AI Data Retention Compared: How Long Each Vendor Keeps Your Chats

TL;DR: AI vendors run several retention clocks at once and advertise only the shortest. Deleted Claude chats leave back-end storage within 30 days, while safety classification scores can be kept for seven years. Perspective AI's private mode starts no server-side clock, and it sits in one subscription with every other model rather than in a separate privacy plan.

Retention is where AI privacy policies stop being vague and start producing numbers, which makes it the most useful thing to compare. The catch is that every vendor runs several clocks simultaneously and quotes whichever is shortest. This page separates them, gives the published figure for each, and records the source URL and the date it was read, so the comparison stays checkable rather than becoming folklore. Retention is also the clearest reason to hold the vendors in one subscription rather than one at a time: you can switch to the model whose clock you can live with instead of replacing a plan every time a policy moves. It lives in our comparisons section. Perspective AI is included on the same terms as every other vendor, including where its answer is a design rather than a duration.

How Long Do AI Companies Keep Your Chats?

Between 72 hours and 7 years, depending on the clock. Google auto-deletes Gemini activity at 18 months by default. Anthropic clears a deleted chat from back-end storage within 30 days, and can hold safety classification scores for 7 years.

Those two sentences describe the same industry and sound like different ones. The reason is that "how long do they keep it" is five questions.

Five Clocks Run At Once

Every published policy we read distinguishes, explicitly or implicitly, between these five. They start at different moments, run for different lengths, and only the first is under your control.

• The delete clock. How long after you press delete before the copy leaves back-end storage.
• The activity clock. How long a kept conversation sits in your history if you never touch it.
• The training clock. How long a copy lives inside a model training pipeline once it has been ingested.
• The human review clock. How long material pulled for quality review is kept, which is frequently decoupled from your deletion.
• The safety clock. How long content and derived scores are kept when a conversation is flagged for policy enforcement.

A vendor can be entirely truthful about clock one while clock five runs for years. Nothing about that is hidden, but it is spread across separate documents and separate paragraphs, which does the work of hiding it.

The Delete Clock: The First Thirty Days

Anthropic publishes the clearest version of this one. A deleted conversation is "removed from your chat history immediately" and "deleted from our back-end storage systems within 30 days." Thirty days is a common industry figure and it is a reasonable one: backup rotation and incident recovery both need a window.

What is worth noticing is that this is the number vendors reach for when asked how long they keep your data, and it is the one that answers the question least. It describes what happens after you take an action most users never take, to a copy that is only one of several.

The Activity Clock: Eighteen Months By Default

Google's Gemini Apps run the most legible activity clock. The default auto-delete is 18 months, changeable to 3 months, 36 months, or indefinite. Turn Keep Activity off, or use a temporary chat, and the window collapses to 72 hours, retained so the service can respond to you using the last 24 hours as context and to protect Google, its users and the public.

That 72-hour figure is the shortest fixed retention number in this comparison, and it is available today to anyone with a Google account. It is also the one most obscured by the interface, since it only applies once you have turned off a setting that ships on. Mistral takes the opposite approach and lets you set the clock directly: its chat retention policy offers Never, 30 days, 60 days, 90 days, 180 days or 1 year, after which chats and their artifacts are deleted automatically.

The Training Clock: Measured In Years

Once a conversation enters a training pipeline it stops obeying the delete clock. Anthropic is specific: if you allow your chats to be used to improve Claude, "we may retain your data in a de-identified format for up to 5 years in our model training pipelines." Feedback you submit through the thumbs up and down buttons is retained for 5 years as well.

Five years is a long time next to thirty days, and the gap is the point. These are not competing claims, they are different objects. Which vendors put your chats into that pipeline in the first place, and what the setting is called, is the subject of who trains on your conversations.

The Human Review Clock: The One That Survives Deletion

Google states that chats reviewed by human reviewers, along with related data such as your language, device type, location information or feedback, "are not deleted when you delete your activity" and are instead retained for up to three years.

This is the single most consequential sentence in any policy on this page, because it breaks the mental model the delete button creates. You cannot tell whether a given conversation was sampled for review, there is no interface that shows you, and deleting the conversation does not reach it. Any personal retention strategy built on "I delete my history" is silently missing this pool.

The Safety Clock: Seven Years, The Longest Figure Published

The longest documented retention number we found belongs to the vendor with the tidiest deletion policy. Anthropic states that if a chat is flagged by automated trust and safety systems as violating the Usage Policy, it retains "inputs and outputs for up to 2 years and trust and safety classification scores for up to 7 years."

Two years of content and seven years of derived scores, from a company that deletes an ordinary conversation from back-end storage in thirty. Anthropic is not the outlier here for having a long safety clock, it is the outlier for publishing the number. Most vendors describe enforcement retention as "as necessary" and leave it there, which is why the comparison below has more prose than digits in places.

The Full Comparison, With Sources

Of the six consumer chat products whose policy we could read, three publish a numeric retention schedule for chat content: Anthropic, Google and Mistral. Four further vendor policies returned HTTP 403 to an ordinary browser request on the same day, including OpenAI's ChatGPT consumer help centre, so their rows are absent rather than estimated.

Reading A Retention Policy In Four Minutes

The skill transfers to any vendor, including ones that launch after this page goes stale.

1. Search the document for a digit followed by "days", "months" or "years". If there are none, the vendor has not committed to anything and the rest is atmosphere.
2. Find the word "flagged" or "violat". This is where the long clock hides, and it is almost never in the section headed retention.
3. Find "review" and check whether the sentence connects it to your deletion. Google's does, in the negative. Most policies simply do not say.
4. Check the storage jurisdiction. DeepSeek's policy states it stores personal data in the People's Republic of China. No retention setting alters a jurisdiction, and for some readers this field outranks every number above it.

Do it once on the tool you already pay for. If the exercise changes which subscriptions you want to keep, what the vendors in this table charge is the next arithmetic, and our ranking of the best AI for privacy sorts the same field by how much of each claim you can verify rather than by duration.

The Mode Where No Clock Starts

Every number on this page describes how long a server keeps something it was given. The alternative is not a shorter number, it is a different shape: an architecture where the server is never handed the transcript. Perspective AI's private mode works that way. Its published documentation describes private mode as browser-held and locally encrypted, with each request relayed statelessly so usage is metered without the prompt or the reply being kept. Clearing the browser clears the record, which is why no deletion request is offered for those conversations, and none is needed. The mechanism-by-mechanism breakdown maps each of those sentences to the document it came from.

Two honest boundaries. Your prompt still has to reach the model that answers it, which no product on this page avoids. What survives in every mode is the account itself: who you are, what you pay, and how much you have used, measured as quantity and never as content. What you get for $14.99/mo is not the absence of a record, it is the ability to decide per conversation whether one exists, across the catalog covering GPT, Claude, Gemini, Grok and DeepSeek. Standard mode remains the right default when you want your threads to follow you between devices.

FAQ