For three years the largest AI companies argued that anything published on the internet was fair game for training their models. Content owners objected, sued, and mostly lost. Now Anthropic, OpenAI and Google are complaining that rivals are harvesting the outputs of their own models to build cheaper competitors, and asking the courts and the industry to make it stop. The uncomfortable part, for any UK business deciding how much to bet on a single frontier model, is not the hypocrisy. It is what the reversal reveals: the intelligence these companies spent billions to build does not stay put, and a moat that leaks is worth less than the price tag suggests.
The story underneath the story
The flashpoint is a technique called distillation: using the outputs of one model to train or improve another. In its original form it is routine and uncontroversial. A lab runs its own large model, uses the answers to train a smaller, cheaper one, and ships the compact version. What has changed is who is on the receiving end. Anthropic now says competitors are extracting intelligence from its top models at scale, turning years of research into a shortcut for whoever asks the right questions in the right volume. OpenAI and Google have issued similar warnings.
The business fear is legitimate and easy to state. If a rival can reconstruct much of your model’s capability for a fraction of what it cost you to build, the economic logic of spending billions on the frontier starts to wobble. That is a real threat to a real business model. It is also, viewed from any distance, an almost exact mirror of what these same companies did to everyone who publishes online.
Strategic Reality: The labs’ complaint and the publishers’ complaint are the same complaint. Anthropic says rivals extract intelligence from its models against its terms of service. Website owners spent three years saying Anthropic extracted intelligence from them against theirs. Whichever side you find more sympathetic, the mechanism is identical, and so is the outcome: once value is exposed to the network, someone finds a way to capture it.
The symmetry runs deeper than rhetoric. The labs frame incoming distillation as a security problem, pointing to swarms of automated requests hammering their models to pull intelligence out. Publishers have described exactly that experience from the other direction, watching AI crawlers push their hosting costs up whilst sending almost nothing back. By Business Insider’s account, Anthropic’s crawlers hit webpages thousands of times for every single referral the company sends back to the open web. Sites are not only having their content used without permission; they are paying, in bandwidth and infrastructure, for the privilege of supplying it.
Why the industry cannot draw the line
The tell is that the AI industry cannot agree on where legitimate distillation ends and theft begins. There is the benign version, where a lab distils from its own models. There is what Anthropic calls a “distillation attack”, where a competitor distils from someone else’s. And there is a large grey middle that nobody has cleanly defined.
| Web scraping for training | Distillation from a rival model | |
|---|---|---|
| What is taken | Published text, images, code | Model outputs (answers, completions) |
| Consent | Rarely sought, usually refused | Against terms of service |
| The taker’s argument | Fair use | ”It’s different” / also possibly fair use |
| The loser’s argument | Breach of terms, cost shifting | Breach of terms, cost shifting |
| Who is complaining now | Publishers, since ~2023 | Frontier labs, since ~2025 |
The grey middle is already producing friction inside the field. Some researchers worry that Anthropic’s aggressive posture will chill legitimate distillation of every kind, including the ordinary internal sort that makes smaller, cheaper, more efficient models possible. The open-source researcher Nathan Lambert has a name for the mood: “distillation panic.” Meanwhile the legal ground is anything but settled. As Business Insider notes, distilling another company’s model may itself qualify as fair use, the very doctrine the labs relied on to justify scraping the web. The argument cuts both ways, and the people who sharpened it may not like where the blade lands.
Competitive Reality: A capability advantage at the frontier now has a half-life. The moment a leading model is widely available through an API, its outputs become a training signal for whoever wants to close the gap. Efforts to lock this down tend to backfire or simply provoke more elaborate workarounds. “It’s always a kind of a cat-and-mouse game,” Oxford China Policy Lab researcher Zilan Qian told Business Insider. As long as the outputs are in the world, people will find a way to reach them.
What this changes for the people buying AI, not building it
Most UK organisations are not training frontier models. They are buying access to them, building products on top of them, or publishing the content that trains them. For all three groups, the distillation fight is not a spectator sport.
| Stakeholder | What the leak changes | So what |
|---|---|---|
| UK firms buying AI capability | Frontier performance diffuses downward to cheaper models faster | Capability you needed a premium contract for last year may be affordable this year |
| Software firms building on a single model | The provider’s moat is thinner than its pricing implies | Portability and multi-model design become risk management, not luxuries |
| UK publishers and content owners | The “you can’t stop us” argument now has an awkward precedent | Leverage for licensing and terms enforcement has quietly improved |
| Open-weight and challenger models | Distillation narrows the gap to the frontier | More credible alternatives, more pricing pressure on incumbents |
The through-line is commoditisation. When intelligence leaks from the top of the market toward the bottom, the buyer is usually the beneficiary. Capability that commanded a premium migrates into cheaper models, open-weight releases, and smaller specialised systems. For a UK business weighing a long, expensive commitment to one provider on the strength of a current benchmark lead, that lead is a depreciating asset. Paying a large premium for a gap that the market is actively eroding is a poor trade.
