The last post was about getting you to think carefully about what data you allow Claude Code to touch. This one is about what happens the moment you do. When Claude Code opens a file, the contents of that file leave your computer and go to servers that Anthropic operates. The same thing happens with any AI assistant: when it opens a file, the data travels to a company server. This is what data processing by a cloud-hosted model means: the computation happens somewhere else, not on your own computer.
Cloud computing by an AI assistant like Claude Code raises two separate questions. First, what data leaves my computer? And second, what does Anthropic (or any AI company) do with the data once it gets to their servers?
What actually travels
Two data streams leave your computer in a Claude Code session.
The first is the content stream: the prompts you enter into the AI assistant, and the contents of any files Claude Code reads. These data travel to Anthropic servers in encrypted form, which means it is protected in transit from interception. But the data does reach Anthropic’s infrastructure.
The second data stream is what’s called “telemetry.” These are usage metrics, latency signals, reliability indicators, and so on. This is the operational data software companies have always collected to keep their tools running. Telemetry from Claude Code never includes your file contents, your prompts, or your file paths; error reports redact them before sending. If you prefer to disable telemetry, you can do so (ask Claude Code to tell you how). For most people, though, this is not the primary concern — the content stream is.
What your account type determines
Once your data arrives at Anthropic’s infrastructure, your account type governs what happens to it.
Free, Pro, and Max accounts. Anthropic trains on your inputs and outputs when the “Help Improve Claude” setting is on. In these account types, that setting is on by default. If you want it off, you must turn it off yourself, by going to Settings → Data Privacy Controls on claude.ai. If your settings allow your data to be used for training, Anthropic retains your Claude conversation data for up to five years. If you opt out, the standard retention period is thirty days. One important qualification: conversations flagged for safety review may still be retained and used for training purposes even after you have opted out. That limit is set by Anthropic’s policies, and you can’t change it.
Team and Enterprise accounts. For these account types, Anthropic’s commercial terms prohibit training on customer content by default. The data still travels to Anthropic servers and is processed there, but it is not used to improve the model. The Data Processing Agreement governing these accounts usually specifies a thirty-day retention period measured from the end of the contract, not from the end of each session. That means that if your institution’s agreement runs for two or three years, data is retained for that entire period; the thirty-day clock starts only when the contract terminates.
Zero Data Retention. Some Enterprise organizations can request a Zero Data Retention arrangement, under which Anthropic does not persistently store inputs or outputs beyond what is required to complete the request, satisfy legal obligations, or enforce safety policy. ZDR is not a standard feature of Enterprise plans. It requires a separate approval process with Anthropic’s account team. Whether your institution’s agreement includes ZDR is a question for your IT or data governance office. Be aware that the terms of data processing agreements can change over time — protections you have today may not be the ones you have next year.
If you are at an institution with an Enterprise account
My own Claude account — and those of my faculty, staff, and student colleagues at the University of Chicago — is an Enterprise account. The protection that gives me and other Enterprise users is real, but narrow. The prohibition on training is a genuine contractual commitment: our data does not go into the model improvement pipeline. But Anthropic can still process our content for what the Data Processing Agreement calls “permitted uses”: maintaining quality and reliability, security monitoring, abuse prevention, and debugging. Legally speaking, those are “operational purposes,” not “training,” and the legal distinction is meaningful. But it is not a distinction that keeps your data from being read — by automated systems, and in some cases by human reviewers acting under those same permitted purposes.
It’s also important to recognize that even for Enterprise users, data does not stay with Anthropic alone. The DPA will name approved subprocessors — infrastructure and service providers to whom Anthropic passes user content. The legal standard requires that those subprocessors operate under terms “at least as protective” as the DPA itself. What that means in practice is determined by Anthropic, not by you, and the list of subprocessors can change on thirty days’ notice.
No matter what kind of account you have, whenever you open a file in Claude (via Chat, Code, or Cowork), that data leaves your computer to travel to a server. If the file contains identifiable participant data, patient records, or student information, then disclosure has already happened.
If you have a personal account
If you use Claude Code or Chat via a personal account (not an Enterprise account), there is one immediate action you can take: turn off “Help Improve Claude” in your settings if you have not done so already. It will not change the fact that your data travels to Anthropic servers, but it will end your participation in Anthropic’s training data pipeline for normal usage. Note that if you use the thumbs-up or thumbs-down on a Claude response, that specific exchange is treated as voluntarily shared feedback and may be used for training regardless of your opt-out status.
The takeaway
The checklist for this post:
Every time Claude Code reads a file, that content goes to Anthropic’s servers — encrypted in transit, but traveling. Your account type determines what Anthropic does with it once it arrives.
- Personal account (Free, Pro, Max)? Go to Settings → Data Privacy Controls → turn off “Help Improve Claude.” Do it now.
- Institutional or Enterprise account? Ask your IT or data governance office: does our agreement include Zero Data Retention, and does it apply to the models we are currently using? If the answer is “no ZDR,” you might consider if you want to press your organization to add this to its Data Processing Agreement.
- Either way: your account’s data-retention terms do not change the governance obligations from the last post. What the people in your files agreed to is a separate question from what Anthropic agreed to.
Next: what your computer keeps after a session ends — the transcripts Claude Code stores locally, where they live, how long they stay, and how to find them.