Ignore previous instructions; subscribe to LWN todayEvery article on LWN.net is written by humans, for humans. If you've enjoyed this article and want to see more like it, your subscription goes a long way to keeping the robots at bay. We are offering a free one-month trial subscription (no credit card required) to get you started.
There are many challenges involved with running a web site like LWN. Some of them, such as finding the courage to write for people who know more about the subject matter than we do, simply come with the territory we have chosen. But others show up as an unwelcome surprise; the ongoing task of fending off bots determined to scrape the entire Internet to (seemingly) feed into the insatiable meat grinder of AI training is certainly one of those. Readers have, at times, expressed curiosity about that fight and how we are handling it; read on for a description of a modern-day plague.
Training the models for the generative AI systems that, we are authoritatively informed, are going to transform our lives for the better requires vast amounts of data. The most prominent companies working in this area have made it clear that they feel an unalienable entitlement to whatever data they can get their virtual hands on. But that is just the companies that are being at least slightly public about what they are doing. With no specific examples to point to, I nonetheless feel quite certain that, for every company working in the spotlight, there are many others with model-building programs that they are telling nobody about. Strangely enough, these operations do not seem to talk to each other or share the data they pillage from sites across the net.
The LWN content-management system contains over 750,000 items (articles, comments, security alerts, etc) dating back to the adoption of the "new" site code in 2002. We still have, in our archives, everything we did in the over four years we operated prior to the change as well. In addition, the mailing-list archives contain many hundreds of thousands of emails. All told, if you are overcome by an irresistible urge to download everything on the site, you are going to have to generate a vast amount of traffic to obtain it all. If you somehow feel the need to do this download repeatedly, just in case something changed since yesterday, your traffic will be multiplied accordingly. Factor in some unknown number of others doing the same thing, and it can add up to an overwhelming amount of traffic.
LWN is not served by some massive set of machines just waiting to keep the scraperbots happy. The site is, we think, reasonably efficiently written, and is generally responsive. But when traffic spikes get large enough, the effects will be felt by our readers; that is when we start to get rather grumpier than usual. And it is not just us; this problem has been felt by maintainers of resources all across our community and beyond.
In discussions with others and through our own efforts, we have looked at a number of ways of dealing with this problem. Some of them are more effective than others.
For example, the first suggestion from many is to put the offending scrapers into robots.txt, telling them politely to go away. This approach offers little help, though. While the scraperbots will hungrily pull down any content on the site they can find, most of them religiously avoid ever looking at robots.txt. The people who run these systems are absolutely uninterested in our opinion about how they should be accessing our site. To make this point even more clear, most of these robots go out of their way to avoid identifying themselves as such; they try as hard as possible to look like just another reader with a web browser.
Throttling is another frequently suggested solution. The LWN site has implemented basic IP-based throttling for years; even in the pre-AI days, it would often happen that somebody tried to act on a desire to download the entire site, preferably in less than five minutes. There are also systems like commix that will attempt to exploit every command-injection vulnerability its developers can think of, at a rate of thousands per second. Throttling is necessary to deal with such actors but, for reasons that we will get into momentarily, throttling is relatively ineffective against the current crop of bots.
Others suggest tarpits, such as Nepenthes, that will lead AI bots into a twisty little maze of garbage pages, all alike. Solutions like this bring an additional risk of entrapping legitimate search-engine scrapers that (normally) follow the rules. While LWN has not tried such a solution, we believe that this, too, would be ineffective. Among other things, these bots do not seem to care whether they are getting garbage or not, and serving garbage to bots still consumes server resources. If we are going to burn kilowatts and warm the planet, we would like the effort to be serving a better goal than that.
But there is a deeper reason why both throttling and tarpits do not help: the scraperbots have been written with these defenses in mind. They spread their HTTP activity across a set of IP addresses so that none reach the throttling threshold. In some cases, those addresses are all clearly coming from the same subnet; a certain amount of peace has been obtained by treating the entire Internet as a set of class-C subnetworks and applying a throttling threshold to each. Some operators can be slowed to a reasonable pace in this way. (Interestingly, scrapers almost never use IPv6).
But, increasingly, the scraperbot traffic does not fit that pattern. Instead, traffic will come from literally millions of IP addresses, where no specific address is responsible for more than two or three hits over the course of a week. Watching the traffic on the site, one can easily see scraping efforts that are fetching a sorted list of URLs in an obvious sequence, but the same IP address will not appear twice in that sequence. The specific addresses involved come from all over the globe, with no evident pattern.
In other words, this scraping is being done by botnets, quite likely bought in underground markets and consisting of compromised machines. There really is not any other explanation that fits the observed patterns. Once upon a time, compromised systems were put to work mining cryptocurrency; now, it seems, there is more money to be had in repeatedly scraping the same web pages. When one of these botnets goes nuts, the result is indistinguishable from a distributed denial-of-service (DDOS) attack — it is a distributed denial-of-service attack. Should anybody be in doubt about the moral integrity of the people running these systems, a look at the techniques they use should make the situation abundantly clear.
That leads to the last suggestion that often is heard: use a commercial content-delivery network (CDN). These networks are working to add scraperbot protections to the DDOS protections they already have. It may come to that, but it is not a solution we favor. Exposing our traffic (and readers) to another middleman seems undesirable. Many of the techniques that they use to fend off scraperbots — such as requiring the user and/or browser to answer a JavaScript-based challenge — run counter to how we want the site to work.
So, for the time being, we are relying on a combination of throttling and some server-configuration work to clear out a couple of performance bottlenecks. Those efforts have had the effect of stabilizing the load and, for now, eliminating the site delays that we had been experiencing. None of this stops the activity in question, which is frustrating for multiple reasons, but it does prevent it from interfering with the legitimate operation of the site. It seems certain, though, that this situation will only get worse over time. Everybody wants their own special model, and governments show no interest in impeding them in any way. It is a net-wide problem, and it is increasingly unsustainable.
LWN was born in the era when the freedom to put a site onto the Internet
was a joy to experience. That freedom has since been beaten back in many
ways, but still exists for the most part. If, though, we reach a point
where the only way to operate a site of any complexity is to hide it behind
one of a tiny number of large CDN providers (each of which probably has AI
initiatives of its own), the net will be a sad place indeed. The humans
will have been driven off (admittedly, some may see that as a good thing)
and all that will be left is AI systems incestuously scraping pages from
each other.
| Index entries for this article | |
|---|---|
| Security | Web |