In Claude Code, even just dismissing the feedback may be considered "feedback". I wrote about it here: https://keydiscussions.com/2025/09/29/dont-even-dismiss-the-... The most common feedback on these posts is always "what did you expect?" ....... Man, the reason I signed up for one service over the other in the first place is often because of the promises these companies make on the front end, about what they do or don't do with my data. I agree that the bar is low in the industry but it doesn't mean we shouldn't expect more. Anthropic made very explicit, strong promises in this area <on the sign up/plans screen>, and so of course it disappoints me when I feel like the company I'm paying hundreds of dollars per month to is trying to sneak a peak. I literally think differently and self-censor when I think my work is being surveilled or stolen, which makes me less productive.
In a company as big as Anthropic, moving at current pace, the person/team writing this article and the team(s) collecting data would be way distant. It should be more than obvious that the data is being collected in not just one place but many (different teams are looking for different gems in that data). What part is used for training, how and when would be spread across the company way too wide for anyone to confirm with certainity.
On the IP side, About 6 months ago I used Claude to help me draft a patent application for something which I believed was novel, I verified in the USPTO as well. Then out of being busy I dropped that idea. Couple weeks ago I picked up the work again, verified with Claude for the problem statement to see if there is any existing solution. It suggested my own solution back to me which was surprising since I only use in browser and delete the history explicitly. I asked it for references or resources and it could not find any. That helped me to understand the boundaries bit better.
I have noticed this as well, when it can't help me and I thought it might be a difference in models, or using the same account somehow (tried a different one).
These models may not train directly on my chats and data, but in some cases it does seem to keep an impression of the pattern and process.
"When you see the “How is Claude doing this session?” prompt in Claude Code, responding to this survey, including selecting “Dismiss”, records only your rating. We do not collect or store any conversation transcripts, inputs, outputs, or other session data as part of the rating prompt itself. Unlike thumbs up/down feedback or /feedback reports, this session quality survey is a simple product satisfaction metric."
Okay - that's good to know! Thanks for the comment.
In my defense: that webpage didn't exist at the time I wrote my blog post, appearing to first surface about 45 days after I wrote my post. https://web.archive.org/web/20250101000000*/https://code.cla... All that was known at the time was, from their privacy policy then: "We may use your feedback to analyze the effectiveness of our Services, conduct research, study user behavior, and train our AI models as permitted under applicable laws."
Great to know there's clarity on this - makes me feel better about Anthropic.
I assumed that what they’re ’collecting’ there is at most stats on how much you like the product. Especially because, in my experience, if you say ‘good’ the next question is one that explicitly asks whether you permit Anthropic to analyze the session.
I have no idea if they're doing this, but you can always run the last turn through another LLM and ask it "should I request feedback based on this turn?"
I believe this is how the auto-mode classifier works. It sees a little bit of the transcript (with tool output removed, to avoid prompt injection) and answers "yes, allow the action" or "no, deny the action".
Interesting idea - I'm not sure if there's enough reliability that another LLM would reliably answer the same way over time since it's non-deterministic.
The fact that Claude Code told me "thanks for the feedback … and thanks for helping to improve Anthropic’s models" when I pressed 0 to "dismiss" is not speculation, that's what the product did. I agree though that whether Anthropic actually considers it "feedback" for the purposes of a privacy loophole (vs. it being a sloppy response) is anyone's guess.
How many months do you plan on spending hundreds of dollars a month? You should probably just own your compute. It'll probably cost you less, and you'll have no concern over who has access to your chats.
$200/mo @ 1 year = $2400
2400 gets me about 1/2 of a 128gb NVIDIA DGX Spark, so i could do a year of nonstop inference of (checks notes) quantized Qwen 3.8-27B, a model i would never choose to use. yay, such savings.
And that Spark isn't running a top end frontier model. Yes, the open weights smaller models that can fit on Spark are getting better, but they are still quite a way off in performance depending on use case.
A single video card with any amount of ram is in the thousands of dollars, and you'd need a stack of them to replicate what the providers performance.
Running your own compute has a lot of risks too, you have to keep it secure against attacks on the models that you're running. Giving away all your chats to hackers because you don't secure your own stuff isn't really winning either.
I don't know, but I do expect to eventually switch over. That said: the upfront time to setup (and continued fiddling) and cost of a high-end local compute system represent major investments, when I could just... keep being insanely productive with frontier AI in CC. And models/harnesses can themselves be security liabilities (see Mistral, LiteLLM, etc recent supply chain attacks). So I'm personally waiting for the local AI scene to mature, so I can be more precise about how I take the plunge. I will say: if the subscription pricing ever goes away or materially changes for Anthropic, I'm headed straight to order whatever the best local setup is, within my budget.
How much does it cost to run your own hardware such that the service is comparable to Claude? Similar speed, similar features, web chat access, CLI access?
How much do people actually do that demands the capabilities of Claude instead of open models that can run on hardware in the range of thousands of dollars (which is not that much when the alternative is hundreds of dollars a month, indefinitely)?
