There's something important here in that a public good like Metabrainz would be fine with the AI bots picking up their content -- they're just doing it in a frustratingly inefficient way.
It's a co-ordination problem: Metabrainz assumes good intent from bots, and has to lock down when they violate that trust. The bots have a different model -- they assume that the website is adversarially "hiding" its content. They won't believe a random site when it says "Look, stop hitting our API, you can pick all of this data in one go, over in this gzipped tar file."
Or better still, this torrent file, where the bots would briefly end up improving the shareability of the data.
> they assume that the website is adversarially "hiding" its content. They won't believe a random site when it says "Look, stop hitting our API, you can pick all of this data in one go, over in this gzipped tar file."
I'm not sure why you're personifying what is almost certainly a script that fetches documents, parses all the links in them, and then recursively fetches all of those.
When we say "AI scraper" we're describing a crawler controlled by an AI company indiscriminately crawling the web, not a literal AI reading and reasoning about each page... I'm surprised this needs to be said.
Yeah AI scrapers is one of the reasons why i have closed my public website https://tvnfo.com and only left donors site online. It’s not only because of AI scrapers but i grew tired of people trying to scrape the site eating a lot of reasorcers this small project don’t have. Very sad really it was publicly online since 2016. Now it’s only available for donors. Running a tiny project on just $60 a month. If this was not my hobby i would close it completely long time ago :-) Who know if there is more support in the future i might reopen public site again with something like anubes bot protection. But i thought it was only small sites like mine who gets hit hard, looks like many have similar issues. Soon nothing will be open or useful online. I wonder if this was the plan all along whoever pushing AI on massive scale.
> They won't believe a random site when it says "Look, stop hitting our API, you can pick all of this data in one go, over in this gzipped tar file."
What mechanism does a site have for doing that? I don't see anything in robots.txt standard about being able to set priority but I could be missing something.
The only real mechanism is "Disallow: /rendered/pages/*" and "Allow: /archive/today.gz" or whatever and there is no communication that the latter is the former. There is no machine-standard AFAIK that allows webmasters to communicate to bot operators in this detail. It would be pretty cool if standard CMSes had such a protocol to adhere to. Install a plugin and people could 'crawl' your Wordpress from a single dump or your Mediawiki from a single dump.
Well, to be fair, I did say "is not read beyond the code itself", header is not the code, so retry-after is a perfectly valid answer. I vaguely remember reading about it, but I don't recall seeing it used in practice. MDN link shows that Chrome derivatives support that header though, which makes it pretty darn widespread
I'm not entirely sure why people think more standards are the way forward. The scrapers apparently don't listen to the already-established standards. What makes one think they would suddenly start if we add another one or two?
There is no standard, well-known way for a website to advertise, "hey, here's a cached data dump for bulk download, please use that instead of bulk scraping". If they were, I'd expect the major AI companies and other users[0] to use that method for gathering training data[1]. They have compelling reasons to: it's cheaper for them, and cultivates goodwill instead of burning it.
This also means that right now, it could be much easier to push through such standard than ever before: there are big players who would actually be receptive to it, so even few not-entirely-selfish actors agreeing on it might just do the trick.
--
[0] - Plenty of them exist. Scrapping wasn't popularized by AI companies, it's standard practice of on-line business in competitive markets. It's the digital equivalent of sending your employees to competing stores undercover.
[1] - Not to be confused with having an LLM scrap specific page for some user because the user requested it. That IMO is a totally legitimate and unfairly penalized/villified use case, because LLM is acting for the user - i.e. it becomes a literal user agent, in the same sense that web browser is (this is the meaning behind the name of "User-Agent" header).
You do realize that these AI scrapers are most likely written by people who have no idea what they're doing right? Or they just don't care? If they were, pretty much none of the problems these things have caused would exist. Even if we did standardize such a thing, I doubt they would follow it. After all, they think they and everyone else has infinite resources so they can just hammer websites forever.
I realise you are making assertions for which you have no evidence. Until a standard exists we can't just assume nobody will use it, particularly when it makes the very task they are scraping for simpler and more efficient.
> They won't believe a random site when it says "Look, stop hitting our API, you can pick all of this data in one go, over in this gzipped tar file."
Is there a mechanism to indicate this? The "a" command in the Scorpion crawling policy file is meant for this purpose, but that is not for use with WWW. (The Scorpion crawling policy file also has several other commands that would be helpful, but also are not for use with WWW.)
There is also the consideration to know what interval they will be archived that can be downloaded in this way; for data that changes often, you will not do it every time. This consideration is also applicable for torrents, since a new hash will be needed for a new version of the file.
The SQLite team faced a similar problem last year, and Richard Hipp (the creator of SQLite) made almost the same comment:
"The malefactor behind this attack could just clone the whole SQLite source repository and search all the content on his own machine, at his leisure. But no: Being evil, the culprit feels compelled to ruin it for everyone else. This is why you don't get to keep nice things...."
This criticism is uninteresting since it's attempting to anthropomorphize a badly coded but well-funded web crawler on a mission to crawl the entire web, over and over again.
There are people behind the web crawler. If they’re so well funded they can exert a little effort to not so badly inconvenience people as they steal their training data.
I am with you in that this rhetoric is getting exhausting.
In this particular case though I don't think "evil” is a moral claim, more shorthand for cost externalizing behavior. Hammering expensive dynamic endpoints with millions of unique requests isn’t neutral automation, it's degrading a shared public resource. Call it evil, antisocial, or extractive, the outcome is the same.
Sounds like you have zero empathy for the real costs AI is driving and feelings that this creates for website owners. How about you pony up and pay for your scraping?
AI is destroying the free internet along with everything else
My web host suspended my website account last week due to a sudden large volume of requests to it - effectively punishing me for being scraped by bots.
I've had to move to a new host to get back up, but what hope does the little guy have? it's like GPU and ram prices, it doesn't matter if I pay 10x 100x or 1000x more than I did, the AI companies have infinite resources, and they don't care what damage they do in the rush to become the no 1 in the industry
The cynic in me would say it's intentional, destroy all the free sites so you have to get your info from their ai models, price home users out of high end hardware so they have to lease the functions from big companies
Wow, so to prevent AI scrapers from harvesting my data I need to send all of my traffic through a third party company that gets to decide who gets to view my content. Great idea!
I use iocaine[0] to generate a tarpit. Yesterday it served ~278k "pages" consisting of ~500MB of gibberish (and that's despite banning most AI scrapers in robots.txt.)
Unfortunately and you kind of have to count this as the cost of the Internet. You've wasted 500Mb of bandwidth.
I've had colocation for eight years+. My monthly b/w cost is now around 20-30Gb a month given to scrapers where I was only be using 1-2Gb a month, years prior.
I pay for premium bandwidth (it's a thing) and only get 2TB of usable data. Do I go offline or let it continue?
Unfortunately, Cloudflare often destroys the experience for users with shared connections, VPNs, exotic browsers… I had to remove it from my site after too many complaints.
Cloudflare works fine with public relay - they and Fastly provide infrastructure for that service (one half of the blinded pair) so it’s definitely something they test.
the more time passes the more i'm convinced that the solution is to—somehow—force everyone to have to go through something like common crawl
i don't want people's servers to be pegged at 100% because a stupid dfs scraper is exhaustively traversing their search facets, but i also want the web to remain scrapable by ordinary people, or rather go back to how readily scrapable it used to be before the invention of cloudflare
as a middle ground, perhaps we could agree on a new /.well-known/ path meant to contain links to timestamped data dumps?
That's sorta what MetaBrainz did - they offer their whole DB as a single tarball dump, much like what Wikipedia does. I downloaded it in the order of an hour; if I need a MusicBrainz lookup, I just do a local query.
For this strategy to work, people need to actually use the DB dumps instead of just defaulting to scraping. Unfortunately scraping is trivially easy, particularly now that AI code assistants can write a working scraper in ~5-10 minutes.
I'll propose my pie in the sky plan here again. We should overhaul the copyright system completely in light of AI and make it mostly win-win for everyone. This is predicated on the idea that the NIST numbers set is sort of the "hello world" dataset for people wanting to learn machine vision and having that common data set is really handy. Numbers made up off the top of my head/subject to tuning but the basic idea is this:
1) Cut copyright to 15-20 years by default. You can have 1 extension of an additional 10-15 years if you submit your work to the "National Data Set" within say 2-3 years of the initial publication.
