Here are some delicious ironies about AI.
Big tech companies have long argued that information available on the internet can be used to develop and output AI models. They call it fair use. The content owner tried to prevent this, but was unsuccessful.
Now, Anthropic, OpenAI, and Google are discovering what the rest of the internet has already learned through painful experience. So once you put something online, you can’t stop it because people will find a way to use it in ways you don’t like.
The latest flashpoint is something called “distillation,” where the output of one AI model is used to improve another AI model. Anthropic says its competitors are harvesting its work at scale, turning billions of dollars of research into shortcuts to their competitors. OpenAI and Google have also issued similar warnings recently.
The fear is palpable. There’s no need to spend billions of dollars building the best AI models when someone else can reproduce much of this intelligence at a fraction of the cost.
That’s a legitimate business concern. But here’s the tricky part.
symmetry
From 30,000 feet, distillation looks a lot like what AI companies are doing to other parts of the internet. Scrape web content for free and without permission. Turn it into a product to sell. Claim it’s fair use. I hope my lawyer can work out the details later.
Anthropic says its competitors are extracting intelligence from its top models. Website owners have claimed for the past three years that Anthropic has extracted information from their websites. Both sides claim this violates their terms of service. Symmetry is hard to ignore.
And despite touting itself as the most ethical AI company, Anthropic is by far the worst company here. For every referral the company sends back to the web, data-sucking bots crawl the web page thousands of times.
bot on both sides
Anthropic, OpenAI, and especially Google see this as a cybersecurity issue, pointing to swarms of bots “attacking” models to extract intelligence. However, they are doing the same for many websites, and site owners are finding their operating costs to be high due to the high amount of bot crawling activity. Some websites not only use your content without permission, but also pay dearly for the privilege.
AI researchers say distillation is different from web scraping. However, the AI industry cannot even decide where to draw the line, whether distillation is OK or not.
Distillation has an inherently harmless form in which laboratories use the output from their own models to create different (often smaller) models. Then there’s what Anthropic calls “distilled attacks.” This is where rivals use someone else’s AI output to develop or improve their own products.
But even here, the lines are blurry, and some AI researchers are now concerned that Anthropic’s aggressive stance could have a negative impact on distillation of any kind. Open source AI expert Nathan Lambert calls this “distilled panic.”
So let me wrap this up from an AI giant’s perspective. They can extract intelligence from the web for free and without permission. It’s different than distillation, and that’s okay. Ah, but not when you use the content in a way they don’t like during distillation.
“Cat and cat”
This distorted debate is being destroyed by the cruel realities of the modern internet. Anthropic has spent months ramping up access to top models to prevent competitors from learning too much. These efforts either backfired or simply created more elaborate workarounds.
Once information is online, smart people will figure out ways to collect it, remix it, and profit from it. That goes for blogs, photos, software code, videos, and yes, the valuable model output of AI giants.
“It’s always a cat-and-mouse game,” Jiran Qian, a research fellow at the Oxford China Policy Institute, told Business Insider. As long as the output of an AI model is out there, “people will probably find a way to access it.”
In fact, extracting another company’s AI models may even be fair use. These legal arguments can go both ways.
Welcome to the new internet, Anthropic, OpenAI, and Google. Get used to it.
Sign up for BI’s Tech Memo newsletter here. Please contact us by email. abarr@businessinsider.com.