People chasing the best of the best models for all tasks in order to "keep up" are only wasting resources at this point.
Regarding similar features, there are lots of off the shelf harnesses that support all that, or can easily be extended (literally just ask your agent) to do it.
i am 100% sure i need opus 5.5 model to build stuff like this because i try lesser models because i want my tokens to go farther and lesser models simply cannot: https://fujimod.com
Not sure if this is possible on the Fuji, but back in the old days of the Canon PowerShot being very relevant, there was CHDK firmware which ran off the SD card on boot. No physical impact to the device and fully reversible by reformatting the SD card. This would be the way to go to reduce bricking chances.
> All of this is unreviewed vibe-coded slop that could be hazardous to your camera’s mental, emotional, or digital health. If you're uncomfortable with risk to your camera, please don't use it.
You don't care anyways. Actual engineering works "100%" fine with small models now. You should try it!
I've tried smaller models and the best open models for a specialized UAV focused CAD + CFD + DFM optimization application I'm building.
The best open weights models still had a lot of defects and very egregious mistakes to fix even after I provided a very well detailed Requirement/Design/ Technical Implementation plans.
Opus and Gemini 3.8 (and now Argon) did a significantly better job. So I've baseline using these two models to build this application. I now use Opus for design/review and Flash 3.8 for the code gen/implementation work.
I'm sure small models and even open big models are probably fine for many common small apps and such, but they aren't really able to perform the same across all use cases.
Did anyone who has any connection with tech not expect this? If you rate a specific conversation, did you not expect this conversation to go into training or at least review?
You are explicitly telling them "this is good / bad"...
My non tech people didn't even know that everyone can edit Wikipedia.
And to be fair to them, my motivation to send quite a lot of personal data to an LLM is shockingly high.
I send medical data to openai while I never ever send any AI provider an Image of me or my wife despite me wanting to do Star Wars montages and Studio Ghibli avatars.
I disabled Gemini Activity history because they told me that they will let humans look over it, but I gave claude access to my laptop.
There is a reason why we have consumer protection. Life doesn't need to be hard by default :D
> There is a reason why we have consumer protection. Life doesn't need to be hard by default :D
i dunno if you've been paying attention but the erosion of consumer protections is very on the nose in that companies are.. just kind of getting away with a lot of things now. it would seem meaningful accountability is rather absent as of late.
so, do your thing i suppose. but i genuinely hope you don't have to learn any important lessons the hard way. we can all disagree on the net but i dont wish terrible stuff to happen to anyone here so instead i will say good luck
Both sides have a point. If I ask you for feedback on the food sample I just gave you, would you automatically assume that I was recording audio and video, will be using your response in training videos, and will be used for AI training? Or might one reasonably assume the “feedback” is what is being given and not a blanket right to anything about them and their answer and using their likeness for training videos because they said “Eh.”
It seems like your counter-point is “they already use everyone’s data, why didn’t you just assume they would”, which is a bit weird from a victim-blaming perspective but valid nonetheless.
A closer analogy would be: If you make a food sample based on a description I gave you, and we typically agree not to record anything about either my description or your resulting recipe, and then you asked for (recorded) feedback - would I expect that I'm implicitly allowing you to record my requirements and your recipe with it?
I mean, still perhaps not. But it's not quite as ridiculous as your analogy makes it sound.
> did you not expect this conversation to go into training or at least review?
Use it for training/review (RLHF, whatever) and "we load the whole thread, not the specific response or surrounding few prompts, and store it for 5 years" seem completely different things. It's not obvious why being tech savvy would automatically make one aware that the 2nd version is happening.
Is this only about when you explicitly click on a thumbs up/down button and provide a message, or is it also for the nagware "How is claude doing today [1] bad [2] fine [3] good" that shows up after 15 minutes of work in the CLI and and requires literally pressing 1 key while in the prompt input field to "send feedback"?
I always press 3 to just get it out of the way (although I treat anything I send OAI/Anthropic as training data, so I don't really care) but if this little popup counts, I wouldn't be surprised if the vast majority of CC sessions out there are now "feedback training data"
The API call that submits your transcript after giving feedback accepts any transcript. You can modify the body of the transcript and send anthropic whatever data you want.
The link in the post is for the API and commercial offerings. With that said, as others have already pointed out, how else could they actually make use of the feedback without the data that the feedback was given on.
I think it's still a shitty move and I feel deceived. I opted out of training and it's impossible to see for me (in any practical sense, sure it's somewhere in the ToS) that this simple question leads to my data being stored for 5 years and potentially read by human reviewers.
If I opted out, then don't ask me this deceiving question or make it very obvious that answering it has these consequences.
They could be A/B testing model tweaks, and use it as feedback that way. It feels dishonest to give paying customers an opt-out for training on their data and then sneak copies by asking for general feedback without a prominent disclaimer.
