AI giants learn what everyone else on the modern internet already knows

In a striking twist of irony, major AI companies, including Anthropic, OpenAI, and Google, are now confronting a predicament that many internet users have long understood: once content is shared online, it can be repurposed in unintended ways. These tech giants have long defended their practices of utilizing publicly available information for AI training, citing 'fair use'. However, they are now experiencing pushback from content creators who feel their work is being exploited without consent. This clash has recently escalated around the concept of 'distillation', a process where outputs from one AI model are used to enhance another. Anthropic has raised alarms over competitors allegedly siphoning off their model's outputs, effectively converting years of investment into a shortcut for others. OpenAI and Google have echoed similar concerns, underscoring the anxiety that arises when substantial investments in AI development can be undermined by rivals tapping into those same resources for a fraction of the cost. The irony is palpable. While these companies have built their models by scraping vast amounts of web content, they are now scrambling to protect their own outputs from similar treatment. Anthropic claims that its rivals are extracting information from their top-performing models, while at the same time, website owners have accused Anthropic of doing the same to their content over the past few years. This ongoing tug-of-war reveals a striking symmetry in the argument, as both sides accuse each other of violating terms of service. Despite positioning itself as an ethical pioneer in AI development, Anthropic's aggressive data collection methods have drawn criticism. Their bots reportedly crawl websites thousands of times, far exceeding the number of referrals sent back to those sites. Meanwhile, OpenAI and Google are framing this conflict as a cybersecurity concern, pointing to the 'attacks' on their models by swarms of bots seeking to extract intelligence. However, critics argue that these companies have similarly bombarded web platforms, raising operational costs for site owners as they deal with the influx of bot traffic. AI researchers maintain that there is a distinction between distillation and web scraping, but the industry remains divided on where to draw the line. Traditional distillation typically involves the use of one’s own model outputs to create new, often smaller models. In contrast, what Anthropic describes as 'distillation attacks' refers to the practice of competitors leveraging external AI outputs to enhance their own systems. The blurred lines have led to concerns that the backlash from Anthropic could stifle legitimate forms of distillation across the board, prompting what open-source AI expert Nathan Lambert terms 'distillation panic'. What’s evident is that the tech giants are navigating a complex landscape where they can freely extract information from the internet while simultaneously decrying the same practices directed at their proprietary outputs. The ongoing developments reveal a harsh truth: once information is online, it is ripe for collection, remolding, and monetization. As Zilan Qian from the Oxford China Policy Lab aptly puts it, this dynamic is a perpetual 'cat-and-mouse game'. As long as AI outputs circulate publicly, there will always be innovative ways to access and utilize that information. For these AI giants, the challenge lies in reconciling their claims of ethical use with the realities of a rapidly evolving digital landscape. As they grapple with these challenges, the implications for the future of AI development and content ownership remain profound.