Key Points

  1. Depending on a service’s settings and policies, chatbot conversations and uploaded files may be processed or used to train AI models.
  2. Passwords, authentication codes, banking details, identifying personal information and confidential intellectual property are among the information users should keep out of chatbot prompts.
  3. Connected services such as Gmail or Slack can make it harder to know which information a chatbot may access, so users should check permissions and remove identifying details.

What Happens to Information You Give AI Chatbots

AI chatbots can process more than the words typed into a prompt. Depending on the service and the user’s settings, conversations, documents, images and other uploaded files may be used to improve or train models. Providers also set their own retention rules, so information may not disappear as soon as a conversation is deleted.

The article says OpenAI’s standard policy provides for deleted conversations and temporary chats to be removed from its systems within 30 days. It also reports that OpenAI briefly retained some user data indefinitely between April and September 2025 in connection with a New York Times lawsuit, before reversing the practice in an October 2025 update. Anthropic says chats are deleted from storage within 30 days; data users agree to retain may, according to the article, be kept in de-identified form for up to five years.

The practical concern is that sensitive information could be exposed through a data leak, accessed by people involved in reviewing data, or made more accessible through connected apps. The article warns that linking a chatbot to services such as Gmail or Slack can make it harder to understand what information the tool can reach.

As a basic rule, do not enter anything you would not want to become public. That includes passwords, login details, one-time authentication codes, bank or card information, identifying personal details and confidential intellectual property. If you need help with a health, legal or financial question, remove names and other identifying details before sharing relevant information.

Why Privacy Settings May Not Be Enough

Settings change how data is handled, but do not remove every risk

Turning off model-training or data-use options may reduce the chance that prompts are used to train a model, but it does not by itself establish that information is never processed or retained. The article also notes that users may have limited visibility into how effective privacy controls are in closed systems. The specific rules differ by provider and account settings, so a single retention period should not be assumed to apply across all AI tools.

Connected apps widen the information boundary

A chatbot linked to email or workplace software may have access to information beyond the text a user enters directly. That makes permissions important: a prompt may seem harmless while the connected service gives the tool a broader view of a user’s data. The article does not establish that every integration accesses information beyond its granted permissions, but it does identify permissions as a reason to be cautious.

De-identification can make sensitive questions safer

Removing names and other details that could identify a person can let users ask for general help without including direct identifiers. This is a risk-reduction step, not a guarantee of anonymity: distinctive circumstances or combinations of details may still identify someone. The safest approach is to share only the minimum information needed to get a useful response.

How to Use Chatbots Without Sharing Sensitive Data

Before submitting a prompt, treat it as information that could be seen beyond the conversation.

  • Do not paste passwords, authentication codes, card numbers or bank details into an AI chatbot; the article identifies these as sensitive information that could be misused if exposed.
  • Keep confidential company material and intellectual property out of consumer chatbot prompts unless your organization has explicitly approved the tool and its data-handling arrangements.
  • Before connecting Gmail, Slack or another service, check what access the integration requests and whether that access is necessary for the task.
  • For health, legal or financial questions, remove names and identifying details and include only the facts needed to frame the question.
  • Check the specific provider’s privacy and retention settings. The article describes different retention practices and notes that settings can affect whether conversations are used for training.