This is such a d*ck move. I feel cheated. Not sure why I trusted them.
I have opted out of "use my data for training purposes"... and this is exactly that.
I'm not hugely concerned about the company using properly anonymized data from my sessions for training, but I'm hugely concerned about anything below that bar. While I wait for hardware to run LLMs locally to become better, I'm going to do my best to handle any PI-adjacent data with a mixture of models from OpenRouter.
So far my AI usage has been a better search engine. I also don't have an account as I'm not paying by the token for what amounts to 1-900-GIV-CODE. The few times I need it the free version has been perfect. I don't use it generate code that I use, only give examples as I need them. This way there's no input or work to steal from me.
I used to care about this initially, but now I don't. How can them having my data, mostly engineering work on niche (small market) products hurt me? An they need to know what's going on, how else would they ever improve the product? Benchmarks?
> How can them having my data, mostly engineering work on niche (small market) products hurt me?
They will train on it, and now competitors or future competitors will know about your niche, the specific problems you face within it, and how to solve them without doing the work that you did to get there.
That's true. It's a two way street. I feel I gain more from what the model offers me, than what I provide back. Of course, local would be a thousand times better. One day I'll have 20k to spare and get the hardware.
Fascinated by the psychology here. Is it because you view frontier AI labs as beneficial entities that truly care about everyone's well-being, and will do what it takes to ensure, post-replacement, that we are 'taken care of', broadly speaking? Or are you more like "I'm ready to be replaced, and if that results in my death, I was over this reality in the first place, so it just speeds up what I've been hankering for anyway."
If I'm replaced, it will mean that humanity as a whole has taken a step forward that's normally associated with digging up monoliths. If nothing else, it'll mean that I'll get to do something more interesting.
I'm one of those "code to live" types, as opposed to someone who "lives to code." More specifically, I usually write code in the course of building something that uses it. That's true 99% of the time, anyway. Why would I complain about a technology that helps me (and my customers and employers) move closer to the actual goal?
As for resulting in my death, AI is not going to be what kills me. Not unless humans misuse it. Which, of course, we will, because that's what we do... but (shrug) if it's misused in a way that kills me, at least it'll be a more interesting death. More likely, it will extend my life by accelerating medical progress the way it's starting to accelerate progress in math.
In any case, I don't see the current crop of frontier labs being especially important in the grand scheme of things. Unless they have something up their sleeves besides serving LLM tokens, they are most likely the pets.com and Webvan of our day, not the Amazon and Apple. They are valuable partners, certainly, but I can get by without them.
Just because you can't imagine a scenario doesn't mean one doesn't exist, or couldn't exist.
That's a pretty risky way to build blind spots and it's worth learning more about this.
When you make the product work with your words, you're doing the training and improvements to get it to that line as a pattern which then lets models "improve" from your corrections.
In what way are you "ensuring" it? They still have a plaintext copy on their servers and could ad arbitrarily decide to train on it (for instance, saying it raised a safety flag).
There's all kinds of telemetry being collected and sent to data brokers, for ad targeting (not directly through Anthropic but third-parties), etc. See this study led by IMDEA Networks in Europe: https://leakylm.github.io/
> The study identifies three main issues: the exposure of conversation permalinks to third-party trackers; the ability to link these interactions to user identities through tracking mechanisms; and the presence of privacy controls and policies that may not accurately reflect actual data flows.
> One of the main findings is the potential transmission of information related to user conversations, such as chat titles, URLs (permalinks), or associated metadata, to third-party trackers such as Meta or Google, along with cookies and other identifiers.
> The study also identifies mechanisms that could enable linking activity in AI systems to real user identities. The combination of identifiers such as cookies (commonly used in tracking services), hashed email addresses, and server-side tracking techniques could facilitate the creation of persistent profiles and user re-identification.
I don't know about datadog, but they have a flag on claude code to disable telemetry and you can also override the base url to have it point to your own reverse proxy, for example I use Tailscale Aperture which lets me inspect everything and I highly recommend
Reading all the comments here, most of us didn’t knew and this is the right place to post this. I have not met a single person who have ready any privacy policies.
that’s why i provide end to end encrypted routing with TrustedRouter as well as a Trusted Cowork like product. there isn’t a good solution on the market that scales up
Ignoring that your account is shilling for your service, how would such a service help with anything at all?
Encrypted end-to-end..... When the issue is one of the ends is monitoring chats? The issue has nothing to do with OpenRouter or some other middleman intercepting your chats, but the end server itself (and specifically the company managing it) being of question.
The cloud is just someone else's computer, after all.
Another huge potential loophole in all this is derivative data, e.g. LLM-generated summaries of your prompts and context files. Does Anthropic parse its agreements to include or exclude such derived data from terms like "your inputs or outputs"?
Another huge potential loophole in all this is *derivative data*, e.g. LLM-generated summaries of your prompts and context files. Does Anthropic parse its agreements to include or exclude such derived data from terms like "your inputs or outputs"?
