4 ms·

As a person who has a couple of sites out there, and witnesses AI crawlers coming and fetching pages from these sites, I have a question: What prevents these c

As a person who has a couple of sites out there, and witnesses AI crawlers coming and fetching pages from these sites, I have a question: What prevents these companies from keeping a copy of that particular page, which I specifically disallowed for bot scraping, and feed it to their next training cycle? Pinky promises? Ethics? Laws? Technical limitations? Leeroy Jenkins?

tempfile 1y ago
The way to prevent people from downloading your pages and using them is to take them off the public internet. There are laws to prevent people from violating your copyright or from preventing access to your service (by excessive traffic). But there is (thankfully) no magical right that stops people from reading your content and describing it.
- bayindirh 1y ago
  Many site operators want people to access their content, but prevent AI companies from scraping their sites for training data. People who think like that made tools like Anubis, and it works. I also want to keep this distinction on the sites I own. I also use licenses to signal that this site is not good to use for AI training, because it's CC BY-NC-SA-2.0. So, I license my content appropriately (No derivative, Non-commercial, shareable with the same license with attribution), add technical countermeasures on top, because companies doesn't respect these licenses (because monies), and circumvent these mechanisms (because monies), and I'm the one to suck this up and shut-up (because their monies)? Makes no sense whatsoever.
  - hombre_fatal 1y ago
    I guess that's a question that might be answered by the NYT vs OpenAI lawsuit at least on the enforceability of copyright claims if you're a corporation like NYT. If you don't have the funds to sue an AI corp, I'd probably think of a plan B. Maybe poison the data for unauthenticated users. Or embrace the inevitability. Or see the bright side of getting embedded in models as if you're leaving your mark.
  - tempfile 1y ago
    Of course some people want that. And at the moment they can prevent it. But those methods may stop working. Will it then be alright to do it? Of course not, so why bother mentioning that they are able to prevent it now - just give a justification. Your license is probably not relevant. I can go to the cinema and watch a movie, then come on this website and describe the whole plot. That isn't copyright infringement. Even if I told it to the whole world, it wouldn't be copyright infringement. Probably the movie seller would prefer it if I didn't tell anyone. Why should I care? I actually agree that AI companies are generally bad and should be stopped - because they use an exorbitant amount of bandwidth and harm the services for other users. At least they should be heavily taxed. I don't even begrudge people for using Anubis, at least in some cases. But it is wrong-headed (and actually wrong in fact) to try to say someone may or may not use my content for some purpose because it hurts my feelings or it messes with my ad revenue. We have laws against copyright infringement, and to prevent service disruption. We should not have laws that say, yes you can read my site but no you can't use it to train an LLM, or to build a search index. That would be unethical. Call for a windfall tax if they piss you off so much.
    - bayindirh 1y ago
      > I can go to the cinema and watch a movie, then come on this website and describe the whole plot. That isn't copyright infringement. This is a false analogy. A correct one would be going to a 1000 movies and creating the 1001th movie with scenes cropped from these 1000 movies and assemble it as a new movie, and this is copyright infringement. I don't think any of the studios would applaud and support you for your creativity. > But it is wrong-headed (and actually wrong in fact) to try to say someone may or may not use my content for some purpose because it hurts my feelings or it messes with my ad revenue. Why does it have to be always about money? Personally it's not. I just don't want my work to be abused and sold to people to benefit a third party without my consent and will (and all my work is licensed appropriately for that). > We should not have laws that say, yes you can read my site but no you can't use it to train an LLM, or to build a search index. This goes both ways. If big corporations can scrape my material without asking me and resell it as an output of a model, I can equally distill their models further and sell it as my own. If companies can scrape my pages to sell my content as theirs, I can scrape theirs and unpaywall them. But that will be copyright infringement, just because they have more money. What angers me is "all is fair game because you're a small fish, and this is a capitalist marketplace" mentality. If companies can paywall their content to humans that don't pay, I can paywall AI companies and demand money or push them out of my lawn, just because I feel like that. The inverse is very unethical, but very capitalist, yes. It's not always about money. P.S.: Oh, try to claim that you can train a model with medical data without any clearance because it'd be unethical to have laws limiting this. It'll be fun. Believe me.
