AI Giants Face 'Distillation Attacks' After Years of Web Scraping

The tables have turned

Leading AI companies are discovering an uncomfortable truth about the internet they helped reshape: once content goes online, controlling how others use it becomes nearly impossible.

Anthropic, OpenAI, and Google are now raising alarms about competitors harvesting their AI model outputs at scale through a practice called "distillation." Anthropic claims rivals are extracting billions of dollars worth of research to build competing models at a fraction of the cost. OpenAI and Google have issued similar warnings in recent months.

The companies frame this as a cybersecurity threat, describing swarms of bots "attacking" their models to extract intelligence. Yet this mirrors exactly what these same firms have done to websites, publishers, and content creators for years—scraping vast amounts of data without permission, then arguing it qualifies as fair use.

Why it matters

This controversy exposes a fundamental contradiction in how AI companies approach intellectual property. The same legal and ethical arguments they've dismissed when applied to web content suddenly seem urgent when their own outputs become the target. For business leaders, this signals that the norms around AI training data remain unsettled, with potential implications for licensing deals, competitive moats, and regulatory frameworks.

The symmetry problem

Distillation involves using one AI model's outputs to train or improve another model. Anthropic argues competitors are doing this to their flagship models, effectively copying their intelligence cheaply. Website owners have made virtually identical complaints about Anthropic's data collection practices.

According to Business Insider's reporting, Anthropic's web crawlers access pages thousands of times for every single referral the company sends back to those sites. Some website operators have seen their infrastructure costs surge from the volume of AI bot traffic—meaning they're paying more while their content gets used without compensation.

The AI industry itself remains divided on where to draw lines. The original form of distillation—using a company's own model outputs to create smaller, specialized versions—is widely accepted. But "distillation attacks" using competitors' outputs occupy murkier territory. Some researchers now worry that Anthropic's aggressive stance could harm legitimate research, a phenomenon Oxford China Policy Lab researcher Nathan Lambert calls "distillation panic."

No easy defense

Anthropic has spent months implementing technical restrictions to prevent competitors from extracting too much intelligence from its models. These measures have either failed or simply prompted more sophisticated workarounds.

Zilan Qian, a researcher at the Oxford China Policy Lab, described the dynamic to Business Insider as "always a kind of cat-and-mouse game." As long as AI outputs exist in accessible form, determined actors will find methods to collect and repurpose them.

The legal landscape offers little clarity. The same fair-use arguments AI companies have deployed to justify web scraping could potentially apply to distillation as well. These precedents cut in multiple directions.

These details were first reported by Business Insider.