SAN FRANCISCO – If you have typed something private into an AI and wondered where it went, one company has now answered. The answer turns on one word it never defines.
Anthropic, which makes Claude, published a short explanation on 14 September, two minutes and nineteen seconds long. It is addressed to organisations that buy Claude, and that shapes what it answers.
Everything below is the company describing its own practice. That makes it the authority on its policy, not proof the policy is followed.
Contents
What is kept, and what is never said about day 31#
The reason comes first.
“The most sophisticated misuse can spread across many sessions and accounts … Seeing that pattern requires storing data for a meaningful period of time.”
Then the rule.
“For this reason, starting with Claude Fable 5, the last thirty days of prompts and Claude’s outputs are stored for safety monitoring.”
It names no customer and no plan.
“Rolling window” implies the oldest day drops off the end. Neither document says the data is deleted. Searched both for delete, deleted, deletion, erase and purge: nothing. The written page uses discard once, and says the company does not analyse each interaction and then “instantaneously discard the data”. A window that moves and a thing destroyed are not the same statement, and only one was made.
Nor does either say whether a flagged conversation leaves the window on schedule.
There is a second clock, and it is not yours.
“Any conversation history a product keeps for you still follows your organization’s own retention settings.”
It separates you from your organisation on purpose. What you can still see is set by somebody else.
The word the whole thing turns on#
The explanation ends with a carve-out.
“To make the transitions smooth, eligible organizations will receive zero data retention on Claude Fable 5 and Fable 5.1 until Enterprise Frontier Safeguards is ready.”
So some customers get zero rather than thirty, and only until the new product ships.
Which customers? The word “eligible” appears twice in the video, twice on the company’s written announcement of 1 September, and is defined in neither. No revenue threshold, no plan name, no product list.
That page lists where the new option will run, from Claude Code to Amazon Bedrock. Running on a product is not the same as being eligible, and the page does not join them.
What it changes is location, not existence. It lets “eligible organizations” keep the data “in their own cloud infrastructure, not Anthropic’s”, and the alert then “goes directly to your team”.
There is no description anywhere of opting out of the monitoring itself. The page names three choices a customer can switch on: where the data sits, who holds the keys, and whether review is fully automated. Not whether anything is kept.
Who can read it#
“by default, no Anthropic personnel can read the data used for monitoring.”
What reads it is software.
“Automated systems then analyze that rolling window of traffic for signals of serious misuse.”
Then the exception.
“Human review can happen only in rare cases, for example when an automated safety system flags content as potentially harmful …”
It ends by limiting them to “a small set of approved reviewers”. How many, and who approves them, is not said.
“Any access is recorded in a tamper-proof log.”
Who may read that log is not said either, and nothing says the person whose conversation was read is ever told.
It also separates monitoring and training.
”… Anthropic does not train on enterprise data without explicit permission …”
Note the word enterprise. As worded, that promise covers an organisation.
No rollout date. Later this fall is a season. Searched both for October, November, December and week: none.
Seventeen outside voices, and not one of them is a check. The written page quotes seventeen named executives, five of them chief information security officers. Four are labelled service partners and one leads a bank industry body. The rest are customers. No regulator, no researcher, no privacy body and no user appears in either document.
And no volume. The video gives a window and never a number: not one figure for conversations, size or accounts.
Nothing about a personal account. Searched the video for consumer, personal, individual, free plan and subscriber: nothing. The protections it does name, no personnel reading by default and a tamper-proof log, are scoped neither way.
So the useful question is not the number. It is the word. If you use Claude through work, ask whether your organisation is eligible, and for how long. If you use it on your own account, neither document is about you.
The test is the same. A claim is worth what its source is worth: how to check one. Compare when a machine decides, where rules are public.
Source: Anthropic, “How data retention works when using Claude”, published 14 September 2026, 2 minutes 19 seconds, 321 words, pulled and read in full, then extracted with locators before drafting. The written announcement linked above is by the same company, and is not a second source.