Hidden Cost: The real exposure for most buyers is not which model is best today. It is architecting your product so tightly around one provider that you cannot move when the price, the terms, or the capability ranking changes, and all three are changing faster than annual contracts assume.
How to position for a market where the moat leaks
None of this argues for waiting on the sidelines. It argues for building on the assumption that no single model stays ahead for long. In rough order of priority:
- Design for portability from the start. Keep prompts, evaluation suites and business logic in a layer that is not welded to one vendor’s API. The goal is to change models in weeks, not quarters, when the economics shift.
- Buy the capability, not the brand. Benchmark against the task you actually have, and re-benchmark on a schedule. Cheaper models routinely catch up on specific jobs long before they lead across the board.
- Price contracts for a moving frontier. Favour shorter commitments and portability clauses over multi-year lock-in bought on today’s leaderboard position. You are buying a depreciating advantage; contract accordingly.
- If you own content, treat access as an asset. The precedent that “published means fair game” is being contested by the very firms that established it. That is the moment to review your terms, your crawler policy, and whether your content is a licensing opportunity rather than a free input.
SME Advantage: A smaller UK firm has no legacy contract to defend and no in-house frontier model to protect. That is an advantage. You can pick the cheapest model that clears the bar for each task, switch the moment a better-value option appears, and let the incumbents fund the research that flows downhill to you.
Four things this analysis can get wrong
The commoditisation reading is the most useful one, but it is not guaranteed. Four failure modes are worth holding in mind.
Distillation genuinely is different from scraping, and the courts may say so. The industry’s inability to draw the line is not proof that no line exists. A ruling that treats model outputs as protectable in ways that web content is not would slow the downward diffusion and restore some of the frontier premium.
The frontier keeps moving, so the gap never fully closes. Distillation copies yesterday’s best model. If the leaders keep opening new capability faster than rivals can distil the last generation, the lead is durable in practice even if any single model’s edge is temporary. Buyers still benefit, but the premium for the true frontier persists.
Access tightens enough to matter. Labs are already restricting how their top models can be queried. If those controls stop being a cat-and-mouse game and start actually working, the leak slows and the moat holds longer than this analysis assumes.
Cheaper is not the same as safe. A distilled or open-weight model that matches a benchmark may not match the governance, support, or reliability a business needs. Chasing the lowest price on capability alone can import risk that shows up later, in exactly the places compliance and reputation live.
Reality Check: “The frontier lead is depreciating” is a planning assumption, not a certainty. Build so you win if it holds, and are not ruined if it does not. That means portability and re-benchmarking, not abandoning strong models for cheap ones on principle.
The strategic takeaway
The deepest point in the Business Insider piece is not about hypocrisy, satisfying as the hypocrisy is. It is structural. Anthropic, OpenAI and Google are colliding with the same reality that shaped the earlier web: once something valuable is online and reachable, clever people collect it, remix it, and profit from it, whatever your terms of service say. The labs learned to exploit that reality against publishers. They are now learning to live inside it, on the losing side.
For a UK business, three things follow. First, a frontier model’s lead is real but perishable, so treat it as a depreciating asset rather than a durable moat when you sign contracts and design products. Second, the diffusion of capability downward is, for a buyer, mostly good news, provided you are built to take advantage of it. Third, if you produce content, the argument the AI industry spent three years winning is one it is now quietly arguing against, and that shift in precedent is worth watching closely.
Take Action: Before your next AI contract renewal, run one test. Ask whether you could switch your main model provider inside a month without rewriting your product. If the answer is no, that is not a technical detail. It is your exposure to a market where the leaders themselves cannot keep their intelligence to themselves.
The AI giants are getting a welcome to the internet they helped build. The firms that come out ahead will be the ones who understood the terms of that internet before the labs did: nothing valuable stays locked up for long, and the smart position is to build for the leak rather than bet against it.
Source and attribution
This analysis draws on “AI giants learn what everyone else on the modern internet already knows” by Alistair Barr, published by Business Insider on 13 July 2026, and its cited comments from Oxford China Policy Lab researcher Zilan Qian and open-source AI researcher Nathan Lambert. Original article available at businessinsider.com.
Editorial analysis and UK business framing by Resultsense. We make sense of AI in the UK — turning research, policy and announcements into what they mean for the people building and buying these systems. For more analysis, explore our insights or get in touch.