Is this really true? I was led to believe my company had an enterprise zero data retention agreement with them and it’s why we didn’t get access to Fable

Is there proof of what you’re saying or is it just a guess?

Read the terms of the ZDR policy with a critical eye. You’ll find that Anthropic retains almost arbitrary rights to retain anything it wants.

https://code.claude.com/docs/en/zero-data-retention

AFAIK there’s no ZDR with Claude models accessed directly via Anthropic. You’d have to go through either Google Vertex, Azure or AWS for true ZDR (at least legally/on paper).

Yeah. I may be naive, but I do trust the major cloud infra providers to offer real ZDR. Though admittedly, I haven't read their terms so it's possible that they also contain egregious loopholes.

https://trustedrouter.com/models/moonshotai/kimi-k3 is a good option if you want that to actually be the case.

oh, I've no doubt the US government and giga corporations can get zero data retention without ten pages of fine print. the rest of us can't.

Unless you spend 5min googling and see that you can do zero retention via AWS Bedrock.

Yeah even the chatgpt teams subscription claims ZDR. I believe the business plan from anthropic does too.

Of course maybe there is some fine print I haven’t read, and obviously I get the point that it may not be trustworthy.

edit: whoops I just checked and the “business”/“teams” plans just agree not to use your data for training

AWS Bedrock has stated that all new Anthropic and OpenAI models will not allow ZDR (only older models).

> zero data retention

Zero data retention is also "trust me dude".

There is no viable way of checking they are actually doing that.

That's assuming they don't put carve-out clauses in, like Anthropic did with Fable, which means data retention is back on the cards, no exceptions.

Also don't forget a zero data retention clause is still subject to the good old "law, or court or administrative order" contract clauses. :)

To get properly close to real zero-retention in a hosted model, you would have to use one of the verifiably private AI that runs in enclaves, e.g. Tinfoil (US) or Privatemode (Germany)[2]. Yes, still not the same as running on your own hardware, but a million lightyears ahead of "zero data retention" "trust me dude" clauses.

[1]https://tinfoil.sh/ [2]https://www.privatemode.ai/

No I know of course, I don’t trust them as far as I can throw them when all of these companies committed the largest copyright theft in human history to build the models.

I just wanted to know if that other person had proof or not, and I guess they didn’t. I would still rather have some semblance of an agreement than not have one at all — if you’re coding on a consumer plan you should just 100% assume anything you write with it will end up in the training set

In context it seems your recommendation is to instead send those data to models within Chinese nation-network space. I’m not here to defend US frontier model companies; your accusation is probably accurate. But I doubt sending data to China is an improvement.

with open weight models, you have three other options

A) use a provider that pinky-swears not to store your data. they obviously don't give a fuck about 'distillation attacks', so they have little motivation to voluntarily monitor and store your queries. reasonably high likelihood of privacy.

B) rent the hardware and run the model yourself. very high likelihood of privacy.

C) buy the hardware and run the model yourself. absolute certainty of privacy.