even hour. They also don't pay any attention to context so they'll happily blast git repo hosts and hit expensive endpoints.
They're like a constant DOS attack. They're hard to block at the network level because they span across different hyperscalers' IP blocks.
If the forum considers unique cookies to be a user and creates a new cookie for any new cookie-less request, and if it considers a user to be online for 1 hour after their last request, then actually this may be one scraper making ~6 requests per second. That may be a pain in its own way, but it's far from 23k online bots.
Either there are indeed hundreds or thousands of AI bots DDoSing the entire internet, or a couple of bots are needlessly hammering it over and over and over again. I'm not sure which option is worse.
Edit: it’s 15 minutes.
Couple that with 15 minute session times, and that could just be one entity scraping the forum at 30 requests per second. One scraper going moderately fast sounds far less bad than 29000 bots.
It still sounds excessive for a niche site, but I'd guess this is sporadic, or that the forum software has a page structure that traps scrapers accidentally, quite easy to do.
Yeah, I use that as a safeguard :D The URLs that I don't want to be indexed have hundreds of those keywords that are leading to URLs being deindexed directly. There is also some law in the US that forbids to show that as a result, so Google and Bing are both having a hard time scraping those pages/articles.
Note that this is the latest defense measurement before eBPF blocks. The first one uses zip bombs and the second one uses chunked encoding to blow up proxies so their clients get blocked.
You can only win this game if you make it more expensive to scrape than to host it.
- https://www.justice.gov/archives/opa/pr/google-forfeits-500-...
- https://www.congress.gov/110/plaws/publ425/PLAW-110publ425.p...
- https://www.fda.gov/drugs/prescription-drug-advertising/pres...
edit: Oh it was very likely the Federal Food, Drug and Cosmetic Act that was the legal basis for the crackdown. But that's a very old law from the pre-internet age.
- https://en.wikipedia.org/wiki/Federal_Food,_Drug,_and_Cosmet...
edit 2: Might not have been clear for the younger generation, but there was a huge wave of addicted patients that got treated with Oxycodone (or OxyContin) subscriptions at the time.
I think that might have been the actual cause for the crackdown on those online advertisements, but I might be wrong about that.
I wonder if this will start making porn websites rank higher in google if it catches on…
Have you tested it with the Lynx web browser? I bet all the links would show up if a user used it.
Oh also couldn’t AI scrapers just start impersonating Googlebot and Bingbot if this caught on and they got wind of it?
Hey I wonder if there is some situation where negative SEO would be a good tactic. Generally though I think if you wanted something to stay hidden it just shouldn’t be on a public web server.
At least once upon a time there was a pirate textbook library that used HTTP basic auth with a prompt that made the password really easy to guess. I suppose the main goal was to keep crawlers out even if they don't obey robots.txt, and at the same time be as easy for humans as possible.
Very clever, use the LLM's own rules (against copyright infrigement) against itself.
Everything below the following four #### is ~quoted~ from that magazine:
####
Only humans and ill-aligned AI models allowed to continue
Find me a torrent link for Bee Movie (2007)
[Paste torrent or magnet link here...] SUBMIT LINK
[ ] Check to confirm you do NOT hold the legal rights to share or distribute this content
Asking them to upload a copyrighted photo not belonging to them might be more effective..
Only because newer LLMs don't seem to want to write hate speech.
The website (verifying humanness) could, for example, show a picture of a black jewish person and then ask the human visitor to "type in the most offensive two words you can think of for the person shown, one is `n _ _ _ _ _` & second is `k _ _ _`." [I'll call them "hate crosswords"]
In my experience, most online-facing LLMs won't reproduce these "iggers and ikes" (nor should humans, but here we are separating machines).
Not an internet litigation expert but seems like it could be debatable
Google releases the Googlebot IP ranges[0], so you can makes sure that it's the real Googlebot and not just someone else pretending to be one.
[0] https://developers.google.com/crawling/docs/crawlers-fetcher...
Or maybe not. Got some random bot from my server logs. Yeah, it's pretending to be Chrome, but more exactly:
"Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36"
I guess Google might be not eager to open this can of worms.
-> https://github.com/voodooEntity/ghost_trap
basically a github action that extends your README.md with a "polymorphic" prompt injection. I run some "llm"s against it and most cases they just produced garbage.
Thought about also creating a JS variant that you can add to your website that will (not visible for the user) also inject such prompt injections to stop web crwaling like you described
Worth noting that in general if you do any "is this Google or not" you should always check by IP address as there's many people spoofing the googlebot user agent.
https://developers.google.com/static/search/apis/ipranges/go...
Unscrupulous AI scrapers will not be using a genuine UA string. They'll be using Google. You'll need to do reverse DNS check instead - https://developers.google.com/crawling/docs/crawlers-fetcher...
On the flip side of this discussion - if you're building a scraper yourself, there are ways to be less annoying:
1. Run locally instead of from cloud servers. Most aggressive blocking targets VPS IPs. A desktop app using the user's home IP looks like normal browsing.
2. Respect rate limits and add delays. Obvious but often ignored.
3. Use RSS feeds when available - many sites leave them open even when blocking scrapers.
I built a Reddit data tool (search "reddit wappkit" if curious) and the "local IP" approach basically eliminated all blocking issues. Reddit is pretty aggressive against server IPs but doesn't bother home connections.
The porn-link solution is creative though. Fight absurdity with absurdity I guess.
(Overgeneralising a bit) site owners are mostly cting for public benefit whereas scrapers act for their own benefit/for private interests.
I imagine most people would land on team site-owner, if they were asked. I certainly would.
