> According to Anthropic’s commercial terms of service, customers are barred from using the service to “build a competing product or service, including to train competing AI models”
That's... quite a license term. I'm a big fan of tools that come with no restrictions on their use in their licenses. I think I'll stick with them.
For years it was a license violation to use Microsoft development tools to build a word processor or spreadsheet. It was also a violation of your Oracle license to publish benchmark results comparing Oracle to other databases.
If you compete with a vendor, or give aid and comfort to their competitors, do not expect the vendor to play nice with you, or even keep you on as a customer.
It wasn't a blanket prohibition, but a restriction on some parts of the documentation and redistributable components. Definitely was weird to see that in the EULA for a toolchain. This was removed later on, though I forget if it's because they changed their mind or removed the components.
These anti-competitive clauses are becoming standard across all major AI providers - Google, Microsoft, and Meta have similar terms. The industry is converging on a licensing model that essentially creates walled gardens for model development.
Yeah, if I remember correctly iTunes had a clause it couldn’t be used for nuclear development.
Not sure what Apple lawyers were imagining but I guess barring Irani scientist from syncing their iPods with uranium refiner schematics set back their programme for decades.
Glad to live in a sane jurisdiction, where provisions made available only after purchase and those that go against typical customer expectations are simply invalid, so I never had to care about EULAs.
Oracle licenses 100% restrict reverse engineering it's product to build a competing once, which is probably the closest to what these AI giants are trying to restrict.
Oracle db products are not meant to build databases, unlike LLM code generator which are meant to build any kind of software, so the restriction sounds a bit different.
Imagine if Oracle was adding a restrictions on what you are allowed to build with Java, that would be a more similar comparison IMO.
You can you just have to ask. And that's not an oracle thing. All the commercial databases have that rule. It's too easy to make misleading benchmarks for such complicated products so that's why they do it.
"Everyone else is doing it" doesn't make it right.
It also makes it dangerous to become dependent on these services. What if at some point in the future, your provider decides that something you make competes with something they make, and they cut off your access?
When that provider's ToS allows them full access to the inputs/outputs you're sending through their system, there is a strong incentive to build something competitive with you once you're proven profitable enough.
I don't know how companies currently navigate that.
Oh I wonder if that applies to me? I've been using claude to do experiments with using SNN's for language models. Doubt anything will come of it... has mostly just been a fun learning experience, but it is technically a "competing product" (one that doesn't work yet)
Exactly this. Strange that this comment got downwoted. AI companies are scrapping the entire internet disregarding copyright and pirating books. Without it, models will be useless.
- Contracts can have unenforceable terms that can be declared null and void by a court, any decision not to renew the contract in future would have no bearing on the current one.
- there are plenty of restrictions on when/ whether you can turn down business for example FRAND contracts or patents don’t allow you choose to not work with a competitor and so on.
People always say "this wouldn't hold up in court" and "the law doesn't work like that" when it comes to contract, but in reality, contracts can mostly contain whatever you want.
I see no reason why Anthropic can't arbitrarily ban OpenAI, regardless of my opinion on the decision. Anthropic hasn't "patented" access to the Claude API; there are no antitrust concerns that I can see; etc.
IMO its all well and good for Anthropic to adhere and or justify a ban under it’s TOS. It’s been super annoying when (myself and others) have been arbitrarily banned with very little communication on how to remedy or adhere to ToS for what feels like it’s intended purpose?
The biggest lesson I learnt from my law degree is that sure you might be legally entitled to it - but you can still be receiving a raw deal and have very little in the way of remedial action.
Nobody was asking if Anthropic can ban OpenAI. I believe they were asking if the contract that can ban using the output to train an AI would hold up in court.
And no, it isn't clear to me that this contract term would hold up in court as Anthopic doesn't have copyright ownership in the AI output. I don't believe you can enforce copyright related contracts without copyright ownership.
I could be wrong of course, but I find it odd this topic comes up from time to time but apparently nobody has a blog post by a lawyer or similar to share on this issue.
> And no, it isn't clear to me that this contract term would hold up in court as Anthopic doesn't have copyright ownership in the AI output
They don't need copyright ownership of the AI output to make part of the the conditions for using their software running on their servers (API access) an agreement not to use it training a competing AI model.
There would have to be a law prohibiting that term, either in general or for a company in the specific circumstances Anthropic is in. (The “specific circumstances” thing is seen, e.g., when a term is otherwise permitted but used but a firm that is also a monopoly in a relevant market as a way of either defending or leveraging that monopoly, and thus it becomes illegal in that specific case.)
"They don't need copyright ownership of the AI output to make part of the conditions for using their software running on their servers (API access) an agreement not to use it training a competing AI model."
You are missing the point.
Copyright law and the copyright act, not general contract law, governs whether a contract provision relating to AI output can be enforced by Anthropic, and since copyright law says Anthropic has no copyright in the output, Anthropic will not win in court.
It's not different than if Anthropic included a provision saying you won't print out the text of Moby Dick. Anthropic doesn't own copyright on Moby Dick and can't enforce a contract provision related to it.
Like I said I can be convinced I'm wrong based on a legal analysis from a neutral party but you seem to be arguing from first principles.
No, I am disagreeing with the point, because it's completely wrong.
> Copyright law and the copyright act, not general contract law, governs whether a contract provision relating to AI output can be enforced by Anthropic
No, it doesn't. There is no provision of copyright law that limits terms of contracts covering AI outputs.
> It's not different than if Anthropic included a provision saying you won't print out the text of Moby Dick.
This is true, but undermines your point.
> Anthropic doesn't own copyright on Moby Dick and can't enforce a contract provision related to it.
Actually, providing services that allow you to produce output can enforce provisions prohibiting reproducing works they don't own the copyright to (and frequently do adopt and enforce rules prohibiting this for things other people own the copyright to).
> Like I said I can be convinced I'm wrong based on a legal analysis from a neutral party but you seem to be arguing from first principles.
You seem to be arguing from first principles that are entirely unrelated to the actual principles, statutory content of, or case law of contracts or copyrights, and I have no idea where they come from, but, sure, believe whatever you want, it doesn't cost me anything.
"In jurisdictions that follow the Second Circuit's more restrictive approach, plaintiffs may be limited to bringing copyright infringement claims when the scope of license terms or other contractual restrictions on the use of works has been exceeded. Plaintiffs who do not own or control the copyright interest in the licensed work, however, will not be able to bring such claims and may be left without an enforcement mechanism under traditional contracting approaches."
>>There is no provision of copyright law that limits terms of contracts covering AI outputs.
This isn't how legal reasoning works in a common law system... to discover the answer you usually find the most similar case to the current fact pattern and then compare it to the current issue.
If you are aware of such a case, even colluqually, point me in the right direction. It might be hard to analogize to another cases though, because Anthropic doesn't have a license for most of the training materials that made their model. I've also read you can't contract around a fair use defense.
If I'm wrong it isn't very helpful to shout "na uh" without further explanation. Give me some search engine keywords and I'll look up whatever you point me towards.
Good luck with that! Most of the relevant model providers include similar terms (Grok, OpenAI, Anthropic, Mistal, basically everyone with the exception of some open model providers).
That's... quite a license term. I'm a big fan of tools that come with no restrictions on their use in their licenses. I think I'll stick with them.