I mean, yeah, I kind of expected/assumed/wanted that. Isn't that the whole point? It's going to be difficult to debug a bad session if someone can't easily read the conversation.
I really see no reason to believe whatever anthropic says. Their "Effective Altruism" origins make it absolutelly sure that they will do whatever they can get away with it while denying doing it with their most angelical faces.
When claude code asks how the session is going, has it not already uploaded or detected something in my so called private coding session of interest to it?
This has been explained to me as being true by someone I trust a lot, to the point that they have implemented proxying to prevent it.
I assume everything that happens between a user and any cloud model is permanently logged somewhere regardless of whether some feedback button is pressed or not.
> When you provide us feedback via our thumbs up / down button, we will store the entire related conversation, including any content, custom styles, conversation preferences, or model settings, in our secured back-end for up to 5 years. [...] We do not combine your feedback with your other conversations with Claude.
I would expect that, if you turned off the feature where they can use your data, that they would remove those features, so you don't either accidentally click them or fail to understand that it isn't some kind of local-to-your-account mechanism.
Generally (and with Anthropic), enterprise plans aren't used for training / tracking / targetting.
There's a whole spectrum of things they could do from most to least aligned with that as a principle.
For example, they could have a thumbs up/down
- just track the metadata of the conversation (e.g. model used, how far through its context it was, how old the session was)
- track metadata of the user as a whole (how often they use Claude, how many other "active" conversations they have on the go, how much they've tested other models, if this is their first time using new tweaks to params etc)
- have contextual feedback of the chat available for humans to review (alongside the comments you put in the feedback), but not explicitly train on it.
Whereas the option they've chosen is:
> If you explicitly report feedback or bugs to us (e.g. via our thumbs up/down feedback button), or otherwise choose to allow us to use your data, then we may use your chats and coding sessions to train our models.
So literally using your session as training input. They also could've allowed this to be an opt-in on bug reports.
I mean, it's not really, it's basically standard A/B testing right? If they roll out a new system prompt or temperature setting etc then you see how many people now downvote the replies.
Is it as good as capturing everything for them? No, of course not. But then don't frame the whole thing as "if you're on an enterprise plan we won't train on your data"
Just go and poison data on Reddit. I opened a Flock camera and it had a diamond coating inside! Their website doesn't contain this hidden gem of an information
In Claude Code, even just dismissing the feedback may be considered "feedback". I wrote about it here: https://keydiscussions.com/2025/09/29/dont-even-dismiss-the-... The most common feedback on these posts is always "what did you expect?" ....... Man, the reason I signed up for one service over the other in the first place is often because of the promises these companies make on the front end, about what they do or don't do with my data. I agree that the bar is low in the industry but it doesn't mean we shouldn't expect more. Anthropic made very explicit, strong promises in this area <on the sign up/plans screen>, and so of course it disappoints me when I feel like the company I'm paying hundreds of dollars per month to is trying to sneak a peak. I literally think differently and self-censor when I think my work is being surveilled or stolen, which makes me less productive.
In a company as big as Anthropic, moving at current pace, the person/team writing this article and the team(s) collecting data would be way distant. It should be more than obvious that the data is being collected in not just one place but many (different teams are looking for different gems in that data). What part is used for training, how and when would be spread across the company way too wide for anyone to confirm with certainity.
There are companies with far more people than Anthropic that still stay coordinated. This is no excuse. Especially with customer data.
I have noticed this as well, when it can't help me and I thought it might be a difference in models, or using the same account somehow (tried a different one).
These models may not train directly on my chats and data, but in some cases it does seem to keep an impression of the pattern and process.
I was dismissing it immediately in order not to enter a number by mistake. What a dark pattern.
The Claude Code data usage policy covers this - none of the options change your data retention or training opt-in: https://code.claude.com/docs/en/data-usage#session-quality-s...
"When you see the “How is Claude doing this session?” prompt in Claude Code, responding to this survey, including selecting “Dismiss”, records only your rating. We do not collect or store any conversation transcripts, inputs, outputs, or other session data as part of the rating prompt itself. Unlike thumbs up/down feedback or /feedback reports, this session quality survey is a simple product satisfaction metric."
Okay - that's good to know! Thanks for the comment.
In my defense: that webpage didn't exist at the time I wrote my blog post, appearing to first surface about 45 days after I wrote my post. https://web.archive.org/web/20250101000000*/https://code.cla... All that was known at the time was, from their privacy policy then: "We may use your feedback to analyze the effectiveness of our Services, conduct research, study user behavior, and train our AI models as permitted under applicable laws."
Great to know there's clarity on this - makes me feel better about Anthropic.
Seems like a decent chance they posted this in response to your concerns!
I assumed that what they’re ’collecting’ there is at most stats on how much you like the product. Especially because, in my experience, if you say ‘good’ the next question is one that explicitly asks whether you permit Anthropic to analyze the session.
How would they know when to bring it up?