2) Content in the National set is well categorized and cleaned up. It's the cleanest data set anyone could want. The data set is used both to train some public models and also licensed out to people wanting to train their own models. Both the public models and the data sets are licensed for nominal fees.
3) People who use the public models or data sets as part of their AI system are granted immunity from copyright violation claims for content generated by these models, modulo some exceptions for knowing and intentional violations (e.g. generating the contents of a book into an epub). People who choose to scrape their own data are subject to the current state of the law with regards to both scraping and use (so you probably better be buying a lot of books).
4) The license fees generated from licensing the data and the models would be split into royalty payments to people whose works are in the dataset, and are still under copyright protection, proportional to the amount of data submitted and inversely proportional to the age of that data. There would be some absolute caps in place to prevent slamming the national data sets with junk data just to pump the numbers.
Everyone gets something out of this. AI folks get clean data, that they didn't have to burn a lot of resources scraping. Copyright holders get paid for their works used by AI and retain most of the protections they have today, just for a shorter time), the public gets usable AI tooling without everyone spending their own resources on building their own data sets, site owners and the like get reduced bot/scraping traffic. It's not perfect, and I'm sure the devil is in the details, but that's the nature of this sort of thing.
This alone will kill off all chances of that ever passing.
Like, I fully agree with your proposal... but I don't think it's feasible. There are a lot of media IPs/franchises that are very, very old but still generate insane amounts of money to this day with active developments. Star Wars and Star Trek obviously, but also stuff like the MCU or Avatar is on its best way to two decades of runtime, Iron Man 1 was released in 2008, or Harry Potter which is almost 30 years old. That's dozens of billions of dollars in cumulative income, and most of that is owned by Disney.
Look what it took to finally get even the earliest Disney movies to enter the public domain, and that was stuff from before World War 2 that was so bitterly fought over.
In order to reform copyright... we first have to use anti-trust to break up the large media conglomerates. And it's not just Disney either. Warner, Sony, Comcast and Paramount also hold ridiculous amounts of IP, Amazon entered the fray as well with acquiring MGM (mostly famous for James Bond), and Lionsgate holds the rights for a bunch of smaller but still well-known IPs (Twilight, Hunger Games).
And that's just the movie stuff. Music is just as bad, although at least there thanks to radio stations being a thing, there are licensing agreements and established traditions for remixes, covers, tribute bands and other forms of IP re-use by third parties.
If someone wants to scrape. I mean not levels of complete internet similar to how google does but at a niche level (like you got a forum you wish to scrape)
I like to create tampermonkey scripts regarding these. They are like more lightweight/easier way to build extensions mostly imo
Now I don't like AI but I don't know anything about scraping so I used AI to generate the scraping code and paste it in tampermonkey and let it run
I recently used this for where I effectively scraped a website which had list of vps servers and their prices and I built myself a list of that to analyze as an example
Also I have to say this that I usually try to look out for databases so much so that on a similar website like this related to something, I contacted them about db but no response, their db of server prices were private and only showed lowest
So I picked the other website and did this. I also scraped all headlines of lowendtalk ever with their links for semi purposes of archival and semi purposes of scraping the headlines and parsing it to LLM to find a list of vps providers as well
Seriously, I can't help but think this has to be part of the answer.
Just something like /llms.txt which contains a list of .txt or .txt.gz files or something?
Because the problem is that every site is going to have its own data dump format, often in complex XML or SQL or something.
LLM's don't need any of that metadata, and many sites might not want to provide it because e.g. Yelp doesn't want competitors scraping its list of restaurants.
But if it's intentionally limited to only paragraph-style text, and stripped entirely of URL's, ID's, addresses, phone numbers, etc. -- so e.g. a Yelp page would literally just be the cuisine category and reviews of each restaurant, no name, no city, no identifier or anything -- then it gives LLM's what they need much faster, the site doesn't need to be hammered, and it's not in a format for competitors to easily copy your content.
At most, maybe add markup for <item></item> to represent pages, products, restaurants, whatever the "main noun" is, and recursive <subitem></subitem> to represent e.g. reviews on a restaurant, comments on a review, comments one level deeper on a comment, etc. Maybe a couple more like <title> and <author>, but otherwise just pure text. As simple as possible.
The biggest problem is that a lot of sites will create a "dummy" llms.txt without most of the content because they don't care, so the scrapers will scrape anyways...
It's a monetary one, specifically, large pools of sequestered wealth making extremely bad long term investments all in a single dubious technical area.
Any new phenomenon driven by this process will have the same deleterious results on the rest of computing. There is a market value in ruining your website that's too high for the fruit grabbers to ignore.
In time adaptations will arise. The apparently desired technical future is not inevitable.
Good idea and perhaps a standard that means we only have to grab deltas or some sort of etag based give me all the database dumps after the one I have (or if something has changed).
Bummer. I have used them a lot when I was ripping my cds. Anonymity is a massive value of the web (at least the appearance of anonymity). I wonder if there is a way to have a central anonymous system that just relays trust, not identity.
So maybe something like you can get a token but its trust is very nearly zero until you combine it with other tokens. Combining tokens combines their trust and their consequences. If one token is abused that abuse reflects on the whole token chain. The connection can be revoked for a token but trust takes time to rebuild so it would take a time for their token trust value to go up. Sort of the 'word of mouth' effect but in electronic form. 'I vouch for 2345asdf334t324sda. That's a great user agent!'
A bit (a lot) elaborate but maybe there is a beginning of an idea there, maybe. Definitely I don't want to loose anonymity (or the perception there of) for services like musicbrainz but at the same point they need some mechanism that gives them trust and right now I just don't know of a good one that doesn't have identity attached.
Its not just AI scrappers doing it by themselves but now users are also being trained to put the link in the claude chat/chat gpt and ask it to summarise it. And off course that would show up on the website end as a scraper.
In fact firefox now allows you to preview the link and get key points without ever going to the link[1]
This is generated on device with llama.cpp compiled to webassembly (aka wllama) and running SmolLM2-360M. [1] How is this different from the user clicking on the link? In the end, your local firefox will fetch the link in order to summarize it, the same way you would have followed the link and read through the document in reader mode.
Like, can we all take a step back and marvel that freaking wasm can do things that 10 years ago were firmly in the realm of sci-fi?
I hope they’ll extend that sort of thing to help filter out the parts of the dom that represent attention grabbing stuff that isn’t quite an ad, but is still off topic/not useful for what I’m working on at the moment (and still keep the relevant links).
They should advertise that. I pretty much reflexively avoid any mention of AI in interfaces because they usually mean "we're sending this all to openthropigoogosoft so I hope you don't have any secrets."
Users are not being trained. Despite the seemingly dominant HN belief to the contrary, people use LLMs for interacting with information (on the web or otherwise) because they work. SOTA LLM services are just that good.
- AI shops scraping the web to update their datasets without respecting netiquette (or sometimes being unable to automate it for every site due to the scale, ironically).
- People extensively using agents (search, summarizers, autonomous agents etc), which are indistinguishable from scraper bots from website's perspective.
- Agents being both faster and less efficient (more requests per action) than humans.
At this point, it's pretty clear that the AI scrapers won't be limited by any voluntary restrictions. Bytedance never seemed to live with robots.txt limitations, and I think at least some of the others didn't either.
- Humans tip humans as a lottery ticket for an experience (meet the creator) or sweepstakes (free stuff)
- Agents tip humans because they know they'll need original online content in the long-term to keep improving.
For the latter, frontier labs will need to fund their training/inference agents with a tipping jar.
There's no guarantee, but I can see it happening given where things are movin.
Someone convinced me last time[0] that these aren't the well-known scrapers we know but other actors. We wouldn't be able to tell, certainly. I'd like to help the scrapers be better about reading my site, but I get why they aren't.
I wish there were an established protocol for this. Say a $site/.well-known/machine-readable.json that instructs you on a handful of established software or allows pointing to an appropriate dump. I would gladly provide that for LLMs.
Of course this doesn't solve for the use-case where the AI companies are trying to train their models on how to navigate real world sites so I understand it doesn't solve all problems, but one of the things I think I'd like in the future is to have my own personal archive of the web as I know it (Internet Archive is too slow to browse and has very tight rate-limits) and I was surprised by how little protocol support there is for robots.
robots.txt is pretty sparse. You can disallow bots and this and that, but what I want to say is "you can get all this data from this git repo" or "here's a dump instead with how to recreate it". Essentially, cooperating with robots is currently under-specified. I understand why: almost all bots have no incentive to cooperate so webmasters do not attempt to. But it would be cool to be able to inform the robots appropriately.
