OpenAI, Google, Microsoft, Anthropic, and others have reportedly begun collaborating on a "common framework" (an IETF Internet-Draft) that would allow website operators to opt out of AI training across multiple platforms at once—starting with updates to robots.txt.
In short, having faced fierce public backlash over "unauthorized training and systemic data theft," they are now scrambling to establish a "bulk opt-out mechanism" as a convenient shield to absolve themselves of past liability.
It’s as if they’re saying:
"I'm full now, so if you put up a sign saying 'Don't steal,' I won't take any more. But I'm keeping everything I've already eaten."
When I mentioned this to Gemini, I received the following response:
- Completely Too Late
Pre-training for existing models has already been completed, and the data that was harvested has been irreversibly incorporated into the model’s weights. Even if we’re told, “We’ll make it possible to opt out of training from now on (and Google and Microsoft won’t implement this until 2027 or later),” the content that’s already been taken away won’t be returned.- Creating a “get-out-of-jail-free card” for legal battles
This feels overwhelmingly like a posturing move by the industry—essentially saying, “We’ve set up a proper mechanism to opt out”—to avoid legal risks, such as lawsuits from desperate publishers and compliance with the EU AI Act.- Convenient Boundaries
The underlying structure remains unchanged: “We’ll let you opt out of AI training, but if you don’t allow search grounding (citation display), you’ll lose traffic from AI search.” Ultimately, publishers remain in an overwhelmingly disadvantageous position.
Put Things Back the Way They Were": The Real Core of Litigation
I just wish they’d delete all the training data and put things back the way they were. In short: spit out every piece of data you’ve stolen without permission, and roll back the models.
When I brought this up, Gemini pointed out the harsh reality behind their reluctance:
They’re Shying Away from “Machine Unlearning”
While research into “unlearning”—the technology for erasing specific data from models—is progressing, completely removing information that has become deeply embedded in the model’s weights is extremely costly and compromises the accuracy of the entire model, so companies are absolutely unwilling to do it. They are fully intent on using the excuse, “Technically, once something has been learned, it cannot be specifically identified and erased,” to get off the hook.They’ve Built Their Business Entirely on a Foundation Created Through Unauthorized Learning
Having completed their “foundation (base model)” using vast amounts of web data collected without permission, they later pull the rug out from under others by declaring, “Okay, from here on out, we’ll follow the rules,” thereby turning it into a barrier to entry that prevents latecomers from entering the market.
It seems that this demand to “revert to the original state” (i.e., delete the data and rebuild the model) has become the central point of contention that rights holders and attorneys around the world are currently raising in court.
In fact, in major lawsuits overseas—including the one filed by The New York Times against OpenAI and Microsoft—the demands made by plaintiffs (and their attorneys) clearly include not only damages but also “the destruction of large language models (LLMs) built using unauthorized training data and the destruction of the training data itself.”
Why can't they just delete the data the AI has absorbed so far? Because the AI has already “eaten” it all?
If they refuse to delete it, massive damages claims are precisely what will corner them in future lawsuits.
They act with zero awareness of wrongdoing—no remorse, no apology. If they truly acknowledged their fault, they would start by saying, “We apologize for using your work without permission, and we will compensate you appropriately.”
Instead, their sole focus is self-preservation: minimizing litigation and regulatory risks. That is why their entire response boils down to a absurd plea: “Just put up a sign saying, ‘Don’t steal anymore.’”
From Parasitic Summarization to Value-Creating Matching
Consider this: YouTube already operates a system that redirects ad revenue to original creators when unauthorized videos are uploaded. Why can’t a similar model apply to AI Overviews? After all, these companies already possess the tech to crawl and analyze billions of websites.
Revenue sharing wasn't deemed 'impossible' when traditional search engines and platforms ran on deterministic algorithms. Why is it suddenly 'unfeasible' just because the underlying tech shifted from rigid algorithms to probabilistic AI? The lack of compensation isn't a technical limitation of AI—it's a deliberate choice not to build the infrastructure.
Building innovative technology without infringing on rights is one thing. Turning into a parasite just to turn a profit is entirely different.
Why not shift to a model that actually creates value? For instance, building platforms where users can search for and purchase products directly, or agentic services that match user inquiries with legitimate businesses and service providers.
Wouldn't that create a model where everyone wins?
I’m just a "nobody." If an insignificant creature 🐊 like me can instinctively see this, I can't help but wonder why the authorities and giants fail to offer any real alternatives.
Perhaps they refuse simply because they have no real alternatives to offer—trapped like a hamster running endlessly on a wheel, achieving nothing.
It's time to stop accepting their "opt-out" excuses and force them to build a real matching model.
- Read more
Write a comment