You have found the setting, switched it off, and you feel sorted.
Turn off training and Claude will not learn from your chats.
That is true, and it is the main thing.
But there is one honest exception worth knowing about.
What is the exception?
It applies to conversations flagged by Anthropic's safety classifiers.
A classifier is an automated check on a chat.
It decides whether the chat may be breaking the Usage Policy.
A flagged conversation can still be used, even with training off.
The purpose is narrow: detecting abuse and enforcing the rules.
That includes improving Anthropic's internal trust and safety models.
Those are the systems that catch misuse, not the Claude you chat with.
How long is a flagged chat kept?
Longer than an ordinary one, and the numbers are worth knowing.
- Inputs and outputs are kept up to 2 years
- Trust and safety scores are kept up to 7 years
Inputs and outputs means what you typed and what Claude replied.
Scores are the ratings the system assigned, not the conversation itself.
Seven years sounds enormous.
Then you remember how long a bank keeps a suspicious transaction record.
Safety systems spot repeat patterns, and patterns take years to appear.
Should this worry me?
For normal use, no.
This is a safety mechanism, not a backdoor to train on your data.
It only touches flagged, policy-violating conversations.
Your chats about spreadsheets, travel plans and homework are not in that bucket.
The reason to know about it is honesty, not fear.
An app that says "off means off" with no footnote is hiding something.
Anthropic publishes the footnote, which is a point in its favour.
What is the practical takeaway?
Turn training off, and trust it for ordinary conversations.
Then use Claude the way you use any tool that keeps a record.
Stay well inside the Usage Policy and the exception never applies to you.
That is the whole point of it.
Things to Remember
- Keep training off as your default; it still covers all your normal chats.
- Understand that flagged, policy-violating chats are the single exception.
- Note the retention: inputs and outputs up to 2 years, safety scores up to 7 years.
- Treat this as abuse detection, not as data collection through the back door.
- Stay inside the Usage Policy and this exception never touches you.