Gregory's ServerReading all these articles about the high traffic from AI bots, here is my experience from a small German hoster: It's a huge problem, and it is getting worse very fast. There are bots like claudebot from Antrophic which are so aggressive that they block whole systems, outranking actual attacks for example against WordPress. And they change their AWS cloud IPs so often that blocking doesn't help for long and they also ignore robots.txt. Our on-call is regularly alerted for such issues.
 90 comments
  @obviousdwest @morihofi @tsia_ there are a lot of providers offering such services and I would prefer to fix the source of the peoblem and not the symptoms. @leah @obviousdwest @tsia_ I just discovered in my Matomo analytics that some traffic came from clients with a referer to https://www.xtraffic.plus. Some sort of traffic generator for SEO optimisation (How even should that work) @morihofi @leah @obviousdwest (knowing some of those SEO people I wouldn’t be surprised if it didn’t work at all)   Maybe: couldn't you reduce the number of hits per minute and block the IP for a certain period of time? In WP, for example, you can do this with Wordfence @werawelt @morihofi @leah nah, typical visitors do bursts (page, CSS, fonts, images, js), so you’d hurt them too much .oO(maybe one could just block all “cloud” services, since they don’t list their customers’ actual ranges in WHOIS and change them too often anyway… if enough people did this, maybe we could get good IP subnet to customer mapping in WHOIS…) @morihofi @leah >50% of my traffic has been non-human for a long time, though that has included search engine spiders etc. I have a target of keeping them to < 50%. See top line in table. https://www.earth.org.uk/note-on-site-technicals.html#Stats The AI bots are a huge slice added to that. In my opinion we should strongly regulate the AI stuff and just forbid most of it, like it's forbidden to run a very dirty coal plant too. For everything else we do it a little like in the pharmaceutical sector. You have to prove that your shiny new product is better/a _real_ benefit compared to the old one. I would bet that only very few very specialized use cases would be the result. But this would require an evaluation that is not based on capitalist logic and this just won't happen.  @leah It ought to be possible to sue people who disregard robots.txt under some kind of computer-related crime law.  @leah I wonder if there wouldn’t be a market for a DDoS protection-like service that detects heavy traffic from a single IP and starts returning garbage data. Ignore robots.txt with your ML scraper, get poisoned data in your ML scraper. Seems like a fair outcome to me. @leah I work for a large software firm with huge amounts of online content - which the AI firms have publicly said they trained on. We've since indicated through our various mechanisms they are not allowed to scrape content... But they very clearly are. We just moved to g4 analytics and started removing bot traffic from our metrics. It was nearly 70% of our traffic, so your numbers are dead on. @leah And here's the thing, they're scraping and surfacing content that's our IP. They're literally stealing from us and redirecting traffic from our business.This is definitely criminal conduct, even if we don't have laws that mark it as such. @T1gerlilly @leah I had a similar experience with my Wordpress site hosted in Canada and my static site hosted in Austria @leah  @leah At minimum, it should follow the European model that says do what thou wilt until there's significant evidence of harm. We have all that and more with AI (also fossil fuels).  @leah i deployed config snippets available for inclusion blocking OpenAI, Claude and some others by useragent as they really misbehave. Pasting random stuff into search boxes etc. Just a single line in the vhost and the get http code for Removed. @rappet that's different depending on the case and bot. See traffic as number of requests if it helps. @leah I have added several lines of code to my .htaccess-file and also adjusted the firewall on the server where my website is hosted. That's it. @leah  @martinus @leah Did the same: All info in French here: https://www.didiermary.fr/bloquer-ai-bots-chatgpt-openai/ @martinus @leah I disagree. If you tell a remote entity, eg by registered mail to the CEO and board or similar, that they are forbidden from accessing your server at all for any reason ("withdrawing implied rights of access") then in the UK at least all continued access is unauthorised and illegal. It is an approach that I have used a few times to keep out persistently badly behaving identifiable bots and spiders. @EarthOrgUK @leah I do not approach any CEO, I just use the method described and guess what? My webserver statistics show me that scraping is, more or less, done. There is still some scraping, but those are either bots from individuals that are not mentioned in the .htaccess file or in the firewall, or brand new bots that I wasn't aware of. @leah I see a future where many sites go password-protected with a simple signup process that includes agreeing that you are not a bot. @ZippyWonderdust @leah If they choose to ignore robots.txt, what makes you think they’ll care any more about a pinkie promise about not being a bot? @ZippyWonderdust @leah Bots have a higher accuracy solving captchas than humans. Captchas are effectively a great way to ensure the only people you’re locking out are _actual people_.  @leah The 21st century sucks. @leah I’m gonna make all my websites blocked until he user actively clicks into the site @leah yes Bytedance (TikTok) is pounding really hard on my clients sites. Right now I'm blocking by user-agent on Cloudflare WAF.  @leah yeah, for the small projects I host and do ops for I see an increasing “ouroboros”-effect: “AI” crawlers hammer us on the one hand for crawling and then we get SEO spammers post “AI” generated spam at the same time. 😞 @leah I know - as a hoster, it doesn't help you much, but maybe you can recommend https://perishablepress.com/8g-firewall/ to your customers. It works great for me :) @leah Would it be possible to deliver very bad content (that still makes sense), instead of the actual content of your website, when the IP belongs to a bot? The training set will be dirty, the resulting model will be catastrophic, and you might be blacklisted by AI-tech companies. (it's not a true solution – but long term, actually it is. The training set is only as good as the trust you can put in it, and if they can't trust the content, what can they do with it?) @leah @ohne_sonne I very much like the idea of somehow Rickrolling all of the AI crawlers!   @leah I think the only solution is poisoning the AIs training data. You have to present decoy material to up the risk that they will train their AI with bad information. It makes it much more difficult for them to maintain the integrity of their service if they have to be constantly retraining bad material out of their AIs.  @coreysnipes @leah A good way to generate fake data that is difficult to discern as fake is to use a Markov random word generator. It will produce sentences that seem real but are nonsense. @coreysnipes @leah I think the beauty of using something like a Markov generator is that it will be very difficult for a DNN to generalize the Markov chain, thus taking capacity away from useful data when being incorporated into the DNN. It's like memorizing the phone book but not realizing it's just random numbers. @leah I’ve put my site behind a free Cloudflare account with their web application firewall, and began to “challenge” all traffic from major hosting companies (AWS, Hetzner, Google, Microsoft, and more over time as I kept monitoring my logs). Took some fine tuning to whitelist legit services of course. Bandwidth usage has dropped by two thirds now… Crawlers such as “DataForSEOBot” or “FriendlyCrawler” and others fly pretty much under everyone’s radar still. It’s a huge mess. @alexskunz @leah some ignore Crawl-Delay in robots.txt, too… grr… I 429 those with Because, fuck Cloudflare. @mirabilos I know that CF has a mixed reputation. From a small website owner's perspective, it's a great help to curb this madness. (there's a lot more going on than just ignoring crawl delays. 3rd parties that identify themselves as GoogleBot etc. for example are super easy to block with CF — hard to do that with .htaccess) @alexskunz @leah this sounds like good advice and also helps to explain why I see that challenge a lot more these days (Comcast and cell ips) @iolaire yes, it's a bit unfortunate. I try to not "challenge" dial-up/residential/cell but some combinations (outdated browser etc.) might trigger a challenge. Cloudflare provides a "challenge solve rate" (CSR) and it's a little bit high today for me — 1.2% of presently 751 challenged requests were solved. Eight slightly inconvenienced visitors is better than getting scraped to death, with resource warnings from my hosting company, etc.   @leah Yup. I run some small sites within a university and they're similarly a shockingly large problem. On one site I've resorted to blocking all of AWS Singapore, because for whatever inscrutable reason a ByteDance bot is hitting every single page every 2-5 minutes, 24/7, every request apparently in a full headless browser (so loading all media and such), and every request from a different IP. This is, mind you, a website that sees maybe a dozen total content updates per year -- in a busy year.  I don't bother with robots.txt nowadays. I run nginx and I just have fun with the bots: @mookie@mookiesplace.com @leah@chaos.social I also mess with bots on my site, ai crawlers get to see a very inaccurate news site.    @leah There are some numbers in Cloudflare's writeup here: https://blog.cloudflare.com/declaring-your-aindependence-block-ai-bots-scrapers-and-crawlers-with-a-single-click @leah Thank you for pointing this out. To increase its relevance we would need some actual data to support your statements. So I guess that’s a case for investigative journalism and science. Anyone here who can do the job? 😎 @leah ouch :( The following Apache .htaccess fragment seems to have worked well for a phpBB forum I administer: == (of course it is even better when they don't make it to Apache at all, but I got the impression that they're detecting they're not getting useful responses anymore and then stop trying)  @leah I've been supporting a small company with their website issues - their dev has been complaining about the database server for a long time... turns out ~80% of their traffic was Claude, and the remaining ~20% was Facebook's bot. After blocking these two based on their user agent, the CPU usage on the system has gone from "pinned at 100%" to basically idle, and response times have fallen back into the acceptable range (previously there were regular spikes to many minutes) |
I would go so far to say that for small sites 30-50% of traffic is now artificial bot traffic. 20-75% for bigger sites with much content. It's a huge waste of resources they outsource to all of us in terms of money and environmental damage.
And this is part of the much bigger footprint all this AI shit has beside the irresponsible big amounts of resources they waste officially.