  - zzo38computer 1y ago
    I don't want AI companies to scrape my sites (or use the files I wrote) for training data either, but that is not specifically what I am trying to stop (unless the files are supposed to be private and unpublished). I should not stop them from using the files for what they want, once they have them. (I also specifically do not want to block use of lynx, curl, Dillo, etc.) What I want to stop is excessive crawling and scraping of my server. Once they have the file they can do what they want with it. Another comment (44786237) mentions that robots.txt is only for restricting recursive access; I agree and that is what should be blocked. They also should not access the same file several times quickly even though it should be unnecessary to do so, just as much as they should not access all of the files. (If someone wants to make a mirror of the files, there may be other ways, e.g. in case there is a archive file available to download many at once (possibly, in case if the site operator made their own index and then did it this way). If it is a git repository, then it can be cloned.)
  - account42 1y ago
    > Many site operators want people to access their content, but prevent AI companies from scraping their sites for training data. That is unfortunately not a distinction that is currently legally enforceable. Until that changes all other "solutions" are pointless and only cause more harm. > People who think like that made tools like Anubis, and it works. It works to get real humans like myself to stop visiting your site while scrapers will have people whose entire job is to work around such "protections". Just like traditional DRM inconveniences honest customers and not pirates. And to be clear, what you are advocating for is DRM. > I also want to keep this distinction on the sites I own. I also use licenses to signal that this site is not good to use for AI training, because it's CC BY-NC-SA-2.0. If AI crawlers cared about that we wouldn't be talking about this issue. A license and only give more permissions than there are without one.
    - bayindirh 1y ago
      > It works to get real humans like myself to stop visiting your site If we talk about Anubis, it's pretty invisible. You wait a couple of seconds in the first visit, and don't get challenged for a couple of weeks, at least. With more tuning some of the sites using Anubis work perfectly well without ever seeing Anubis' wall while stopping AI crawlers. > And to be clear, what you are advocating for is DRM. Yes. It's pretty ironic that someone like me who believes in open access prefers a DRM solution to keep companies abusing the small fish, but life is an interesting phenomenon, and these things happen. > Until that changes all other "solutions" are pointless and only cause more harm. As an addendum to above paragraph, I'm not happy that I have to insert draconian measures between the user and the information I want to share, but I need a way to signal that I'm not having their ways to these faceless things. What do you propose? Taking my sites offline? Burning myself in front of one of the HQs? > If AI crawlers cared about that we wouldn't be talking about this issue. A license and only give more permissions than there are without one. AI crawlers default to "Public Domain" when they find no licenses. Some of my lamest source code repositories made into "The Stack" because I forgot to add COPYING.md. A fork of a GPLv2 tool I wrote some patches also got into "The Stack", because COPYING.md was not in the root folder of the repository. I'd rather add licenses (which I can accept) to things rather than leave them as-is, because AI companies also eagerly grab things without license. All licenses I use mandate attribution and continuation of license, at least, and my blog doesn't allow any derivations of from what I have written. So you can't ingest it into a model to be derived and remixed with something else.
accrual 1y ago
Thanks for sharing your experience. A little off-topic but I'd like to start hosting some personal content, guides/tutorials, etc. Do you still see authentic human traffic on your domains, is it easy to discern? I feel like I missed the bus on running a blog pre-AI.
- kldg 1y ago