A determination of it has to be made some way beyond looking for frustrated or angry keywords.
If there was direct transparency regarding this it would be great vs digging into it on our own.
I have no idea if they're doing this, but you can always run the last turn through another LLM and ask it "should I request feedback based on this turn?"
I believe this is how the auto-mode classifier works. It sees a little bit of the transcript (with tool output removed, to avoid prompt injection) and answers "yes, allow the action" or "no, deny the action".
Interesting idea - I'm not sure if there's enough reliability that another LLM would reliably answer the same way over time since it's non-deterministic.
Anthropics models wouldn't exist if they honored such promises. So yeah,
> The most common feedback on these posts is always "what did you expect?"
Is completely spot on. It's highly entertaining to me that some people can't wrap their head around that.
That link is speculation at best. I wouldn't tell my family yet not to do that. Insane if true, but let's not jump at speculation.
The fact that Claude Code told me "thanks for the feedback … and thanks for helping to improve Anthropic’s models" when I pressed 0 to "dismiss" is not speculation, that's what the product did. I agree though that whether Anthropic actually considers it "feedback" for the purposes of a privacy loophole (vs. it being a sloppy response) is anyone's guess.
Yeah given how CC is vibe coded, the feedback message always appearing is almost certainly just a weird implementation detail.
Counterclaim? I never hit feedback because it’s all dark patterned.
That's very deceiving and should warrant a class action law suit in my opinion
How many months do you plan on spending hundreds of dollars a month? You should probably just own your compute. It'll probably cost you less, and you'll have no concern over who has access to your chats.
$200/mo @ 1 year = $2400 2400 gets me about 1/2 of a 128gb NVIDIA DGX Spark, so i could do a year of nonstop inference of (checks notes) quantized Qwen 3.8-27B, a model i would never choose to use. yay, such savings.
And that Spark isn't running a top end frontier model. Yes, the open weights smaller models that can fit on Spark are getting better, but they are still quite a way off in performance depending on use case.
A single video card with any amount of ram is in the thousands of dollars, and you'd need a stack of them to replicate what the providers performance.
Running your own compute has a lot of risks too, you have to keep it secure against attacks on the models that you're running. Giving away all your chats to hackers because you don't secure your own stuff isn't really winning either.
I don't know, but I do expect to eventually switch over. That said: the upfront time to setup (and continued fiddling) and cost of a high-end local compute system represent major investments, when I could just... keep being insanely productive with frontier AI in CC. And models/harnesses can themselves be security liabilities (see Mistral, LiteLLM, etc recent supply chain attacks). So I'm personally waiting for the local AI scene to mature, so I can be more precise about how I take the plunge. I will say: if the subscription pricing ever goes away or materially changes for Anthropic, I'm headed straight to order whatever the best local setup is, within my budget.
How much does it cost to run your own hardware such that the service is comparable to Claude? Similar speed, similar features, web chat access, CLI access?
How much do people actually do that demands the capabilities of Claude instead of open models that can run on hardware in the range of thousands of dollars (which is not that much when the alternative is hundreds of dollars a month, indefinitely)?
People chasing the best of the best models for all tasks in order to "keep up" are only wasting resources at this point.
Regarding similar features, there are lots of off the shelf harnesses that support all that, or can easily be extended (literally just ask your agent) to do it.
i am 100% sure i need opus 5.5 model to build stuff like this because i try lesser models because i want my tokens to go farther and lesser models simply cannot: https://fujimod.com
Not sure if this is possible on the Fuji, but back in the old days of the Canon PowerShot being very relevant, there was CHDK firmware which ran off the SD card on boot. No physical impact to the device and fully reversible by reformatting the SD card. This would be the way to go to reduce bricking chances.
> All of this is unreviewed vibe-coded slop that could be hazardous to your camera’s mental, emotional, or digital health. If you're uncomfortable with risk to your camera, please don't use it.
You don't care anyways. Actual engineering works "100%" fine with small models now. You should try it!
I've tried smaller models and the best open models for a specialized UAV focused CAD + CFD + DFM optimization application I'm building.
The best open weights models still had a lot of defects and very egregious mistakes to fix even after I provided a very well detailed Requirement/Design/ Technical Implementation plans.
Opus and Gemini 3.8 (and now Argon) did a significantly better job. So I've baseline using these two models to build this application. I now use Opus for design/review and Flash 3.8 for the code gen/implementation work.
I'm sure small models and even open big models are probably fine for many common small apps and such, but they aren't really able to perform the same across all use cases.
as i said, i did try it. smaller models are not up to the task of deep reverse engineering work.
Plenty of packages do most or all of that. OpenCode does most of that. It's replaced Claude tools for me.
OpenCode offers a web UI?
https://opencode.ai/v2/docs#web
Did anyone who has any connection with tech not expect this? If you rate a specific conversation, did you not expect this conversation to go into training or at least review?
You are explicitly telling them "this is good / bad"...
