If turning other peoples copyrighted work into a model is transformative enough to be protected then so is distilling that model into a different, better, model.
Whether we think they're paying enough is another question, but "I'm paying for content so can protect it" doesn't seem inconsistent.
We may decide that giving models away for free means they don't have to license content (judging by HN comments), but currently that doesn't seem to be the case as Meta is facing lawsuits for its open models.
(Obligatory stratechery piece: https://stratechery.com/2026/whos-afraid-of-chinese-models/ )
The same principle can be applied to distillation - it is a fair use. You just shouldn't use illegal ways to access the models being distilled.
To the commenter below: if it is illegal - has the police/FBI report been made? Otherwise it is just a civil court matter.
It does seem to be becoming the norm for AI companies to licence premium content in America, judging by the deals they're making. It doesn't seem to be done by the international distillers. It's a cost that American open models will seem to have to pay but not international.
International distillers doesn't use that premium content, so they don't pay for it. They do pay for their access to the models they are distilling. Thus providing the revenue stream to those models. Thus those models make profit off the content they used for training. The content they mostly have't paid for.
>It's a cost that American open models will seem to have to pay but not international.
It goes both ways - American companies and their business are protected by American laws and have access to the market protected by those laws, etc.
This doesn't seem to be true. They are training on their own scraped data overwhelmingly (we can extract copyright data from, eg, deepseek). They couldn't get nearly enough tokens through the American APIs to train a model on alone.
> American companies and their business are protected by American laws and have access to the market protected by those laws
Absolutely. Currently international providers are selling inference on the American market though, I don't know how that will sit legally the way things are currently going.
This is a very surprising claim to me (and I imagine many small website owners who keep getting scraped by Anthropic and OpenAI).
Do you have a source?
https://digiday.com/media/a-timeline-of-the-major-deals-betw...
I don't really understand why you think they're relevant, given that this conversation is about the training itself.
Even the news orgs say explicitly in the press releases that it's about training on their archive
eg. http://ap.org/media-center/press-releases/2023/ap-open-ai-ag...
---
edit, examples:
Wiley https://newsroom.wiley.com/press-releases/press-release-deta...
Shutterstock https://investor.shutterstock.com/news-releases/news-release...
Axel Springer https://openai.com/index/axel-springer-partnership
Stack Overflow: https://stackoverflow.co/partnerships
Disney (for characters in video. Video is especially where licensing is a big difference internationally right now) https://openai.com/index/disney-sora-agreement
etc.
The news corp one had a leaked price ($250mill), so they don't seem to be insignificant. These would have to be included in API prices I presume.
In any case the laws are being written now, but I doubt these will have worse protection than software does, which has far better protections than copyright
Software is protected by copyright. Some software may also be protected by patents, but last time I checked, AI generated output of any kind was not patentable.
Also note that the OpenAI/Anthropic argument is that the model training is sufficiently transformative to satisfy the fair use of the original content for training.
By that same argument, when distilling the distillers aren't using the original content the OpenAI/Anthropic models were trained on - the distillers are interacting only with the "sufficiently transformed" content of the OpenAI/Anthropic models and are normally paying for that.
There is also that old phonebook rule that facts can't be copyrighted. So, if i asked the model about bunch of phone numbers, i can publish the resulting list, can train my model on it, etc. Such approach doesn't allow to reproduce copyrighted works of course - and as we know the AI output isn't copyrightable, so it looks like basically any output i get i can use whatever way i like.
Also, if model output distillation is shown as some form of reverse engineering I assume the DMCA can apply
I agree that you can't patent a book, but I would point out that you can patent an idea, which may only appear in a book or journal article.
For example, a patent describing a chemical process. The actual idea of how to do it is public domain, go look up the patent. Print it out. Do whatever with those words. Its fine. Building a plant to go do that chemical process to make that same output chemical in that same way, that's IP infringement. Its not the words, its the idea.
Who are you saying owns that IP? The people who trained the model? The people who ran the model? The people who wrote the prompt? The person who paid for all of that to happen?
If the model output is owned by the person prompting it and paying for the tokens, what's the problem here?
If the model output is owned by the trainer of the model, that's a big nasty can of worms.
I mean otherwise it’s a very slippery slope, effectively it would give Anthropic the ownership of any code generated by its models..