To archive Metabrainz there is no way but to browse the pages slowly page-by-page. There's no machine-communicable way that suggests an alternative.
Well, imagine the best case and that you're a cooperative bot writer who does not intend to harm website owners. Okay, so you follow robots.txt and all that. That's straightforward.
But it's not like you're writing a "metabrainz crawler" and a "metafilter crawler" and a "wiki.roshangeorge.dev crawler". You're presumably trying to write a general Internet crawler. You encounter a site that is clearly a HTTP view into some git repo (say). How do you know to just `git clone` the repo in order to have the data archived as opposed to just browsing the HTTP view.
As you can see, I've got a lot of crawlers on my blog as well, but it's a mediawiki instance. I'd gladly host a Mediawiki dump for them to take, but then they'd have to know this was a Mediawiki-based site. How do I tell them that? The humans running the program don't know my site exists. Their bot just browses the universe and finds links and does things.
In the Metabrainz case, it's not like the crawler writer knows Metabrainz even exists. It's probably just linked somewhere in the web the crawler is exploring. There's no "if Metabrainz, do this" anywhere in there.
The robots.txt is a bit of a blunt-force instrument, and friendly bot writers should follow it. But assuming they do, there's no way for them to know that "inefficient path A to data" is the same as "efficient path B to data" if both are visible to their bot unless they write a YourSite-specific crawler.
What I want is to have a way to say "the canonical URL for the data on A is at URL B; you can save us both trouble by just fetching B". In practice, none of this is a problem for me. I cache requests at Cloudflare, and I have Mediawiki caching generated pages, so I can easily weather the bot traffic. But I want to enable good bot writers to save their own resources. It's not reasonable for me to expect them to write a me-crawler, but if there is a format to specify the rules I'm happy to be compliant.
> To archive Metabrainz there is no way but to browse the pages slowly page-by-page. There's no machine-communicable way that suggests an alternative.
Why does there have to be a "machine-communicable way"? If these developers cared about such things they would spend 20 seconds looking at this page. It's literally one of the first links when you Google "metabrainz"
You expect the developers of a crawler to look at every site they crawl and develop a specialized crawler for them? That’s fine if you’re only crawling a handful of sites, but absolutely insane if you’re crawling the entire web.
I feel the pain — it’s very difficult to detect many of the less ethical scrapers. They use residential IP pools, rotate IPs, and provide valid user agents.
And many proxy/scraping providers now use real browsers that can automatically bypass cloudflare captchas as well, and a bot with a real browser similarly won't be clicking on any invisible links, so... I am skeptical just how long this will make an appreciable difference.
I wish this wasn't necessary, but the next steps forward are likely:
a) Have a reverse proxy that keeps a "request budget" per IP and per net block, but instead of blocking requests, causing the client to rotate their IP, the requests get throttled/slowed down, without dropping them.
b) Write your API servers in more efficient languages. According to their Github, their backend runs on Perl and Python. These technologies have been "good enough" for quite some time, but considering current circumstances and until a better solution is found, this may not be the case anymore and performance and cpu cost per request does matter these days.
c) Optimize your database queries, remove as much code as possible from your unauthenticated GET request handlers, require authentication for the expensive ones.
There isn't. There never was one, because vast majority of websites are actually selfish with respect to data, even when that's entirely pointless. You can see this even here, with how some people complain LLMs made them stop writing their blogs: turns out plenty of people say they write for others to read, but they care more about tracking and controlling the audience.
Anyway, all that means there was never a critical mass of sites large enough for a default bulk data dump discovery to become established. This means even the most well-intentioned scrappers cannot reliably determine if such mechanism exist, and have to scrap per-page anyway.
> turns out plenty of people say they write for others to read
LLMs are not people. They don't write blogs so that a company can profit from their writing by training LLMs on it. They write for others to read their ideas.
LLMs aren't making their owners money by just idling on datacenters worth of GPU. They're making money by being useful for users that pay for access. The knowledge and insights from writings that go into training data all end up being read by people directly, as well as inform even more useful output and work benefiting even more people.
I don’t know why anyone would still be trying to pull data off the open internet. Too much signal to noise. So much AI influence already baked into the corpus. You are just going to be reinforcing existing bias. I’m more worried about the day Amazon or Hugging Face take down their large data sets.
I wish more resources were available legitimately. There is a dataset I need for legitimate research that I cannot even find a way to contact the repo owners.
Mind you I take effort to not be burdensome by downloading only what I need and taking time between each request of a couple seconds, and the total data usage is low.
Ironically, I supposed you could call it "AI" what I'm using it for, but really it's just data analytics.
I haven't really looked but I wonder if there are any IP reputation services tracking AI crawlers the same way they track tor relays and vpns and such. If those databases were accurate it seems like you could prevent those crawlers from ever hitting the site. Maybe they change too often/too quickly.
A year or two ago I personally encountered scraping bots that were scraping every possible resultant page from a given starting point. So if it scraped a search results page it would also scrape every single distinct combination of facets on that search (including nonsensical combinations e.g. products that match the filter "products where weight<2lbs AND weight>2lbs")
We ended up having to block entire ASNs and several subnets (lots from Facebook IPs, interestingly)
If you have a lot of pages, AI bots will scrape every single one on a loop - wiki's generally don't have anywhere near the number of pages as an incremented entity primary id. I have a few million pages on a tiny website and it gets hammered by AI bots all day long. I can handle it, but it's a nuisance and they're basically just scraping garbage (statistics pages of historical matches or user pages that have essentially no content).
Many of them don't even self-identify and end up scraping with shrouded user-agents or via bot-farms. I've had to block entire ASNs just to tone it down. It also hurts good-faith actors who genuinely want to build on top of our APIs because I have to block some cloud providers.
I would guess that I'm getting anywhere from 10-25 AI bot requests (maybe more) per real user request - and at scale that ends up being quite a lot. I route bot traffic to separate pods just so it doesn't hinder my real users' experience[0]. Keep in mind that they're hitting deeply cold links so caching doesn't do a whole lot here.
[0] this was more of a fun experiment than anything explicitly necessary, but it's proven useful in ways I didn't anticipate
How many requests per second do you get? I also see a lot of bot traffic but nowhere near to hit the servers significantly, and i render most stuff on the server directly.
It's not "written too slow" if you e.g. only get 50 users a week, though. If bots add so much load that you need to go optimise your website for them, then that's a bot problem not a website problem.
Yes yes, definitely people don’t know what they’re doing and not that they’re operating on a scale or problem you are not. Metabrainz cannot cache all of these links as most of them are hardly ever hit. Try to assume good intent.
There’s a lot of factors. Depends how well your content lends itself to being cached by a CDN, the tech you (or your predecessors) chose to build it with, and how many unique pages you have. Even with pretty aggressive caching, having a couple million pages indexed adds up real fast. Especially if you weren’t fortunate enough to inherit a project using a framework that makes server side rendering easy.
Require some special header for accessing them, without needing a API token if it is public data. HTTPS will not necessarily be required. Scrapers can still use it but it seems unlikely unless it becomes common enough; but if they do then you can remove that and require proper authentication.
Another is to use X.509 client certificates for authentication, which is more secure than using API keys anyways; however, this will require that you have a X.509 certificate, and some people might not want that, so due to that, perhaps it should not be mandatory.
Grateful for Metabrainz putting in this work to keep the service up. We really ought to have some kind of "I am an AI!" signal to route the request properly into a long queue...
Nowadays people complain about AI scrapers with the same vain as they complained about search indexers a way back when. Just a few years later, people had stopped caring too much about storage access and bandwidth, and started begging search engines to visit their websites. Every trick on the planet Earth, SEO optimization, etc.
Looking forward to the time when everybody suddenly starts to embrace AI indexers and welcome them. History does not repeat itself but it rhymes.
This is an interesting approach. Archive.org could be such a solution, kind of. Not its cold storage as it's now, but a warm access layer. Sponsorship by AI companies would a good initiative for the project.
Search indexing historically has had several of orders less impact on bandwidth and processing costs to website maintainers.
My recommendation is to copy the text in this article and pass it LLM to summarize this article's key points, since it appears you missed the central complaint of the article.
It's much easier to detect a single account abusing your API and ban them/require payment. Trying to police an endpoint open to the internet is like playing g whackamole
Some sort of hashing and incremental serial versioning type standards with http servers would allow hitting a server up for incremental udpates, allow clean access to content, with rate limits, and even keep up with live feeds and chats and so on.