  if you do analytics, it is not so hard, but then you need to store user data (if not directly, then worse, with a third party), which should be viewed as a liability. I see ~2/3 human traffic, ~1/3 bot traffic (I just parse user agent strings and count whitelisted browsers as human), but my main landing page is all dynamic-populated webgl. I just asked Gemini what it sees on website, and it states "The page appears to be loading, with the text "Loading room data...".[1] There are also labels for "BG", "FG", and "CURSOR", and a background weather animation." -so I can be feel reasonably confident I don't need to worry about AI, for now; it needs a machine-friendly frontend. you could go proper insanomode, too. remaking The Internet is trivial if you don't care about existing web standards -- replacing HTTP with your own TCP implementation, getting off html/js/css, etc. being greenfield, you can control the protocol, server, and client implementation, and put it in whatever language you want. I made a stateful Internet implementation in Python earlier for proof-of-concept, but I want to port it and expand on it in rust soon (just for fun; I don't do serious biznos). you'll very likely have 100% human traffic then, even if you're the only person curious and trusting enough to run your client.
  - 71bw 1y ago
    > I made a stateful Internet implementation in Python earlier for proof-of-concept Is there a repo or some other form of public access? I'd like to see this.
    - kldg 1y ago
      it's not in a shareable state; is unsafe as-is. can share general idea and sample "webpage" files, though. the server ("lodge") passes JSON to the client from what are called .branch files. the client receives JSON, parses it, then builds the UI and state representation from the JSON, then stored in that client's memory (self.current_doc and self.page_state in python client). branches can invoke waterwheel (.ww) files hosted on the lodge. waterwheel files on the lodge contain scripts which define how patches (as JSON) are to be sent to the client. the client updates its state based on the JSON patch it receives. sample .branch and .ww from python implementation (in pastebin so to not make everyone have to scroll through this): https://pastebin.com/A0DEZDmR https://pastebin.com/A0DEZDmR
tintor 1y ago
technical limitations / data poisoning measures
Aeolun 1y ago
> What prevents these companies from keeping a copy of that particular page, which I specifically disallowed for bot scraping, and feed it to their next training cycle? What prevents anyone else? robots.txt is a request, not an access policy.
- utbabya 1y ago
  This honor system mostly worked at scale because interests align, which seems to be no longer the case. Does information no longer wants to be free now? Maybe internet, just like social media was just a social experiment at the end, albeit a successful one. Thanks GenAI.
  - windexh8er 1y ago
    Can the Terms of Service of individual content creators leverage a "death of a thousand cuts" model to produce a legal honeypot which would require organizations like Perplexity to be bound up in 10s of thousands of conciliation court cases? Big Tech has hidden behind ToS for years. Now, it seems as though it only works for them, but not against. It seems as though this would be easy to orchestrate and prove forcing these companies into a legal nightmare or risk insolvent business stature due to the high load of cases filed against. Why couldn't something like this be used to flip the table? A conciliation brigading, of sorts.
    - Eisenstein 1y ago
      Because lawyers are expensive and big tech companies have lots of them. Because it takes a ton of time and effort to sue someone. Because you need to show standing, which means you need to be able to demonstrate you lost something of value by their actions. Because the power imbalance is heavily weighted towards a corporation. Because the way to deal with such things should be legislation and not court decisions. And lots more reasons...
      - windexh8er 1y ago
        That's exactly why I said conciliation court. None of what you've outlined is required nor is it expensive. But, for each case, the defendant is still required to show up. I've successfully used conciliation court against large corporations in the past which is why I question it here. And while this should be able to be handled via legislation it won't be. Beyond that a workaround could force that to happen.
1024core 1y ago
It's your server. You're free to do whatever you want. You can serve different versions of the page depending on the UserAgent (has been done many times before). You can put up a paywall depending on UserAgent or OS (has been done). In short, it's a 2-way street: the client on the other end of the TCP pipe makes a request, and your server fulfills the request as it sees fit.
miki123211 1y ago
the fact that it would be discovered almost immediately. If you give them a URL that does not appear in Google, ask them to visit that URL specifically, and then notice the content from that URL in the training data, it's proof that they're doing this, which would be quite damaging to them.
- Freak_NL 1y ago
  > […] it's proof that they're doing this, which would be quite damaging to them. Is it? It's damning, but is it damaging at all? I'm not getting the impression that anyone's data being available for training if some bot can get to it is just how things are now, rather than an unsettled point of contention. There's too much money invested in this thing for any other outcome, and with the present decline of the rule of law…
autoexec 1y ago
Nothing, and that's why I expect they all do it.

