This policy applies across all providers. Here is the warning in Cursor: https://i.redd.it/7sfyker2ya6h1.png

Note that Anthropic has committed not to train models on logged data, so I don’t understand some of the concerns here. What exactly is your threat model? That Anthropic would train models contrary to their terms of service? That you trust them enough not to log your data prior to this, but not enough to trust their stated limits on how logged data will be used now?

Edit: I am partially convinced by some of the replies. However, it is worth noting that this change primarily affects Enterprise users. Data from consumer plans is already retained for 30 days. Source: https://privacy.claude.com/en/articles/10023548-how-long-do-...

> you trust them enough not to log your data prior to this, but not enough to trust their stated limits on how logged data will be used now

It doesn't really matter how much you happen trust another party. In the regulatory world it only matters what contracts they will sign that guarantee their compliance. We do have those with AWS, we don't with Anthropic. If Anthropic physically captures the data, they just moved themselves outside the boundary of parties who we can do business with. Unless they want to sign a contract and implement all the corresponding compliance measures. They are insane if they think that's a good deal for them to do all that right now in every jurisdiction where AWS operates, when AWS has already spent a decade building it up.

> Note that Anthropic has committed not to train models on logged data, so I don’t understand some of the concerns here. What exactly is your threat model? That

Like Meta had committed to respect your privacy. Replace the name of the company with any of the top 50 companies in the world and go back how many have hold their promises - or just doing fine when breaking the rules. There is no legislation in the U.S. that can bankrupt the company for violating this? So there are no guarantees.

Meta openly torrented books and nobody asked them to remove/destroy their AI models. Similarly, for Anthropic, it was just a business cost. They were allowed to keep the models. No real consequences for breaking the rules.

It adds another provider that you have to trust with your data. Previously the assumption is that AWS was securely handling your data and you may have the data on AWS to start with anyways. Now you have two providers handling your data which doubles your risk if you trust them equally. If you think AWS has more robust data controls than Anthropic then it more than doubles your risk.

You may also have data management requirements such as allowed storage and transit countries as well as various certifications and contracts that you now need to extend to the second data processor.

Basically if you are already using AWS just adding the AWS-only bedrock model is legally easy and doesn't really change your security posture. If you need to now also log your data to Anthropic it makes the choice much more complicated.

Both can be true simultaneously. Anthropic can probably be trusted not to train on our Fable sessions, but eroding ZDR as the industry standard still sets a dangerous precedent.

There's a parallel between data retention and general mass surveillance. Sure, both systems can be used for purely benign purposes, with appropriate safeguards in place. But history shows that surveillance systems are alarmingly easy to co-opt for nefarious means, and model providers do have a heck of an incentive to leverage retained data for internal means.

This is worth protesting, even if I believe this policy itself does not immediately compromise my privacy.

> Note that Anthropic has committed not to train models on logged data, so I don’t understand some of the concerns here. What exactly is your threat model? That Anthropic would train models contrary to their terms of service? That you trust them enough not to log your data prior to this, but not enough to trust their stated limits on how logged data will be used now?

It is a different thing when they say they don't store your data.

And when they say they store your data for 30 days and review it for "issues", it makes your "spider sense" tingle. Who and how will review it, what are the "issues" they are looking for, etc. It is to vague and they can keep it this "dangerous" model for themselves.

Someone has never dealt with HIPAA laws and it shows.

Once you start storing anything, whether credit card numbers or AI inputs, then there is possibility (if not in fact probability) that you'll be hacked and it will leak.

Given Anthropic's failure to secure their own source code, do you really trust them to secure yours?

> Note that Anthropic has committed not to train models on logged data, so I don’t understand some of the concerns here. What exactly is your threat model?

First of all, will they respect that promise in the future? Because, you know… they already received your data and by some legal quirks they are already required to store it for so many years. “What’s your threat model”, uhh, sending confidential information to a third party.

It’s okay if you do this with your own personal property. But if you are working on client projects, what, are you going to start shipping customer data under nda without consent? Good luck in the court.

We shipped software to governments and some big companies where this is a big no-no. Try to explain to your clients that during the development process some pieces were sent to Antrophic, and they might keep it for whatever reasons.

here's how they train on your data:

an inference request comes in

claude fable RESTful API service does the stuff, some backend systems run the prefill and batch decode, and your conversation is cached for 5 minutes in some prefix cache.

the request is also sent to claude paraphraser, which does almost exactly the same thing as the compactor and rewrites your conversation.

then they record the paraphrased conversation and train on that. it keeps the salient parts of the conversation, like whatever internal knowledge you have, and disposes of anything that could have been correlated with the earlier conversation, which is easy to do because verification is a string comparison.