My non tech people didn't even know that everyone can edit Wikipedia.
And to be fair to them, my motivation to send quite a lot of personal data to an LLM is shockingly high.
I send medical data to openai while I never ever send any AI provider an Image of me or my wife despite me wanting to do Star Wars montages and Studio Ghibli avatars.
I disabled Gemini Activity history because they told me that they will let humans look over it, but I gave claude access to my laptop.
There is a reason why we have consumer protection. Life doesn't need to be hard by default :D
> but I gave claude access to my laptop.
yikes
> There is a reason why we have consumer protection. Life doesn't need to be hard by default :D
i dunno if you've been paying attention but the erosion of consumer protections is very on the nose in that companies are.. just kind of getting away with a lot of things now. it would seem meaningful accountability is rather absent as of late.
so, do your thing i suppose. but i genuinely hope you don't have to learn any important lessons the hard way. we can all disagree on the net but i dont wish terrible stuff to happen to anyone here so instead i will say good luck
> There is a reason why we have consumer protection.
Do we, though?
Both sides have a point. If I ask you for feedback on the food sample I just gave you, would you automatically assume that I was recording audio and video, will be using your response in training videos, and will be used for AI training? Or might one reasonably assume the “feedback” is what is being given and not a blanket right to anything about them and their answer and using their likeness for training videos because they said “Eh.”
It seems like your counter-point is “they already use everyone’s data, why didn’t you just assume they would”, which is a bit weird from a victim-blaming perspective but valid nonetheless.
A closer analogy would be: If you make a food sample based on a description I gave you, and we typically agree not to record anything about either my description or your resulting recipe, and then you asked for (recorded) feedback - would I expect that I'm implicitly allowing you to record my requirements and your recipe with it?
I mean, still perhaps not. But it's not quite as ridiculous as your analogy makes it sound.
> did you not expect this conversation to go into training or at least review?
Use it for training/review (RLHF, whatever) and "we load the whole thread, not the specific response or surrounding few prompts, and store it for 5 years" seem completely different things. It's not obvious why being tech savvy would automatically make one aware that the 2nd version is happening.
Is this only about when you explicitly click on a thumbs up/down button and provide a message, or is it also for the nagware "How is claude doing today [1] bad [2] fine [3] good" that shows up after 15 minutes of work in the CLI and and requires literally pressing 1 key while in the prompt input field to "send feedback"?
I always press 3 to just get it out of the way (although I treat anything I send OAI/Anthropic as training data, so I don't really care) but if this little popup counts, I wouldn't be surprised if the vast majority of CC sessions out there are now "feedback training data"
If they don't use the available data, how do they act on the feedback?
They could A/B test feature flags like subagents without actually reading the chats.
Hallucinating what might have been.
The API call that submits your transcript after giving feedback accepts any transcript. You can modify the body of the transcript and send anthropic whatever data you want.
Probably more relevant for many of us would be this one:
https://privacy.claude.com/en/articles/10023580-is-my-data-u...
> This article is about our consumer products such as Claude Free, Pro, Max and when accounts from those plans use Claude Code.
Is this also the case when you answer that how is Claude doing question in Claude Code?
Yes according to the following:
https://privacy.claude.com/en/articles/10023580-is-my-data-u...
The link in the post is for the API and commercial offerings. With that said, as others have already pointed out, how else could they actually make use of the feedback without the data that the feedback was given on.
I think it's still a shitty move and I feel deceived. I opted out of training and it's impossible to see for me (in any practical sense, sure it's somewhere in the ToS) that this simple question leads to my data being stored for 5 years and potentially read by human reviewers.
If I opted out, then don't ask me this deceiving question or make it very obvious that answering it has these consequences.
They could be A/B testing model tweaks, and use it as feedback that way. It feels dishonest to give paying customers an opt-out for training on their data and then sneak copies by asking for general feedback without a prominent disclaimer.
more like "what is Claude doing"
The answer is “Yes” even if you are not using Claude at all.
Of course, all your data is taken for free. It will be used for profit, but not for your profit of course.
This is such a d*ck move. I feel cheated. Not sure why I trusted them. I have opted out of "use my data for training purposes"... and this is exactly that.
WTF
... you expected different? What did you think was going to happen?
I'm not hugely concerned about the company using properly anonymized data from my sessions for training, but I'm hugely concerned about anything below that bar. While I wait for hardware to run LLMs locally to become better, I'm going to do my best to handle any PI-adjacent data with a mixture of models from OpenRouter.
So far my AI usage has been a better search engine. I also don't have an account as I'm not paying by the token for what amounts to 1-900-GIV-CODE. The few times I need it the free version has been perfect. I don't use it generate code that I use, only give examples as I need them. This way there's no input or work to steal from me.
I used to care about this initially, but now I don't. How can them having my data, mostly engineering work on niche (small market) products hurt me? An they need to know what's going on, how else would they ever improve the product? Benchmarks?