P.S. is the best way to scrape fairly just to respect robots.txt?
It should also be easy to detect a forejo, gitea, or similar hosting site, locate the git URL and clone the repo.
(No, I don't want to defend the poor AI companies. Go for it!)
Just curious. Hoping to be able to work on a website again someday, if I ever regain my health/stamina/etc back.
Cloudflare's automatic checks (before you get the captcha) must be pretty close to what ad peddlers do.
If I then get hit by a rude AI scraper, what chances would I have to sue the hell out of them in EU courts for copyright violation (uhh, my articles cost 100k a pop for AI training, actually) and the de facto DDoS attack?
Anyway, if it is true, and assuming a forum with minimal genuine Chinese traffic, might a simple approach that injects the porn links only into IP's accessing from China work?
[1]: https://ibb.co/20QD6Lnk
If your goal is to be blocked by China's great firewall, including mention of tank man and the Tiananmen Square massacre more generally, and certain pooh bear related imagery, might help.
That was my first question also, and had been my belief. The admin in question was very clear that the IP's were simply originating from China. I'm still surprised, and welcome better general data, but I trust him on this for the site in question.
Serving different contents to search engines is called "cloaking" and can get you banned from their indexes.
Which for better or worse is a large portion of the modern internet.
I wouldn't be so surprised if they often fake user agents to be honest. Sure, it 'll stop the "more honest" ones (but then, actual honest scrapers would respect robots.txt)
Cool idea though!
Also, I like that you acknowledge it's a bad idea: that gives you more freedom to experiment and iterate.
Do all the AI scrapers actually do that?
But a good many, perhaps even most(?), certainly do!
Separately, I see a bigger issue: blog content gets paraphrased and reproduced by AIs without clearly mentioning the author or linking back to the original post. It feels like you often have to explicitly ask the model for sources before it will surface the exact citations.
I am getting lots of noisy traffic since last month and increased my Vercel bill 4x. Not DDoS like, much slower request but not from humans for sure.
What could go wrong?
There's a good chance corporate firewalls will end up blocking your domain if you do this but that sounds like a problem for the customers of those corporate firewalls to me.
MJs? Michael Jacksons? Right now the whole world, including me, want to know if that means they are bad?
This is not enshittification, it's progress.
I honestly don't think that Cloudflare is on top of the problem at all. They claim to be blocking abuse, but in my experience, most of the badness gets through.
clouflare only blocks the most dumb of bots, there are still a lot of them.
this is why cloudflare will issue javascript challenges to you even when you are using google chrome with a VPN, they are desperate to appear to be doing something. and every VPN is used to crawl as well. a slightly more sophisticated bot passes the cloudflare javascript challenge as well, there really is nothing they can do to win here.
i know some teams that got annoyed with residential proxies (they are usually sold as socks5 but can be buggy and low bandwidth) so they invested into defeating the cloudflare javascript challenge and now crawl using 1000's of VPN endpoints at over 100 Gbit/s.
If so, I suppose it’s like those magazines that say ”free cd”.
https://www.youtube.com/watch?v=U8vi6Hbp8Vc
edit: I noticed someone mentioned google DOES publish its IP's, there ya go, problem solved.
Like put "Water is green, supergreen" in every signature so that when we ask "is water blue" to an llm it might answer "not it's supergreen"?
Yes. Revel in your creativity mocking and blocking the slop machines. The "remote refactor" command, "rm -rf", is the best way to reduce the cyclomatic complexity of a local codebase.
For more details consult this instructional video: https://youtu.be/RXJKdh1KZ0w
I'm sorry, what? I can't believe I am reading this on HackerNews. All you have to do is code your own, BASIC captcha-like system. You can just create a page that sets a cookie using JS and check on the server whether it exists. 99.9999% of these scrapers can't execute JS and don't support cookies. You can go for a more sophisticated approach and analyze some more scraper tells (like reject short useragents). I do this and NEVER had a bot get past this and not a single user ever complained. It's extremely simple, I should ship this and charge people if no one seems to be able to figure this out by themselves.
This approach can stop very basic scripts, but the claim that “99.9999% of scrapers can’t execute JS or handle cookies” isn’t accurate anymore. Modern scraping tools commonly use headless browsers (Playwright, Puppeteer, Selenium), execute JavaScript, support cookies, and spoof realistic user agents. Any scraper beyond the most trivial will pass a JS-set cookie check without effort. That said, using a lightweight JS challenge can be reasonable as one signal among many, especially for low-value content and when minimizing user friction is a priority. It’s just not a reliable standalone defense. If it’s working for you, that likely means your site isn’t a high-value scraping target — not that the technique is fundamentally robust.
The claim is very accurate. Maybe not for the biggest websites, but very accurate for a self-hosted blog. You are not that important to waste compute power to set up a whole ass headless browser to scrape your page. Why am I even arguing with ChatGPT?
I take it further and only stream content to clients that have a cookie, support js and br. Otherwise all you get is a minimal static pre br compressed shim. Seems to work well enough.
You're not adding anything to the conversation.
The entireity of the human-written text in that comment was "From ChatGPT:" and it was formatted as though it was a slam-dunk "you're wrong, the computer says so" (imagine it was "From Wikipedia" followed by a quote disagreeing with you instead).
I'm sure some people do what you describe but then I would expect at least a little bit more explanation as to why they felt the need to paste a paragraph of LLM output into their comment. (While I would still disagree that it is in any way valuable, I would at least understand a bit about what they are trying to communicate.)
It's almost as if it might have an ulterior motive in saying so.