tempfile 1y ago

The way to prevent people from downloading your pages and using them is to take them off the public internet. There are laws to prevent people from violating your copyright or from preventing access to your service (by excessive traffic). But there is (thankfully) no magical right that stops people from reading your content and describing it.

bayindirh 1y ago
Many site operators want people to access their content, but prevent AI companies from scraping their sites for training data. People who think like that made tools like Anubis, and it works. I also want to keep this distinction on the sites I own. I also use licenses to signal that this site is not good to use for AI training, because it's CC BY-NC-SA-2.0. So, I license my content appropriately (No derivative, Non-commercial, shareable with the same license with attribution), add technical countermeasures on top, because companies doesn't respect these licenses (because monies), and circumvent these mechanisms (because monies), and I'm the one to suck this up and shut-up (because their monies)? Makes no sense whatsoever.
- hombre_fatal 1y ago
  I guess that's a question that might be answered by the NYT vs OpenAI lawsuit at least on the enforceability of copyright claims if you're a corporation like NYT. If you don't have the funds to sue an AI corp, I'd probably think of a plan B. Maybe poison the data for unauthenticated users. Or embrace the inevitability. Or see the bright side of getting embedded in models as if you're leaving your mark.
- tempfile 1y ago
  Of course some people want that. And at the moment they can prevent it. But those methods may stop working. Will it then be alright to do it? Of course not, so why bother mentioning that they are able to prevent it now - just give a justification. Your license is probably not relevant. I can go to the cinema and watch a movie, then come on this website and describe the whole plot. That isn't copyright infringement. Even if I told it to the whole world, it wouldn't be copyright infringement. Probably the movie seller would prefer it if I didn't tell anyone. Why should I care? I actually agree that AI companies are generally bad and should be stopped - because they use an exorbitant amount of bandwidth and harm the services for other users. At least they should be heavily taxed. I don't even begrudge people for using Anubis, at least in some cases. But it is wrong-headed (and actually wrong in fact) to try to say someone may or may not use my content for some purpose because it hurts my feelings or it messes with my ad revenue. We have laws against copyright infringement, and to prevent service disruption. We should not have laws that say, yes you can read my site but no you can't use it to train an LLM, or to build a search index. That would be unethical. Call for a windfall tax if they piss you off so much.
  - bayindirh 1y ago
    > I can go to the cinema and watch a movie, then come on this website and describe the whole plot. That isn't copyright infringement. This is a false analogy. A correct one would be going to a 1000 movies and creating the 1001th movie with scenes cropped from these 1000 movies and assemble it as a new movie, and this is copyright infringement. I don't think any of the studios would applaud and support you for your creativity. > But it is wrong-headed (and actually wrong in fact) to try to say someone may or may not use my content for some purpose because it hurts my feelings or it messes with my ad revenue. Why does it have to be always about money? Personally it's not. I just don't want my work to be abused and sold to people to benefit a third party without my consent and will (and all my work is licensed appropriately for that). > We should not have laws that say, yes you can read my site but no you can't use it to train an LLM, or to build a search index. This goes both ways. If big corporations can scrape my material without asking me and resell it as an output of a model, I can equally distill their models further and sell it as my own. If companies can scrape my pages to sell my content as theirs, I can scrape theirs and unpaywall them. But that will be copyright infringement, just because they have more money. What angers me is "all is fair game because you're a small fish, and this is a capitalist marketplace" mentality. If companies can paywall their content to humans that don't pay, I can paywall AI companies and demand money or push them out of my lawn, just because I feel like that. The inverse is very unethical, but very capitalist, yes. It's not always about money. P.S.: Oh, try to claim that you can train a model with medical data without any clearance because it'd be unethical to have laws limiting this. It'll be fun. Believe me.