> How can them having my data, mostly engineering work on niche (small market) products hurt me?
They will train on it, and now competitors or future competitors will know about your niche, the specific problems you face within it, and how to solve them without doing the work that you did to get there.
That's true. It's a two way street. I feel I gain more from what the model offers me, than what I provide back. Of course, local would be a thousand times better. One day I'll have 20k to spare and get the hardware.
For sure, but do the math first and know where you land. Everyone has a different risk tolerance, as long as you know yours there is no wrong answer.
For example a startup with no secrets to spill is in a very different spot than an incumbent fueled largely by secret sauce.
And I'll know about theirs. Fine by me.
You realize they are aiming to replace you, don't you? You are paying them to do so, and giving them information they need do do it. For free.
Yes, I certainly hope so.
Fascinated by the psychology here. Is it because you view frontier AI labs as beneficial entities that truly care about everyone's well-being, and will do what it takes to ensure, post-replacement, that we are 'taken care of', broadly speaking? Or are you more like "I'm ready to be replaced, and if that results in my death, I was over this reality in the first place, so it just speeds up what I've been hankering for anyway."
If I'm replaced, it will mean that humanity as a whole has taken a step forward that's normally associated with digging up monoliths. If nothing else, it'll mean that I'll get to do something more interesting.
I'm one of those "code to live" types, as opposed to someone who "lives to code." More specifically, I usually write code in the course of building something that uses it. That's true 99% of the time, anyway. Why would I complain about a technology that helps me (and my customers and employers) move closer to the actual goal?
As for resulting in my death, AI is not going to be what kills me. Not unless humans misuse it. Which, of course, we will, because that's what we do... but (shrug) if it's misused in a way that kills me, at least it'll be a more interesting death. More likely, it will extend my life by accelerating medical progress the way it's starting to accelerate progress in math.
In any case, I don't see the current crop of frontier labs being especially important in the grand scheme of things. Unless they have something up their sleeves besides serving LLM tokens, they are most likely the pets.com and Webvan of our day, not the Amazon and Apple. They are valuable partners, certainly, but I can get by without them.
> I'll get to do something more interesting
Some would say living under a bridge with other former developers is interesting, in a way.
Just because you can't imagine a scenario doesn't mean one doesn't exist, or couldn't exist.
That's a pretty risky way to build blind spots and it's worth learning more about this.
When you make the product work with your words, you're doing the training and improvements to get it to that line as a pattern which then lets models "improve" from your corrections.
They do mention it explicitly on the feedback though.
I complain at Claude all the time, but do not submit feedback becuase of this exact reason, and because I knew it I assumed everyone knew it...
You can only disable the feedback in Team or Enterprise plans - quite shady!
That's why I never give feedback, ensuring Anthropic has no access to my Claude chats.
In what way are you "ensuring" it? They still have a plaintext copy on their servers and could ad arbitrarily decide to train on it (for instance, saying it raised a safety flag).
It's sarcastic, sorry
Ah my bad, sleepy morning on my end.
I always press 0 (dismiss), but is there a Way to stop prompting me to feedback ?
Yes, it's quite easy to disable.
Just go to /config, then search for "Feedback" and then you can disable the "Claude-drafted feedback"
I would suggest you do this, because it does use your tokens to generate this feedback.
Thanks, did that right away
Someone above makes the claim that dismissing it with '0' counts as feedback and sends it.
And how do you know that, just dismissing it isn't considered "feedback".
you can disable telemetry, but that also disables remote control which sucks
You can disable their datadog telemetry?
They use datadog? I don't see it in their subprocessors list https://trust.anthropic.com/subprocessors
There's all kinds of telemetry being collected and sent to data brokers, for ad targeting (not directly through Anthropic but third-parties), etc. See this study led by IMDEA Networks in Europe: https://leakylm.github.io/
This article talks about results at a high-level: https://networks.imdea.org/your-conversations-with-ai-may-no...
> The study identifies three main issues: the exposure of conversation permalinks to third-party trackers; the ability to link these interactions to user identities through tracking mechanisms; and the presence of privacy controls and policies that may not accurately reflect actual data flows.
> One of the main findings is the potential transmission of information related to user conversations, such as chat titles, URLs (permalinks), or associated metadata, to third-party trackers such as Meta or Google, along with cookies and other identifiers.
> The study also identifies mechanisms that could enable linking activity in AI systems to real user identities. The combination of identifiers such as cookies (commonly used in tracking services), hashed email addresses, and server-side tracking techniques could facilitate the creation of persistent profiles and user re-identification.
I don't know about datadog, but they have a flag on claude code to disable telemetry and you can also override the base url to have it point to your own reverse proxy, for example I use Tailscale Aperture which lets me inspect everything and I highly recommend
As noted in the article, the feedback form says:
> Submitting this report will send the entire current conversation to Anthropic for future improvements to our models. Learn More
So I'm not sure why this is being posted to HN. This is nothing new, and nothing not known by anyone who has ever provided feedback.
