Comment by nikcub
3 hours ago
> I pretty sure OpenAI and Anthropic are doing the same or worse.
No they're not. It would end both companies if they were ever found to be doing that.
Their terms are clear - if you use the coding plans they can[0] train in return. Enterprise and API, absolutely not.
The argument here is that with the Chinese labs you have zero legal recourse.
[0] opt-in, thanks
>> No they're not. It would end both companies if they were ever found to be doing that. Their terms are clear - The argument here is that with the Chinese labs you have zero legal recourse.
Their terms are not worth shit considering they are reselling you stolen copyrighted data. Even in they terms they started clearly say they retain your data for "safety reasons" for however long they want. Perhaps you didn't watch the space with Anthropic going back and forth with ToS updates(we retain your data for 30 days...stike that and add 30 days or more or no or ..whatever) like my own alpha website.
Your argument boils down to "they've done something I find objectionable, so that means everything they say must be lies".
I'm not comfortable with how these models were trained. I have quite a bit of open source code out there, and I personally see such training as copyright and license laundering.
But that's not how the law sees it, and I grudgingly accept that, regardless of how I may feel, and I don't let my feelings on the matter make me think irrationally when it comes to whether or not these AI companies honor the terms they provide.
Sure, they might be breaking their promises, training on our data when they say they won't. But I do think they most likely aren't, and that it would be corporate suicide if they were and it ever came out.
There is an enormous difference between:
* Exploiting ambiguity around fair use at a large scale before the law catches up and then jointly lobbying with your competition to make sure your interpretation of the law becomes reality.
* Explicitly signing a contract with enterprises to respect their IP and then proceeding to break that contract with your own customers.
The former is firmly in the gray area of legality and doesn't directly hurt your own customers. The latter is both an unambiguous contract violation and a flagrant attack on your own customers' most valuable asset.
https://www.anthropic.com/legal/privacy
> Personal data we collect or receive to train our models
> • Data that our users or crowd workers provide, including Inputs and Outputs from our Services (unless users opt out)
> • Feedback that users explicitly provide about our Services
> • Materials flagged for safety, security, or policy review
While I don’t have visibility into individual corp contracts, hitting tab on a FIM is ‘feedback’, so it is not so clear cut.
2 replies →
retention for 'safety' -> AI race as national security -> training on your data for 'national security' aka safety
It's simple mental calisthenics. If you are handing an organization whose entire business model is built on stealing data with spurious reasoning, what do you actually expect they will do? Don't be a fool.
2 replies →
> Explicitly signing a contract with enterprises to respect their IP and then proceeding to break that contract with your own customers.
You mean all the conditions that are attached to Fable use? My enterprise is deliberately holding off because those are unacceptable.
1 reply →
Anthropic paid a large settlement for the copyrighted data they pirated. So far, US courts have found that it's perfectly fine to train AIs on copyrighted data for which you have legal access.
> Even in they terms they started clearly say they retain your data for "safety reasons" for however long they want.
The discussion was about training, not data retention. Two very different concerns.
And if you're a decent sized customer, most providers have a route to not even retaining the data for safety/security reasons. The reason Anthropic had issues is because they do have a path to "no data storage" for Sonnet/Opus, but not for Fable. Which is why at work we have access to the former, but not the latter.
Whether the terms are worth shit doesn't matter. If they're training on data from paying customers who have requested otherwise and it gets out (which it would, eventually), SAP, Accenture, Deloitte and other huge companies with well-funded legal teams would nuke them from orbit. This is a different area of law from the copyright stuff, different rules/norms/expectations/consequences apply.
They're not training on your data, they're training on "please anonymise this conversation" data.
So because it would wreck they if others found out, it’s unlikely?
Which is more likely? That past behavior is an indication of future behavior, or that they because they could be eliminated from being found out it’s unlikely they’d do that thing. (By the way it’s also likely they’d are eliminated if they dont train their data with every advantage over their competitors possible). So I think it’s naive to think the incentives reward not doing the malicious thing now.
I would think they are not but Alex Karp CEO of Palantir seems to imply that they are:
https://youtu.be/0A3sGymV6kY?si=ti7uSZtYqJ3vKpGM
I found it a little shocking TBH
Alex Karp says a lot of things
Are we talking about the company sending back private information through its client to « fight » model distillation?
Yes.
Enterprise contracts are checked and agreed by lawyers. The contract states no training.
If the provider fucks up, there are actual monetary damages defined for breach of contract.
It's an unenforceable clause. The affected party has no means to prove that a breach has happened.
They could use an agent to summarise the source material, and then train models on those summaries, and claim that some sort of clean-room training has happened?
You've been to too many meetings with PMs and directors saying "An agent could very easily do this"
> if you use the coding plans they train in return.
No, you have to opt-in to that. There's a privacy toggle on account settings.
For OpenAI, you have to use the Enterprise plan at API pricing in order for them not to train on your data.
Source: https://chatgpt.com/codex/pricing/?type=team
Yes, my comment only applies to Claude - I have no knowledge of OpenAI policy.
I think the risk of not doing is more existential than doing so and getting caught. Wouldn’t you agree?
Edit: And the point of the poster is they have already demonstrated a track record of lying and misconduct, so how can you trust their word now? What have they done to show you they have taken responsibility for past actions and changed?
lmao, wasn't xAI caught doing this recently? moreover at least moonshot is being honest about it.
they train on your requests by paraphrasing them (which means rewriting them but keeping all the saliency) and removing their association with you
i don't know why this is so controversial, their terms are written to perfectly fit this training regime
if you are using bedrock, until very recently, they didn't see your requests and could not paraphrase. but too many people were using bedrock for too much stuff they wanted to see. so that's why the terms for bedrock changed for fable 5. this was the core of the palantir / defense dept drama with anthropic.
Anthropic constantly uses dark patterns to steal training data from customers (like the “how is claude doing” spam, data retention loosening when the safeguards false positive, etc).
How is that a dark pattern? What is the light pattern for getting feedback from users?
There are multiple ways to use feedback. Personally I would often be fine with actual human reading my feedback, taking in account the points I made and evaluating how it should affect their feature development and future roadmap.
Now what I would expect AI companies to do is to take things which were submitted as feedback and pretty much adding to training:
"Do more of this: <copy of the whole response which was flagged as good in feedback>"
"Do less of this: <copy of the whole response which was flagged as bad in feedback>"
It's paraphrased, but the point is that they will most likely use it more-or-less as-is and thus whatever is in there will be part of the model's training set rather than someone picking up the parts from response that are important and only including them (which happens with traditional feedback).
2 replies →