Something like this in practice breaks a lot of the adtech surveillance and telemetry, and makes use of local storage, and incidentally empowers p2p sharing with a verifiable source of truth, which incentivizes things like IPFS and big p2p networks.
The biggest reason we don't already have this is the exploitation of user data for monetization and intrusive modeling.
It's easy to build proof of concept instances of things like that and there are other technologies that make use of it, but we'd need widespread adoption and implementation across the web. It solves the coordination problem, allows for useful throttling to shut out bad traffic while still enabling public and open access to content.
The technical side to this is already done. Merkle trees and hashing and crypto verification are solid, proven tech with standard implementations and documentation, implementing the features into most web servers would be simple, and it would reduce load on infrastructure by a huge amount. It'd also result in IPFS and offsite sharing and distributed content - blazing fast, efficient, local focused browsers.
It would force opt in telemetry and adtech surveillance, but would also increase the difference in appearance between human/user traffic and automated bots and scrapers.
We can't have nice things because the powers that be decided that adtech money was worth far more than efficiency, interoperability, and things like user privacy and autonomy.
I actually deleted my web site early 2025 and removed the A record from DNS because of AI scraper traffic. It had been up for 22 years. Nothing important or particularly useful on it but it's an indicator of the times.
Bear in mind that some portion of this could be human directed research. I'm doing a research project right now with 1000 things that I'm building profiles on; to build a full profile requires an agent to do somewhere around 100 different site lookups. Where APIs exist, I've registered API keys and had the agent write a script to query the data in that manner, but that required me to be deliberate about it. Non technie plebs aren't likely to be directed to use an API by the agent.
Many years ago AWS came up with a "Requester Pays" model for their S3 storage, where you can make a request for someone else's object using your own account and it would charge the transfer cost to your own account instead of theirs.
I wonder if a model similar to this (but decentralized/federated or something) could be used to help fight bots?
At some point they must become more cost efficient by pure market economics mechanisms. That implies less load on sites. Much of the scraping that I see is still very dumb/repetative. Like Googlebot in like 2001.
(Blocking Chinese IP ranges with the help of some geoip db helps a lot in the short term. Azure as a whole is the second largest source of pure idiocy.)
> The /metadata/lookup API endpoints (GET and POST versions) now require the caller to send an Authorization token in order for this endpoint to work.
> The ListenBrainz Labs API endpoints for mbid-mapping, mbid-mapping-release and mbid-mapping-explain have been removed. Those were always intended for debugging purposes and will also soon be replaced with a new endpoints for our upcoming improved mapper.
> LB Radio will now require users to be logged in to use it (and API endpoint users will need to send the Authorization header). The error message for logged in users is a bit clunky at the moment; we’ll fix this once we’ve finished the work for this year’s Year in Music.
Seems reasonable and no big deal at all. I'm not entirely sure what "nice things" we can't have because of this. Unauthenticated APIs?
random idea, instead of blocking scrapers, maybe detect them (via user-agent, request patterns, ignoring robots.txt) and serve them garbage data wrapped in dad jokes.
if (isSuspiciousScraper(req)) {
return res.json({
data: getDadJoke(),
artist: "Rick Astley", // always
album: "Never Gonna Give You Up"
});
}
This is frustrating to watch. MetaBrainz is exactly the kind of project AI companies should be supporting—open data, community-maintained, freely available for download in bulk.
Instead they’re:
∙ Ignoring robots.txt (the bare minimum web courtesy)
∙ Bypassing the provided bulk download (literally designed for this use case)
∙ Scraping page-by-page (inefficient for everyone)
∙ Overloading volunteer-run infrastructure
∙ Forcing the project to add auth barriers that hurt legitimate users
The irony: if they’d just contacted MetaBrainz and said “hey, we’d like to use your dataset for training,” they’d probably get a bulk export and maybe even attribution. Instead, they’re burning goodwill and forcing open projects to lock down.
This pattern is repeating everywhere. Small/medium open data projects can’t afford the infrastructure to handle aggressive scraping, so they either:
1. Add authentication (reduces openness)
2. Rate limit aggressively (hurts legitimate API users)
3. Go offline entirely (community loses the resource)
AI companies are externalizing their data acquisition costs onto volunteer projects. It’s a tragedy of the commons, except the “commons” is deliberately maintained infrastructure that these companies could easily afford to support.
Have you considered publishing a list of the offending user agents / IP ranges? Might help other projects protect themselves, and public shaming sometimes works when technical measures don’t
Metabrainz is a great resource -- I wrote about them a few years ago here: https://www.eff.org/deeplinks/2021/06/organizing-public-inte...
There's something important here in that a public good like Metabrainz would be fine with the AI bots picking up their content -- they're just doing it in a frustratingly inefficient way.
It's a co-ordination problem: Metabrainz assumes good intent from bots, and has to lock down when they violate that trust. The bots have a different model -- they assume that the website is adversarially "hiding" its content. They won't believe a random site when it says "Look, stop hitting our API, you can pick all of this data in one go, over in this gzipped tar file."
Or better still, this torrent file, where the bots would briefly end up improving the shareability of the data.
> they assume that the website is adversarially "hiding" its content. They won't believe a random site when it says "Look, stop hitting our API, you can pick all of this data in one go, over in this gzipped tar file."
I'm not sure why you're personifying what is almost certainly a script that fetches documents, parses all the links in them, and then recursively fetches all of those.
When we say "AI scraper" we're describing a crawler controlled by an AI company indiscriminately crawling the web, not a literal AI reading and reasoning about each page... I'm surprised this needs to be said.
It doesn’t need to be said.
Yeah AI scrapers is one of the reasons why i have closed my public website https://tvnfo.com and only left donors site online. It’s not only because of AI scrapers but i grew tired of people trying to scrape the site eating a lot of reasorcers this small project don’t have. Very sad really it was publicly online since 2016. Now it’s only available for donors. Running a tiny project on just $60 a month. If this was not my hobby i would close it completely long time ago :-) Who know if there is more support in the future i might reopen public site again with something like anubes bot protection. But i thought it was only small sites like mine who gets hit hard, looks like many have similar issues. Soon nothing will be open or useful online. I wonder if this was the plan all along whoever pushing AI on massive scale.
> They won't believe a random site when it says "Look, stop hitting our API, you can pick all of this data in one go, over in this gzipped tar file."
What mechanism does a site have for doing that? I don't see anything in robots.txt standard about being able to set priority but I could be missing something.
The only real mechanism is "Disallow: /rendered/pages/*" and "Allow: /archive/today.gz" or whatever and there is no communication that the latter is the former. There is no machine-standard AFAIK that allows webmasters to communicate to bot operators in this detail. It would be pretty cool if standard CMSes had such a protocol to adhere to. Install a plugin and people could 'crawl' your Wordpress from a single dump or your Mediawiki from a single dump.
A sitemap.xml file could get you most of the way there.
It’s not great, but you could add it to the body of a 429 response.
Genuinely curious: do programs read bodies of 429 responses? In the code bases that I have seen, 429 is not read beyond the code itself
Sometimes! The server can also send a retry-after header to indicate when the client is allowed to request the resource again: https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/...
… which isn't part of the body of a 429…
Well, to be fair, I did say "is not read beyond the code itself", header is not the code, so retry-after is a perfectly valid answer. I vaguely remember reading about it, but I don't recall seeing it used in practice. MDN link shows that Chrome derivatives support that header though, which makes it pretty darn widespread
Almost certainly not by default, certainly not in any of the http libs I have used
The mechanism is putting some text that points to the downloads.
So perhaps it's time to standardize that.
I'm not entirely sure why people think more standards are the way forward. The scrapers apparently don't listen to the already-established standards. What makes one think they would suddenly start if we add another one or two?
There is no standard, well-known way for a website to advertise, "hey, here's a cached data dump for bulk download, please use that instead of bulk scraping". If they were, I'd expect the major AI companies and other users[0] to use that method for gathering training data[1]. They have compelling reasons to: it's cheaper for them, and cultivates goodwill instead of burning it.
This also means that right now, it could be much easier to push through such standard than ever before: there are big players who would actually be receptive to it, so even few not-entirely-selfish actors agreeing on it might just do the trick.
--
[0] - Plenty of them exist. Scrapping wasn't popularized by AI companies, it's standard practice of on-line business in competitive markets. It's the digital equivalent of sending your employees to competing stores undercover.
