An AI chatbot is not a private diary

The inference

A chatbot can feel like a private place to think. But a consumer AI conversation is still data sent to a service, governed by that product’s settings and policies. That does not mean every conversation is routinely read or sent to authorities. It does mean a chat window is not the same thing as an encrypted notebook, a confidential therapy session, or a conversation protected by professional privilege.

The practical rule: use AI to organize a concern, but do not treat the chat as a private vault. Share the minimum detail needed, check the exact product’s controls, and move deeply personal writing to a place designed for private records.

Your words can pass through more than the model: the account, storage, review, and connected tools matter too.

Why this is in the news

A Florida woman was arrested after investigators said she used Claude to write that she planned to attack the Lee County Sheriff’s Office. WINK News, citing the arrest report, said the service’s safety systems flagged the conversation, a human reviewer escalated it, and law enforcement was contacted. The woman is charged, and the allegations have not been adjudicated. (WINK News)

The point is not that an ordinary journal entry will trigger a police report. This case involved an alleged specific threat. The useful lesson is narrower: a conversation with a cloud service may be subject to automated safety systems, human review under limited circumstances, and emergency disclosure rules. Anthropic’s published policy describes the circumstances in which it may disclose information to law enforcement, including emergencies involving imminent physical harm or death. Read the rules for the service you use, not just the reassuring feel of its interface. (Anthropic policy)

“Not used for training” is not “private”

Privacy is several different questions, often compressed into one switch. Turning off model improvement can affect whether new conversations are used to train models. It does not by itself answer how long data is retained, whether it appears in history or memory, whether a safety review can happen, who has account access, or what happens after a legal or emergency request.

Question	What it controls	What it does not automatically mean
Training / model improvement	Whether your data may be used to improve models	That the conversation is never stored or reviewed
Chat history and retention	Whether a conversation remains in your account and for how long	That every copy disappears instantly when you delete it
Memory / personalization	Whether information may be carried into future chats	That a new chat starts with no other account-level context
Human review and safety	Whether a conversation may be reviewed for service or safety reasons	That every conversation is routinely read by an employee
Legal or emergency disclosure	When information may be shared outside the service	That ordinary journaling is automatically reported

The controls differ by service and can change. OpenAI, for example, describes controls for model improvement, temporary chats, and managing account data. Anthropic publishes separate privacy and government-request policies. Treat these as product-specific settings, not an industry-wide promise. (OpenAI privacy controls, OpenAI data controls, Anthropic privacy policy)

A safer way to use AI for reflection

You do not have to choose between sharing everything and never using AI. Change the job you give it. Instead of pasting an intimate, identifying account of a conflict, ask it to help with a generalized situation. Instead of uploading an entire journal, write a short summary yourself.

Before you share, remove:

• Full names, addresses, phone numbers, account details, and specific workplaces.
• Other people’s private information, especially details they did not consent to share.
• Passwords, access codes, private medical records, legal documents, or financial account data.
• Exact details that could identify someone when combined, such as a small town, a rare job title, and a precise date.

Use a substitute such as “a family member,” “my manager,” or “a health concern.” Ask for a framework, questions to consider, or a draft you will revise. If the task can be done with a fictionalized example, use that instead of the original story.

For personal journaling, a paper notebook or a local notes app may be a better fit. Check device lock, cloud sync, backups, and shared-device access; local storage is not automatically safe if a phone or account is shared or compromised. For therapy, legal advice, or medical care, use the appropriate professional relationship rather than treating a chatbot as a confidential substitute.

Put it to work: a 30-second preflight

Before sending a personal prompt, pause and answer:

What am I trying to get help with?
What is the least personal version of this prompt that still works?
Which exact account and product am I using?
Are history, memory, training, and sharing settings what I expect?
Would I be comfortable if this text were reviewed under the service's policy?

If that last answer is no, do not send the original text. Ask a generalized question, write offline, or talk with a trusted person or qualified professional. If you choose to continue, check the current privacy controls for that exact account and revisit them when the service changes its terms.

Do not mistake comfort for confidentiality

Conversational AI is designed to respond naturally, which can make disclosure feel effortless. That is useful for brainstorming and sorting through everyday decisions. It can also make people forget the basic boundary: the service provider receives what you submit, and its policies determine how that information is handled.

Keep AI for the reflection prompt, not the private record. Share only what the task needs, keep identifying details out when possible, and use a channel whose privacy protections match the conversation.