- zzo38computer 1y ago
  I don't want AI companies to scrape my sites (or use the files I wrote) for training data either, but that is not specifically what I am trying to stop (unless the files are supposed to be private and unpublished). I should not stop them from using the files for what they want, once they have them. (I also specifically do not want to block use of lynx, curl, Dillo, etc.) What I want to stop is excessive crawling and scraping of my server. Once they have the file they can do what they want with it. Another comment (44786237) mentions that robots.txt is only for restricting recursive access; I agree and that is what should be blocked. They also should not access the same file several times quickly even though it should be unnecessary to do so, just as much as they should not access all of the files. (If someone wants to make a mirror of the files, there may be other ways, e.g. in case there is a archive file available to download many at once (possibly, in case if the site operator made their own index and then did it this way). If it is a git repository, then it can be cloned.)
- account42 1y ago
  > Many site operators want people to access their content, but prevent AI companies from scraping their sites for training data. That is unfortunately not a distinction that is currently legally enforceable. Until that changes all other "solutions" are pointless and only cause more harm. > People who think like that made tools like Anubis, and it works. It works to get real humans like myself to stop visiting your site while scrapers will have people whose entire job is to work around such "protections". Just like traditional DRM inconveniences honest customers and not pirates. And to be clear, what you are advocating for is DRM. > I also want to keep this distinction on the sites I own. I also use licenses to signal that this site is not good to use for AI training, because it's CC BY-NC-SA-2.0. If AI crawlers cared about that we wouldn't be talking about this issue. A license and only give more permissions than there are without one.
  - bayindirh 1y ago
    > It works to get real humans like myself to stop visiting your site If we talk about Anubis, it's pretty invisible. You wait a couple of seconds in the first visit, and don't get challenged for a couple of weeks, at least. With more tuning some of the sites using Anubis work perfectly well without ever seeing Anubis' wall while stopping AI crawlers. > And to be clear, what you are advocating for is DRM. Yes. It's pretty ironic that someone like me who believes in open access prefers a DRM solution to keep companies abusing the small fish, but life is an interesting phenomenon, and these things happen. > Until that changes all other "solutions" are pointless and only cause more harm. As an addendum to above paragraph, I'm not happy that I have to insert draconian measures between the user and the information I want to share, but I need a way to signal that I'm not having their ways to these faceless things. What do you propose? Taking my sites offline? Burning myself in front of one of the HQs? > If AI crawlers cared about that we wouldn't be talking about this issue. A license and only give more permissions than there are without one. AI crawlers default to "Public Domain" when they find no licenses. Some of my lamest source code repositories made into "The Stack" because I forgot to add COPYING.md. A fork of a GPLv2 tool I wrote some patches also got into "The Stack", because COPYING.md was not in the root folder of the repository. I'd rather add licenses (which I can accept) to things rather than leave them as-is, because AI companies also eagerly grab things without license. All licenses I use mandate attribution and continuation of license, at least, and my blog doesn't allow any derivations of from what I have written. So you can't ingest it into a model to be derived and remixed with something else.