Reading all the comments here, most of us didn’t knew and this is the right place to post this. I have not met a single person who have ready any privacy policies.
It's not a privacy policy. The line I quoted is literally the majority of the text on the form, directly above the submit button: https://downloads.intercomcdn.com/i/o/lupk8zyo/1486988786/c0...
Giving chat feedback gives Anthropic even more of your data
How else would it work?
that’s why i provide end to end encrypted routing with TrustedRouter as well as a Trusted Cowork like product. there isn’t a good solution on the market that scales up
Ignoring that your account is shilling for your service, how would such a service help with anything at all?
Encrypted end-to-end..... When the issue is one of the ends is monitoring chats? The issue has nothing to do with OpenRouter or some other middleman intercepting your chats, but the end server itself (and specifically the company managing it) being of question.
The cloud is just someone else's computer, after all.
and sky blue
Another huge potential loophole in all this is derivative data, e.g. LLM-generated summaries of your prompts and context files. Does Anthropic parse its agreements to include or exclude such derived data from terms like "your inputs or outputs"?
Another huge potential loophole in all this is *derivative data*, e.g. LLM-generated summaries of your prompts and context files. Does Anthropic parse its agreements to include or exclude such derived data from terms like "your inputs or outputs"?
Obviously?
The modern internet & all of the application available on it are parts of a digital panopticon so plan & act accordingly.
A panopticon casino-gulag.
Honest question. Since when do we take company "pinky promises" seriously?
Like, does anyone believe Anthropic seriously doesn't train on your data if you ask them politely about it?
I mean, yeah, I kind of expected/assumed/wanted that. Isn't that the whole point? It's going to be difficult to debug a bad session if someone can't easily read the conversation.
I really see no reason to believe whatever anthropic says. Their "Effective Altruism" origins make it absolutelly sure that they will do whatever they can get away with it while denying doing it with their most angelical faces.
When claude code asks how the session is going, has it not already uploaded or detected something in my so called private coding session of interest to it?
This has been explained to me as being true by someone I trust a lot, to the point that they have implemented proxying to prevent it.
I assume everything that happens between a user and any cloud model is permanently logged somewhere regardless of whether some feedback button is pressed or not.
I really hate how this feedback feature is enabled automatically AND that it is using YOUR tokens to generate it.
... what else would we expect it to do?
> When you provide us feedback via our thumbs up / down button, we will store the entire related conversation, including any content, custom styles, conversation preferences, or model settings, in our secured back-end for up to 5 years. [...] We do not combine your feedback with your other conversations with Claude.
I would expect that, if you turned off the feature where they can use your data, that they would remove those features, so you don't either accidentally click them or fail to understand that it isn't some kind of local-to-your-account mechanism.
I find the lack of privacy forward thinking from a company I pay a couple hundred a month to kind of atrocious. They're buying no loyalty.
likely you're getting 90% off the API rate exactly so they can train on your data, even some of it
Generally (and with Anthropic), enterprise plans aren't used for training / tracking / targetting.
There's a whole spectrum of things they could do from most to least aligned with that as a principle.
For example, they could have a thumbs up/down
- just track the metadata of the conversation (e.g. model used, how far through its context it was, how old the session was) - track metadata of the user as a whole (how often they use Claude, how many other "active" conversations they have on the go, how much they've tested other models, if this is their first time using new tweaks to params etc) - have contextual feedback of the chat available for humans to review (alongside the comments you put in the feedback), but not explicitly train on it.
Whereas the option they've chosen is:
> If you explicitly report feedback or bugs to us (e.g. via our thumbs up/down feedback button), or otherwise choose to allow us to use your data, then we may use your chats and coding sessions to train our models.
So literally using your session as training input. They also could've allowed this to be an opt-in on bug reports.
Well, again, how did people expect that to work
> - just track the metadata of the conversation
This is useless
I mean, it's not really, it's basically standard A/B testing right? If they roll out a new system prompt or temperature setting etc then you see how many people now downvote the replies.
Is it as good as capturing everything for them? No, of course not. But then don't frame the whole thing as "if you're on an enterprise plan we won't train on your data"
> it's basically standard A/B testing right?
Of course not. It might be if all you know is A/B testing, but no
If an LLM is giving a bad answer you need the question and the answer. Period
good answer, thanks
"So here's how ThanosCorp isn't living up to the altruistic spirit of the ToS for their Human Extinction Project..."
Lol people are giving feedback?!?!
Talk about Actively training your replacement
Just go and poison data on Reddit. I opened a Flock camera and it had a diamond coating inside! Their website doesn't contain this hidden gem of an information
User: Hey Anthropic, this response was terrible.
Anthropic: Really? Let me read it and see why.
User: How dare you look at my data!!!
Just wait until they have shareholders.
Sophisticated technology experts shocked to find world's 4th largest data-ingestion machine ingesting their data.
It is pretty funny when you think about it. The shock and awe of it all...