[1] - Not to be confused with having an LLM scrap specific page for some user because the user requested it. That IMO is a totally legitimate and unfairly penalized/villified use case, because LLM is acting for the user - i.e. it becomes a literal user agent, in the same sense that web browser is (this is the meaning behind the name of "User-Agent" header).
You do realize that these AI scrapers are most likely written by people who have no idea what they're doing right? Or they just don't care? If they were, pretty much none of the problems these things have caused would exist. Even if we did standardize such a thing, I doubt they would follow it. After all, they think they and everyone else has infinite resources so they can just hammer websites forever.
I realise you are making assertions for which you have no evidence. Until a standard exists we can't just assume nobody will use it, particularly when it makes the very task they are scraping for simpler and more efficient.
I'm in favor of /.well-known/[ai|llm].txt or even a JSON or (gasp!) XML.
Or even /.well-known/ai/$PLATFORM.ext which would have the instructions.
Could even be "bootstrapped" from /robots.txt
> They won't believe a random site when it says "Look, stop hitting our API, you can pick all of this data in one go, over in this gzipped tar file."
Is there a mechanism to indicate this? The "a" command in the Scorpion crawling policy file is meant for this purpose, but that is not for use with WWW. (The Scorpion crawling policy file also has several other commands that would be helpful, but also are not for use with WWW.)
There is also the consideration to know what interval they will be archived that can be downloaded in this way; for data that changes often, you will not do it every time. This consideration is also applicable for torrents, since a new hash will be needed for a new version of the file.
> The bots have a different model -- they assume that the website is adversarially "hiding" its content.
this should give us pause. if a bot considers this adversarial and is refusing to respect the site owners wishes, thats a big part of the problem.
a bot should not consider that “adversarial”
> Or better still, this torrent file, where the bots would briefly end up improving the shareability of the data.
that is an amazing thought.
The SQLite team faced a similar problem last year, and Richard Hipp (the creator of SQLite) made almost the same comment:
"The malefactor behind this attack could just clone the whole SQLite source repository and search all the content on his own machine, at his leisure. But no: Being evil, the culprit feels compelled to ruin it for everyone else. This is why you don't get to keep nice things...."
https://sqlite.org/forum/forumpost/7d3eb059f81ff694
This criticism is uninteresting since it's attempting to anthropomorphize a badly coded but well-funded web crawler on a mission to crawl the entire web, over and over again.
"Why don't you just clone the repo?" Seriously?
"Why don't you just clone the repo?" Yes. Why dont you?
If you're gonna grab a repo to make a code theft machine then at least dont ddos the servers while you're at it.
Do you at all understand the concept of global web crawl?
If, not, maybe spend a minute thinking about what it might entail?
This is not about defending the enemy, it's about understanding them.
"The malefactor behind this attack" isn't a complaint about the web crawler.
There are people behind the web crawler. If they’re so well funded they can exert a little effort to not so badly inconvenience people as they steal their training data.
I find this rhetoric so tiring.
Evil? Really? Seems a bit dramatic.
I am with you in that this rhetoric is getting exhausting.
In this particular case though I don't think "evil” is a moral claim, more shorthand for cost externalizing behavior. Hammering expensive dynamic endpoints with millions of unique requests isn’t neutral automation, it's degrading a shared public resource. Call it evil, antisocial, or extractive, the outcome is the same.
> shorthand for cost externalizing behavior
I consider that evil, having no regard for the wellbeing of others for you own greed.
OK. How about shitty and selfish then?
What other word would you use? I find "evil" quite an apt description.
You can be ignorantly evil.
Absolutely evil.
Sounds like you have zero empathy for the real costs AI is driving and feelings that this creates for website owners. How about you pony up and pay for your scraping?
When they routinely do things like take down public libraries yes I consider it evil too.
AI is destroying the free internet along with everything else
My web host suspended my website account last week due to a sudden large volume of requests to it - effectively punishing me for being scraped by bots.
I've had to move to a new host to get back up, but what hope does the little guy have? it's like GPU and ram prices, it doesn't matter if I pay 10x 100x or 1000x more than I did, the AI companies have infinite resources, and they don't care what damage they do in the rush to become the no 1 in the industry
The cynic in me would say it's intentional, destroy all the free sites so you have to get your info from their ai models, price home users out of high end hardware so they have to lease the functions from big companies
Cloudflare has a service for this now that will detect AI scrapers and send them to a tarpit of infinite AI generated nonsense pages.
Wow, so to prevent AI scrapers from harvesting my data I need to send all of my traffic through a third party company that gets to decide who gets to view my content. Great idea!
You don’t need to do anything. You can use any number of solutions or roll your own.
Someone shared an alternative. Must everything in AI threads be so negative and condescending?
You can implement this yourself, who is stopping you?
Citation needed
I use iocaine[0] to generate a tarpit. Yesterday it served ~278k "pages" consisting of ~500MB of gibberish (and that's despite banning most AI scrapers in robots.txt.)
[0] https://iocaine.madhouse-project.org
Can't seem to access this.
It flashes some text briefly then gives me an 418 TEAPOT response. I wonder if it's because I'm on Linux?
EDIT: Begrudgingly checked Chrome, and it loads. I guess it doesn't like Firefox?
Nor Safari on iOS.
Doesn't work on my firefox either.
Friendly fire, I suppose.
Works on my Firefox. Mac and Linux
Unfortunately and you kind of have to count this as the cost of the Internet. You've wasted 500Mb of bandwidth.
I've had colocation for eight years+. My monthly b/w cost is now around 20-30Gb a month given to scrapers where I was only be using 1-2Gb a month, years prior.
I pay for premium bandwidth (it's a thing) and only get 2TB of usable data. Do I go offline or let it continue?
One of the most popular ones is Anubis. It uses a proof of work and can even do poisoning: https://anubis.techaro.lol/
They even mention iocaine. I know, inconceivable!: https://iocaine.madhouse-project.org/
There's also tons of HN posts on the topic with varying solutions:
https://news.ycombinator.com/item?id=45935729
https://news.ycombinator.com/item?id=45711094
https://news.ycombinator.com/item?id=44142761
https://news.ycombinator.com/item?id=44378127
https://forge.hackers.town/hackers.town/nepenthes
> Citation needed
this reply kinda sucks :)
Unfortunately, Cloudflare often destroys the experience for users with shared connections, VPNs, exotic browsers… I had to remove it from my site after too many complaints.
I am sure Cloudflare would have no problem selling you a VPN service.
After all, it's not very far from hosting booters and selling DoS protection.
Well... https://developers.cloudflare.com/warp-client/warp-modes/#wa...
Price is $5/mo
Also iCloud Private Relay.
CloudFlare is making it impossible to browse privately
Cloudflare works fine with public relay - they and Fastly provide infrastructure for that service (one half of the blinded pair) so it’s definitely something they test.
Not sure "TLS added and removed here :)" as a Service is the right tool in the drawer for this.
Poison Fountain: https://news.ycombinator.com/item?id=46577464
cloudflare also blocks my human-is-driving browser all the time
"enahble javascript and cookies to continue"
also unsupported browser
The solution, as always, is noise.
Except for the scrapers that pay cloudflare to exempt them.
Do you have a link to that?
https://blog.cloudflare.com/ai-labyrinth/
Which is still a far worse experience than if Cloudflare's services weren't needed.
Modern scrapers are using headless chromium which will not see the invisible links, so I'm not sure how long this will be effective.
the more time passes the more i'm convinced that the solution is to—somehow—force everyone to have to go through something like common crawl
i don't want people's servers to be pegged at 100% because a stupid dfs scraper is exhaustively traversing their search facets, but i also want the web to remain scrapable by ordinary people, or rather go back to how readily scrapable it used to be before the invention of cloudflare
as a middle ground, perhaps we could agree on a new /.well-known/ path meant to contain links to timestamped data dumps?
That's sorta what MetaBrainz did - they offer their whole DB as a single tarball dump, much like what Wikipedia does. I downloaded it in the order of an hour; if I need a MusicBrainz lookup, I just do a local query.
For this strategy to work, people need to actually use the DB dumps instead of just defaulting to scraping. Unfortunately scraping is trivially easy, particularly now that AI code assistants can write a working scraper in ~5-10 minutes.
I mean this AI data scrapper would need to scan and fetch billions of website
why would they even care over 1 single website ??? You expect instiution to care out of billions website they must scrape daily
This is probably the reason. It’s more effort to special case every site that offers dumps than to just unleash your generic scraper on it.