accrual 1y ago

Thanks for sharing your experience. A little off-topic but I'd like to start hosting some personal content, guides/tutorials, etc. Do you still see authentic human traffic on your domains, is it easy to discern? I feel like I missed the bus on running a blog pre-AI.

kldg 1y ago
if you do analytics, it is not so hard, but then you need to store user data (if not directly, then worse, with a third party), which should be viewed as a liability. I see ~2/3 human traffic, ~1/3 bot traffic (I just parse user agent strings and count whitelisted browsers as human), but my main landing page is all dynamic-populated webgl. I just asked Gemini what it sees on website, and it states "The page appears to be loading, with the text "Loading room data...".[1] There are also labels for "BG", "FG", and "CURSOR", and a background weather animation." -so I can be feel reasonably confident I don't need to worry about AI, for now; it needs a machine-friendly frontend. you could go proper insanomode, too. remaking The Internet is trivial if you don't care about existing web standards -- replacing HTTP with your own TCP implementation, getting off html/js/css, etc. being greenfield, you can control the protocol, server, and client implementation, and put it in whatever language you want. I made a stateful Internet implementation in Python earlier for proof-of-concept, but I want to port it and expand on it in rust soon (just for fun; I don't do serious biznos). you'll very likely have 100% human traffic then, even if you're the only person curious and trusting enough to run your client.
- 71bw 1y ago
  > I made a stateful Internet implementation in Python earlier for proof-of-concept Is there a repo or some other form of public access? I'd like to see this.
  - kldg 1y ago
    it's not in a shareable state; is unsafe as-is. can share general idea and sample "webpage" files, though. the server ("lodge") passes JSON to the client from what are called .branch files. the client receives JSON, parses it, then builds the UI and state representation from the JSON, then stored in that client's memory (self.current_doc and self.page_state in python client). branches can invoke waterwheel (.ww) files hosted on the lodge. waterwheel files on the lodge contain scripts which define how patches (as JSON) are to be sent to the client. the client updates its state based on the JSON patch it receives. sample .branch and .ww from python implementation (in pastebin so to not make everyone have to scroll through this): https://pastebin.com/A0DEZDmR https://pastebin.com/A0DEZDmR

tintor 1y ago

technical limitations / data poisoning measures

Aeolun 1y ago

> What prevents these companies from keeping a copy of that particular page, which I specifically disallowed for bot scraping, and feed it to their next training cycle? What prevents anyone else? robots.txt is a request, not an access policy.

utbabya 1y ago
This honor system mostly worked at scale because interests align, which seems to be no longer the case. Does information no longer wants to be free now? Maybe internet, just like social media was just a social experiment at the end, albeit a successful one. Thanks GenAI.
- windexh8er 1y ago
  Can the Terms of Service of individual content creators leverage a "death of a thousand cuts" model to produce a legal honeypot which would require organizations like Perplexity to be bound up in 10s of thousands of conciliation court cases? Big Tech has hidden behind ToS for years. Now, it seems as though it only works for them, but not against. It seems as though this would be easy to orchestrate and prove forcing these companies into a legal nightmare or risk insolvent business stature due to the high load of cases filed against. Why couldn't something like this be used to flip the table? A conciliation brigading, of sorts.
  - Eisenstein 1y ago
    Because lawyers are expensive and big tech companies have lots of them. Because it takes a ton of time and effort to sue someone. Because you need to show standing, which means you need to be able to demonstrate you lost something of value by their actions. Because the power imbalance is heavily weighted towards a corporation. Because the way to deal with such things should be legislation and not court decisions. And lots more reasons...
    - windexh8er 1y ago
      That's exactly why I said conciliation court. None of what you've outlined is required nor is it expensive. But, for each case, the defendant is still required to show up. I've successfully used conciliation court against large corporations in the past which is why I question it here. And while this should be able to be handled via legislation it won't be. Beyond that a workaround could force that to happen.

1024core 1y ago

It's your server. You're free to do whatever you want. You can serve different versions of the page depending on the UserAgent (has been done many times before). You can put up a paywall depending on UserAgent or OS (has been done). In short, it's a 2-way street: the client on the other end of the TCP pipe makes a request, and your server fulfills the request as it sees fit.

miki123211 1y ago

the fact that it would be discovered almost immediately. If you give them a URL that does not appear in Google, ask them to visit that URL specifically, and then notice the content from that URL in the training data, it's proof that they're doing this, which would be quite damaging to them.

Freak_NL 1y ago
> […] it's proof that they're doing this, which would be quite damaging to them. Is it? It's damning, but is it damaging at all? I'm not getting the impression that anyone's data being available for training if some bot can get to it is just how things are now, rather than an unsettled point of contention. There's too much money invested in this thing for any other outcome, and with the present decline of the rule of law…

autoexec 1y ago

Nothing, and that's why I expect they all do it.