Anthropic rolled out updates to Claude’s Consumer Terms and Privacy Policy that change how long your conversations stick around — but only if you opt in. Existing users had to make a choice on a new model-training setting to keep using Claude, and the choice they made determines whether their chats are kept for 30 days or for five years.
If you use Claude for anything personal — drafting a difficult email, talking through a health question, organizing your finances, summarizing a private document — the difference between those two numbers is worth understanding before you click through another “not now.”
What Actually Changed
A new setting controls both training and retention together. Anthropic now asks consumer users — Free, Pro, and Max plans — whether they want their chats and coding sessions used to help improve Claude. The setting isn’t just about training. Saying yes also extends how long Anthropic keeps that data, from 30 days to five years.
The default is the shorter window. If you don’t actively opt in, your data retention stays at the existing 30-day policy. Anthropic isn’t quietly switching everyone to five years — it’s offering a trade and waiting for you to take it.
It only applies to new or resumed chats. Conversations you had before the policy update aren’t retroactively pulled into the new retention window just because you later opt in. The setting governs activity going forward, not your historical chat archive.
Business accounts are unaffected. Claude Team, Enterprise, the Claude Developer Platform, and anything covered by Anthropic’s Commercial Terms sit outside this change entirely. This is a consumer-tier policy, aimed at the hundreds of millions of people using Claude’s free and personal-paid plans.
Why Anthropic Frames This as a Trade, Not a Default
Most AI companies that train on user conversations have made that the default setting, with an opt-out buried a few menus deep. Anthropic’s approach here is structurally different: the longer retention period is the cost of opting in, not something that happens to you passively.
Anthropic has also tied the request to a stated purpose beyond pure model improvement — strengthening safeguards against scams and abusive use of Claude. Whether that framing changes how you weigh the trade is a personal call, but it’s worth noting the company isn’t presenting this purely as “help us build better AI.” It’s also positioned as a safety mechanism, which is a different pitch than the one most platforms make when asking for your data.
That said, a stated purpose doesn’t change the mechanics. If you opt in, your conversations are retained for five years and can be used to improve Claude’s models — full stop. The safety framing explains Anthropic’s motivation; it doesn’t narrow what actually happens to your data once you’ve agreed.
What “Used to Improve Claude” Actually Covers
This is the part worth being precise about, because “improve the model” is a phrase that covers a wide range of activity, not a single narrow action.
Your conversations can be reviewed by people, not just processed by automated systems. Training pipelines typically involve human review at some stage — quality checks, labeling, edge-case analysis — meaning opting in isn’t only a question of “will a machine see this,” it’s also “could a person at Anthropic see this.”
The five-year window applies regardless of whether your specific conversation ends up directly shaping a model update. Not every retained chat gets used in every training run. But the retention period is the ceiling Anthropic is allowed to hold the data for, not a guarantee about what happens to any single conversation within that window.
Deleting a conversation doesn’t necessarily undo training that already happened. If a chat was used to inform a model update before you deleted it, the deletion removes the stored conversation going forward — it doesn’t reverse what the model already learned from it. This is a structural limitation of how machine learning training works, not something specific to Anthropic, but it’s a meaningful gap between “I deleted it” and “it never affected anything.”
How This Compares to ChatGPT and Gemini
Context matters here, because Anthropic isn’t operating in isolation — it’s one of three major consumer AI assistants making similar decisions around the same period.
OpenAI’s approach to ChatGPT has generally defaulted free-tier conversations toward potential use in training, with an opt-out setting available but not opt-in by design. Google’s Gemini, layered into Workspace and Personal Intelligence features, draws on a broader pool of personal context — Gmail, Photos, Calendar — with its own separate set of controls.
Measured against that landscape, Anthropic’s opt-in structure for the five-year retention trade is a more conservative default than the industry norm, even though the underlying activity — using personal conversations to train a model — is the same category of practice. An opt-in default doesn’t eliminate the privacy question; it does shift more of the actual decision-making onto the user, which only matters if the user is actually paying attention when the prompt appears.
The Setting Worth Checking Today
If you use Claude and don’t remember explicitly choosing an option when this update rolled out, that’s worth fixing today rather than assuming you’re on the conservative default.
Check your model training setting directly in Claude’s privacy settings, rather than relying on memory of what you clicked when a notification appeared. In-app prompts are easy to dismiss quickly without registering exactly what was being asked.
If you want the shorter retention window, you need the training setting turned off — not just “not actively turned on.” Depending on how and when you responded to the original prompt, your account could be sitting in either state, and the only reliable way to know is to check directly.
Reconsider what you’ve typed into Claude historically, independent of this setting. The new retention policy is forward-looking, but it’s a reasonable prompt to think about whether you’ve been using a chat interface as a substitute for somewhere more appropriate to store sensitive information — a password, a scanned ID number, a draft of something you haven’t shared with anyone yet.
Where This Gets Complicated for Document and Photo Uploads
Claude, like ChatGPT and Gemini, accepts file and image uploads — a scanned form, a photo of a handwritten note, a PDF you want summarized. Those uploads are conversation content for the purposes of this policy, meaning the same training-and-retention setting applies to a document you’ve uploaded as it does to a typed message.
That’s easy to overlook, because uploading a file feels more like a one-off task than an ongoing conversation. But if your training setting is on, a personal document you uploaded for a single summarization task is retained on the same five-year timeline as everything else in that conversation — not deleted once the task is done.
For documents and photos you actually want to keep — not just process once and forget — a chat interface was never really the right tool regardless of this policy. AI assistants are built for processing and generating text; they’re not designed as your archive, and a retention policy update is a reasonable nudge to separate the two functions if you’ve been blurring them.
What Hasn’t Changed
It’s worth being clear about what this update doesn’t touch. Anthropic’s underlying security practices, encryption in transit, and infrastructure protections aren’t part of this change — this is a policy update about retention duration and training consent, not a shift in how Claude secures data against external threats.
It also doesn’t change anything about Claude’s commercial or enterprise offerings, which were never part of the consumer training conversation in the first place. If you use Claude through a work account on Team or Enterprise, this entire update is irrelevant to your settings.
The Bigger Pattern
This update is part of a broader shift across the AI industry: as competition intensifies and training data becomes a more contested resource, every major AI company is finding new ways to ask users for permission to use their data, sometimes paired with an incentive and sometimes not. Anthropic’s version is comparatively transparent about the trade — a clear setting, a clear consequence, an opt-in rather than a buried default.
Transparent isn’t the same as inconsequential, though. Five years is a long time for a personal conversation to sit in a retained dataset, reviewable and usable for training, regardless of how clearly the option was presented. The right call depends entirely on what you actually use Claude for — and whether you’ve been treating it as a place to store things that deserve a more permanent, less negotiable home.
If that describes the documents and photos you’ve been uploading to AI tools just to get them off your phone, daftei is built for that instead: encrypted with AES-256 at rest and TLS 1.3 in transit, never used to train any AI model — Anthropic’s, anyone else’s, or its own — and never sold or shown to advertisers.