I'll propose my pie in the sky plan here again. We should overhaul the copyright system completely in light of AI and make it mostly win-win for everyone. This is predicated on the idea that the NIST numbers set is sort of the "hello world" dataset for people wanting to learn machine vision and having that common data set is really handy. Numbers made up off the top of my head/subject to tuning but the basic idea is this:
1) Cut copyright to 15-20 years by default. You can have 1 extension of an additional 10-15 years if you submit your work to the "National Data Set" within say 2-3 years of the initial publication.
2) Content in the National set is well categorized and cleaned up. It's the cleanest data set anyone could want. The data set is used both to train some public models and also licensed out to people wanting to train their own models. Both the public models and the data sets are licensed for nominal fees.
3) People who use the public models or data sets as part of their AI system are granted immunity from copyright violation claims for content generated by these models, modulo some exceptions for knowing and intentional violations (e.g. generating the contents of a book into an epub). People who choose to scrape their own data are subject to the current state of the law with regards to both scraping and use (so you probably better be buying a lot of books).
4) The license fees generated from licensing the data and the models would be split into royalty payments to people whose works are in the dataset, and are still under copyright protection, proportional to the amount of data submitted and inversely proportional to the age of that data. There would be some absolute caps in place to prevent slamming the national data sets with junk data just to pump the numbers.
Everyone gets something out of this. AI folks get clean data, that they didn't have to burn a lot of resources scraping. Copyright holders get paid for their works used by AI and retain most of the protections they have today, just for a shorter time), the public gets usable AI tooling without everyone spending their own resources on building their own data sets, site owners and the like get reduced bot/scraping traffic. It's not perfect, and I'm sure the devil is in the details, but that's the nature of this sort of thing.
> Cut copyright to 15-20 years by default.
This alone will kill off all chances of that ever passing.
Like, I fully agree with your proposal... but I don't think it's feasible. There are a lot of media IPs/franchises that are very, very old but still generate insane amounts of money to this day with active developments. Star Wars and Star Trek obviously, but also stuff like the MCU or Avatar is on its best way to two decades of runtime, Iron Man 1 was released in 2008, or Harry Potter which is almost 30 years old. That's dozens of billions of dollars in cumulative income, and most of that is owned by Disney.
Look what it took to finally get even the earliest Disney movies to enter the public domain, and that was stuff from before World War 2 that was so bitterly fought over.
In order to reform copyright... we first have to use anti-trust to break up the large media conglomerates. And it's not just Disney either. Warner, Sony, Comcast and Paramount also hold ridiculous amounts of IP, Amazon entered the fray as well with acquiring MGM (mostly famous for James Bond), and Lionsgate holds the rights for a bunch of smaller but still well-known IPs (Twilight, Hunger Games).
And that's just the movie stuff. Music is just as bad, although at least there thanks to radio stations being a thing, there are licensing agreements and established traditions for remixes, covers, tribute bands and other forms of IP re-use by third parties.
If someone wants to scrape. I mean not levels of complete internet similar to how google does but at a niche level (like you got a forum you wish to scrape)
I like to create tampermonkey scripts regarding these. They are like more lightweight/easier way to build extensions mostly imo
Now I don't like AI but I don't know anything about scraping so I used AI to generate the scraping code and paste it in tampermonkey and let it run
I recently used this for where I effectively scraped a website which had list of vps servers and their prices and I built myself a list of that to analyze as an example
Also I have to say this that I usually try to look out for databases so much so that on a similar website like this related to something, I contacted them about db but no response, their db of server prices were private and only showed lowest
So I picked the other website and did this. I also scraped all headlines of lowendtalk ever with their links for semi purposes of archival and semi purposes of scraping the headlines and parsing it to LLM to find a list of vps providers as well
Seriously, I can't help but think this has to be part of the answer.
Just something like /llms.txt which contains a list of .txt or .txt.gz files or something?
Because the problem is that every site is going to have its own data dump format, often in complex XML or SQL or something.
LLM's don't need any of that metadata, and many sites might not want to provide it because e.g. Yelp doesn't want competitors scraping its list of restaurants.
But if it's intentionally limited to only paragraph-style text, and stripped entirely of URL's, ID's, addresses, phone numbers, etc. -- so e.g. a Yelp page would literally just be the cuisine category and reviews of each restaurant, no name, no city, no identifier or anything -- then it gives LLM's what they need much faster, the site doesn't need to be hammered, and it's not in a format for competitors to easily copy your content.
At most, maybe add markup for <item></item> to represent pages, products, restaurants, whatever the "main noun" is, and recursive <subitem></subitem> to represent e.g. reviews on a restaurant, comments on a review, comments one level deeper on a comment, etc. Maybe a couple more like <title> and <author>, but otherwise just pure text. As simple as possible.
The biggest problem is that a lot of sites will create a "dummy" llms.txt without most of the content because they don't care, so the scrapers will scrape anyways...
It's not a technical problem you are facing.
It's a monetary one, specifically, large pools of sequestered wealth making extremely bad long term investments all in a single dubious technical area.
Any new phenomenon driven by this process will have the same deleterious results on the rest of computing. There is a market value in ruining your website that's too high for the fruit grabbers to ignore.
In time adaptations will arise. The apparently desired technical future is not inevitable.
Good idea and perhaps a standard that means we only have to grab deltas or some sort of etag based give me all the database dumps after the one I have (or if something has changed).
And then YC funds a startup who plans to leapfrog the competition by doing their own scrape instead using the standard data everyone else has
Bummer. I have used them a lot when I was ripping my cds. Anonymity is a massive value of the web (at least the appearance of anonymity). I wonder if there is a way to have a central anonymous system that just relays trust, not identity.
So maybe something like you can get a token but its trust is very nearly zero until you combine it with other tokens. Combining tokens combines their trust and their consequences. If one token is abused that abuse reflects on the whole token chain. The connection can be revoked for a token but trust takes time to rebuild so it would take a time for their token trust value to go up. Sort of the 'word of mouth' effect but in electronic form. 'I vouch for 2345asdf334t324sda. That's a great user agent!'
A bit (a lot) elaborate but maybe there is a beginning of an idea there, maybe. Definitely I don't want to loose anonymity (or the perception there of) for services like musicbrainz but at the same point they need some mechanism that gives them trust and right now I just don't know of a good one that doesn't have identity attached.
Its not just AI scrappers doing it by themselves but now users are also being trained to put the link in the claude chat/chat gpt and ask it to summarise it. And off course that would show up on the website end as a scraper.
In fact firefox now allows you to preview the link and get key points without ever going to the link[1]
[1] https://imgur.com/a/3E17Dts
> In fact firefox now allows you to preview the link and get key points without ever going to the link[1]
> [1] https://imgur.com/a/3E17Dts
This is generated on device with llama.cpp compiled to webassembly (aka wllama) and running SmolLM2-360M. [1] How is this different from the user clicking on the link? In the end, your local firefox will fetch the link in order to summarize it, the same way you would have followed the link and read through the document in reader mode.
[1] https://blog.mozilla.org/en/mozilla/ai/ai-tech/ai-link-previ...
That’s awesome :-)
Like, can we all take a step back and marvel that freaking wasm can do things that 10 years ago were firmly in the realm of sci-fi?
I hope they’ll extend that sort of thing to help filter out the parts of the dom that represent attention grabbing stuff that isn’t quite an ad, but is still off topic/not useful for what I’m working on at the moment (and still keep the relevant links).
I actually didnt know it was using a local model and that it fetches it locally.
They should advertise that. I pretty much reflexively avoid any mention of AI in interfaces because they usually mean "we're sending this all to openthropigoogosoft so I hope you don't have any secrets."
I'll also add that if you go to the Labs page (in settings) you can enable another local model to semantically search your history
Users are not being trained. Despite the seemingly dominant HN belief to the contrary, people use LLMs for interacting with information (on the web or otherwise) because they work. SOTA LLM services are just that good.
It's three issues:
- AI shops scraping the web to update their datasets without respecting netiquette (or sometimes being unable to automate it for every site due to the scale, ironically).
- People extensively using agents (search, summarizers, autonomous agents etc), which are indistinguishable from scraper bots from website's perspective.
- Agents being both faster and less efficient (more requests per action) than humans.
Resurfacing a proposal I put out on llms-txt: https://github.com/AnswerDotAI/llms-txt/issues/88
We should add optional `tips` addresses in llms.txt files.
We're also working on enabling and solving this at Grove.city.
Human <-> Agent <-> Human Tips don't account for all the edge cases, but they're a necessary and happy neutral medium.
Moving fast. Would love to share more with the community.
Wrote about it here: https://x.com/olshansky/status/2008282844624216293
I like the idea, (original) content creators being credited is good for the entire ecosystem.
Though if LLMs are willingly ignoring robots.txt, often hiding themselves or using third party scraped data- are they going to pay?
At this point, it's pretty clear that the AI scrapers won't be limited by any voluntary restrictions. Bytedance never seemed to live with robots.txt limitations, and I think at least some of the others didn't either.
I can't see this working.
The thesis/approach is:
- Humans tip humans as a lottery ticket for an experience (meet the creator) or sweepstakes (free stuff) - Agents tip humans because they know they'll need original online content in the long-term to keep improving.
For the latter, frontier labs will need to fund their training/inference agents with a tipping jar.
There's no guarantee, but I can see it happening given where things are movin.
Someone convinced me last time[0] that these aren't the well-known scrapers we know but other actors. We wouldn't be able to tell, certainly. I'd like to help the scrapers be better about reading my site, but I get why they aren't.
I wish there were an established protocol for this. Say a $site/.well-known/machine-readable.json that instructs you on a handful of established software or allows pointing to an appropriate dump. I would gladly provide that for LLMs.
Of course this doesn't solve for the use-case where the AI companies are trying to train their models on how to navigate real world sites so I understand it doesn't solve all problems, but one of the things I think I'd like in the future is to have my own personal archive of the web as I know it (Internet Archive is too slow to browse and has very tight rate-limits) and I was surprised by how little protocol support there is for robots.
robots.txt is pretty sparse. You can disallow bots and this and that, but what I want to say is "you can get all this data from this git repo" or "here's a dump instead with how to recreate it". Essentially, cooperating with robots is currently under-specified. I understand why: almost all bots have no incentive to cooperate so webmasters do not attempt to. But it would be cool to be able to inform the robots appropriately.
To archive Metabrainz there is no way but to browse the pages slowly page-by-page. There's no machine-communicable way that suggests an alternative.
0: https://news.ycombinator.com/item?id=46352723
As referenced in the article, there absolutely is an alternative.
https://metabrainz.org/datasets
Linked to from the homepage as “datasets”.
I may be too broadly interpreting what you mean by “machine-communicable” in the context of AI scraping though.
Well, imagine the best case and that you're a cooperative bot writer who does not intend to harm website owners. Okay, so you follow robots.txt and all that. That's straightforward.
But it's not like you're writing a "metabrainz crawler" and a "metafilter crawler" and a "wiki.roshangeorge.dev crawler". You're presumably trying to write a general Internet crawler. You encounter a site that is clearly a HTTP view into some git repo (say). How do you know to just `git clone` the repo in order to have the data archived as opposed to just browsing the HTTP view.
As you can see, I've got a lot of crawlers on my blog as well, but it's a mediawiki instance. I'd gladly host a Mediawiki dump for them to take, but then they'd have to know this was a Mediawiki-based site. How do I tell them that? The humans running the program don't know my site exists. Their bot just browses the universe and finds links and does things.
In the Metabrainz case, it's not like the crawler writer knows Metabrainz even exists. It's probably just linked somewhere in the web the crawler is exploring. There's no "if Metabrainz, do this" anywhere in there.
The robots.txt is a bit of a blunt-force instrument, and friendly bot writers should follow it. But assuming they do, there's no way for them to know that "inefficient path A to data" is the same as "efficient path B to data" if both are visible to their bot unless they write a YourSite-specific crawler.
What I want is to have a way to say "the canonical URL for the data on A is at URL B; you can save us both trouble by just fetching B". In practice, none of this is a problem for me. I cache requests at Cloudflare, and I have Mediawiki caching generated pages, so I can easily weather the bot traffic. But I want to enable good bot writers to save their own resources. It's not reasonable for me to expect them to write a me-crawler, but if there is a format to specify the rules I'm happy to be compliant.
> To archive Metabrainz there is no way but to browse the pages slowly page-by-page. There's no machine-communicable way that suggests an alternative.
Why does there have to be a "machine-communicable way"? If these developers cared about such things they would spend 20 seconds looking at this page. It's literally one of the first links when you Google "metabrainz"
https://metabrainz.org/datasets
You expect the developers of a crawler to look at every site they crawl and develop a specialized crawler for them? That’s fine if you’re only crawling a handful of sites, but absolutely insane if you’re crawling the entire web.
I feel the pain — it’s very difficult to detect many of the less ethical scrapers. They use residential IP pools, rotate IPs, and provide valid user agents.
And many proxy/scraping providers now use real browsers that can automatically bypass cloudflare captchas as well, and a bot with a real browser similarly won't be clicking on any invisible links, so... I am skeptical just how long this will make an appreciable difference.
So what you’re saying is that a CAPTCHA cannot actually T C and H A.
I wish this wasn't necessary, but the next steps forward are likely:
a) Have a reverse proxy that keeps a "request budget" per IP and per net block, but instead of blocking requests, causing the client to rotate their IP, the requests get throttled/slowed down, without dropping them.
b) Write your API servers in more efficient languages. According to their Github, their backend runs on Perl and Python. These technologies have been "good enough" for quite some time, but considering current circumstances and until a better solution is found, this may not be the case anymore and performance and cpu cost per request does matter these days.
c) Optimize your database queries, remove as much code as possible from your unauthenticated GET request handlers, require authentication for the expensive ones.
> Rather than downloading our dataset in one complete download, they insist on loading all of MusicBrainz one page at a time.
Is there a standard mechanism for batch-downloading a public site? I'm not too familiar with crawlers these days.
There isn't. There never was one, because vast majority of websites are actually selfish with respect to data, even when that's entirely pointless. You can see this even here, with how some people complain LLMs made them stop writing their blogs: turns out plenty of people say they write for others to read, but they care more about tracking and controlling the audience.
Anyway, all that means there was never a critical mass of sites large enough for a default bulk data dump discovery to become established. This means even the most well-intentioned scrappers cannot reliably determine if such mechanism exist, and have to scrap per-page anyway.
> turns out plenty of people say they write for others to read
LLMs are not people. They don't write blogs so that a company can profit from their writing by training LLMs on it. They write for others to read their ideas.
LLMs aren't making their owners money by just idling on datacenters worth of GPU. They're making money by being useful for users that pay for access. The knowledge and insights from writings that go into training data all end up being read by people directly, as well as inform even more useful output and work benefiting even more people.
They're not talking about downloading the web pages. The data is available in a bulk download: https://listenbrainz.org/data/
I don’t know why anyone would still be trying to pull data off the open internet. Too much signal to noise. So much AI influence already baked into the corpus. You are just going to be reinforcing existing bias. I’m more worried about the day Amazon or Hugging Face take down their large data sets.
MetaBrainz is a fairly valuable “high signal” dataset though.
I wish more resources were available legitimately. There is a dataset I need for legitimate research that I cannot even find a way to contact the repo owners.
Mind you I take effort to not be burdensome by downloading only what I need and taking time between each request of a couple seconds, and the total data usage is low.
Ironically, I supposed you could call it "AI" what I'm using it for, but really it's just data analytics.
I haven't really looked but I wonder if there are any IP reputation services tracking AI crawlers the same way they track tor relays and vpns and such. If those databases were accurate it seems like you could prevent those crawlers from ever hitting the site. Maybe they change too often/too quickly.
How do they get overloaded? Is the website too slow? I have a quite big wiki online and barely see any impact from bots.
A year or two ago I personally encountered scraping bots that were scraping every possible resultant page from a given starting point. So if it scraped a search results page it would also scrape every single distinct combination of facets on that search (including nonsensical combinations e.g. products that match the filter "products where weight<2lbs AND weight>2lbs")
We ended up having to block entire ASNs and several subnets (lots from Facebook IPs, interestingly)
I have encountered this same issue with faceted search results and individual inventory listings.
If you have a lot of pages, AI bots will scrape every single one on a loop - wiki's generally don't have anywhere near the number of pages as an incremented entity primary id. I have a few million pages on a tiny website and it gets hammered by AI bots all day long. I can handle it, but it's a nuisance and they're basically just scraping garbage (statistics pages of historical matches or user pages that have essentially no content).
Many of them don't even self-identify and end up scraping with shrouded user-agents or via bot-farms. I've had to block entire ASNs just to tone it down. It also hurts good-faith actors who genuinely want to build on top of our APIs because I have to block some cloud providers.
I would guess that I'm getting anywhere from 10-25 AI bot requests (maybe more) per real user request - and at scale that ends up being quite a lot. I route bot traffic to separate pods just so it doesn't hinder my real users' experience[0]. Keep in mind that they're hitting deeply cold links so caching doesn't do a whole lot here.
[0] this was more of a fun experiment than anything explicitly necessary, but it's proven useful in ways I didn't anticipate
How many requests per second do you get? I also see a lot of bot traffic but nowhere near to hit the servers significantly, and i render most stuff on the server directly.
In these discussions no one will admit this, but the answer is generally yes. Websites written in python and stuff like that.
It's not "written too slow" if you e.g. only get 50 users a week, though. If bots add so much load that you need to go optimise your website for them, then that's a bot problem not a website problem.
Yes yes, definitely people don’t know what they’re doing and not that they’re operating on a scale or problem you are not. Metabrainz cannot cache all of these links as most of them are hardly ever hit. Try to assume good intent.
But serving HTML is unbelievably cheap, isn't it?
There’s a lot of factors. Depends how well your content lends itself to being cached by a CDN, the tech you (or your predecessors) chose to build it with, and how many unique pages you have. Even with pretty aggressive caching, having a couple million pages indexed adds up real fast. Especially if you weren’t fortunate enough to inherit a project using a framework that makes server side rendering easy.
The API seems to be written in Perl: https://github.com/metabrainz/musicbrainz-server
Time for a vinyl-style Perl revival ...
There some other possibilities, such as:
Require some special header for accessing them, without needing a API token if it is public data. HTTPS will not necessarily be required. Scrapers can still use it but it seems unlikely unless it becomes common enough; but if they do then you can remove that and require proper authentication.
Another is to use X.509 client certificates for authentication, which is more secure than using API keys anyways; however, this will require that you have a X.509 certificate, and some people might not want that, so due to that, perhaps it should not be mandatory.
Grateful for Metabrainz putting in this work to keep the service up. We really ought to have some kind of "I am an AI!" signal to route the request properly into a long queue...
Companies wouldn’t send it because they know that most websites would block them
Nowadays people complain about AI scrapers with the same vain as they complained about search indexers a way back when. Just a few years later, people had stopped caring too much about storage access and bandwidth, and started begging search engines to visit their websites. Every trick on the planet Earth, SEO optimization, etc.
Looking forward to the time when everybody suddenly starts to embrace AI indexers and welcome them. History does not repeat itself but it rhymes.
We already know the solution: One well-behaved, shared scraper could serve all of the AI companies simultaneously.
The problem is that they're not doing it.
This is an interesting approach. Archive.org could be such a solution, kind of. Not its cold storage as it's now, but a warm access layer. Sponsorship by AI companies would a good initiative for the project.
Except robots.txt was the actual real solution to search indexing...
Search indexing historically has had several of orders less impact on bandwidth and processing costs to website maintainers.
My recommendation is to copy the text in this article and pass it LLM to summarize this article's key points, since it appears you missed the central complaint of the article.
Nothing prevents scraper from creating a free account and sending auth token in API requests.
I'm not saying the API changes are pointless, but still, what's the catch?
It's much easier to detect a single account abusing your API and ban them/require payment. Trying to police an endpoint open to the internet is like playing g whackamole
Some sort of hashing and incremental serial versioning type standards with http servers would allow hitting a server up for incremental udpates, allow clean access to content, with rate limits, and even keep up with live feeds and chats and so on.
Something like this in practice breaks a lot of the adtech surveillance and telemetry, and makes use of local storage, and incidentally empowers p2p sharing with a verifiable source of truth, which incentivizes things like IPFS and big p2p networks.
The biggest reason we don't already have this is the exploitation of user data for monetization and intrusive modeling.
It's easy to build proof of concept instances of things like that and there are other technologies that make use of it, but we'd need widespread adoption and implementation across the web. It solves the coordination problem, allows for useful throttling to shut out bad traffic while still enabling public and open access to content.
The technical side to this is already done. Merkle trees and hashing and crypto verification are solid, proven tech with standard implementations and documentation, implementing the features into most web servers would be simple, and it would reduce load on infrastructure by a huge amount. It'd also result in IPFS and offsite sharing and distributed content - blazing fast, efficient, local focused browsers.
It would force opt in telemetry and adtech surveillance, but would also increase the difference in appearance between human/user traffic and automated bots and scrapers.
We can't have nice things because the powers that be decided that adtech money was worth far more than efficiency, interoperability, and things like user privacy and autonomy.
Seems a mistake to disable the (I assume) faster-to-generate api. Bots will go back to scraping the website itself, increasing load.
Setting the API to require a token and adding a honeypot to the pages themselves seems like a decent solution.
I actually deleted my web site early 2025 and removed the A record from DNS because of AI scraper traffic. It had been up for 22 years. Nothing important or particularly useful on it but it's an indicator of the times.
But why?
No humans, no point.
it's funny how ai is the problem that the cryptocurrencyverse was always in search of...
Bear in mind that some portion of this could be human directed research. I'm doing a research project right now with 1000 things that I'm building profiles on; to build a full profile requires an agent to do somewhere around 100 different site lookups. Where APIs exist, I've registered API keys and had the agent write a script to query the data in that manner, but that required me to be deliberate about it. Non technie plebs aren't likely to be directed to use an API by the agent.
Many years ago AWS came up with a "Requester Pays" model for their S3 storage, where you can make a request for someone else's object using your own account and it would charge the transfer cost to your own account instead of theirs.
I wonder if a model similar to this (but decentralized/federated or something) could be used to help fight bots?
At some point they must become more cost efficient by pure market economics mechanisms. That implies less load on sites. Much of the scraping that I see is still very dumb/repetative. Like Googlebot in like 2001.
(Blocking Chinese IP ranges with the help of some geoip db helps a lot in the short term. Azure as a whole is the second largest source of pure idiocy.)
They seem to have so much bubble money at the moment that the cost of scraping is probably a rounding error in their pocket change.
Exactly.
> The /metadata/lookup API endpoints (GET and POST versions) now require the caller to send an Authorization token in order for this endpoint to work.
> The ListenBrainz Labs API endpoints for mbid-mapping, mbid-mapping-release and mbid-mapping-explain have been removed. Those were always intended for debugging purposes and will also soon be replaced with a new endpoints for our upcoming improved mapper.
> LB Radio will now require users to be logged in to use it (and API endpoint users will need to send the Authorization header). The error message for logged in users is a bit clunky at the moment; we’ll fix this once we’ve finished the work for this year’s Year in Music.
Seems reasonable and no big deal at all. I'm not entirely sure what "nice things" we can't have because of this. Unauthenticated APIs?
I agree its not a big deal. Unauthenticated APIs are nice though, especially for someone who's maybe not as familiar with how APIs work.
It is nonsense since AI is the nicest thing.
random idea, instead of blocking scrapers, maybe detect them (via user-agent, request patterns, ignoring robots.txt) and serve them garbage data wrapped in dad jokes.
This is frustrating to watch. MetaBrainz is exactly the kind of project AI companies should be supporting—open data, community-maintained, freely available for download in bulk. Instead they’re: ∙ Ignoring robots.txt (the bare minimum web courtesy) ∙ Bypassing the provided bulk download (literally designed for this use case) ∙ Scraping page-by-page (inefficient for everyone) ∙ Overloading volunteer-run infrastructure ∙ Forcing the project to add auth barriers that hurt legitimate users The irony: if they’d just contacted MetaBrainz and said “hey, we’d like to use your dataset for training,” they’d probably get a bulk export and maybe even attribution. Instead, they’re burning goodwill and forcing open projects to lock down. This pattern is repeating everywhere. Small/medium open data projects can’t afford the infrastructure to handle aggressive scraping, so they either: 1. Add authentication (reduces openness) 2. Rate limit aggressively (hurts legitimate API users) 3. Go offline entirely (community loses the resource) AI companies are externalizing their data acquisition costs onto volunteer projects. It’s a tragedy of the commons, except the “commons” is deliberately maintained infrastructure that these companies could easily afford to support. Have you considered publishing a list of the offending user agents / IP ranges? Might help other projects protect themselves, and public shaming sometimes works when technical measures don’t