{"id":21630,"date":"2026-09-18T16:32:33","date_gmt":"2026-09-18T20:32:33","guid":{"rendered":"https:\/\/www.psafe.com\/en\/blog\/?p=21630"},"modified":"2026-09-18T16:32:33","modified_gmt":"2026-09-18T20:32:33","slug":"title-what-not-to-share-with-chatgpt-gemini-or-ai","status":"publish","type":"post","link":"https:\/\/www.psafe.com\/en\/blog\/title-what-not-to-share-with-chatgpt-gemini-or-ai\/","title":{"rendered":"What Not to Share With ChatGPT, Gemini, or Any AI"},"content":{"rendered":"

You can ask an AI assistant to summarize a contract, review a document, or explain medical language. Before you upload anything, ask how much of that material the tool actually needs. Knowing <\/span>what not to share with AI<\/b> helps anyone using ChatGPT, Gemini, or another assistant avoid exposing passwords, identity records, other people\u2019s information, and protected work files.<\/span><\/p>\n

Use a three-level rule: some information should never enter a chat; some can be useful only after names and identifiers are removed; confidential business content belongs solely in an organization-approved environment. This guide shows how to make that decision, what privacy controls actually change, and what to do after an accidental disclosure.<\/span><\/p>\n

What not to share with ChatGPT, Gemini, or any AI<\/b><\/h2>\n

Never share passwords, one-time codes, API keys, access tokens, session cookies, or seed phrases. Keep Social Security numbers, full identity documents, financial details, identifiable health records, precise locations, and other people\u2019s data out of consumer AI chats. When a task truly needs context, reduce and redact the material and use only an approved environment.<\/span><\/p>\n

Never share access credentials or secrets<\/b><\/h3>\n

A password, authentication code, or API key is not helpful background information. It is a way into an account or system. The same rule applies to recovery codes, PINs, private keys, session tokens, and cryptocurrency seed phrases. Do not paste them into a chat to ask whether they are valid or secure.<\/span><\/p>\n

Replace real values with placeholders such as <\/span>[PASSWORD]<\/span>, <\/span>[API_KEY]<\/span>, or <\/span>[ONE_TIME_CODE]<\/span>. When troubleshooting code, remove the credential and share only the smallest relevant section.<\/span><\/p>\n

Redact personal documents before using them<\/b><\/h3>\n

An SSN, driver\u2019s license, passport, bank statement, tax return, medical record, insurance identifier, child\u2019s information, or intimate detail can create identity and privacy risks. For many tasks, the assistant does not need the person\u2019s name, document number, address, date of birth, or account details. Review headers, footers, signatures, QR codes, photos, comments, hidden pages, visible notifications, and metadata. Removing the filename alone is not enough.<\/span><\/p>\n

Keep protected work data in an approved environment<\/b><\/h3>\n

Client contracts, private source code, product plans, payroll files, incident reports, legal strategy, and unpublished research do not automatically belong in a consumer chatbot. Even when a business service says that customer inputs and outputs are not used to train models by default, the organization may still restrict which data can be processed, retained, or accessed by administrators.<\/span><\/p>\n

Confirm the approved tool, account, and data classification. If the policy is unclear, use fictional material or follow the organization\u2019s official security, privacy, or IT process. A personally purchased subscription is not the same as an employer-approved workspace.<\/span><\/p>\n

Related:<\/b> What changes in ChatGPT for teens and parents<\/span><\/a><\/p>\n

Turning off training does not make a chat a vault<\/b><\/h2>\n

Training, service processing, history, memory, safety retention, and administrator access are different concepts. A service has to process what you submit to generate a response. Other handling can depend on the product, account, settings, feedback, safety needs, organization rules, and legal requirements.<\/span><\/p>\n

OpenAI describes how content from consumer services may be used to improve models and how users can control that use. It separately explains the current behavior of Temporary Chat. Google documents activity, retention, human review, and connected services in the<\/span> Gemini Apps Privacy Hub<\/span><\/a>.<\/span><\/p>\n

Those controls matter, but they do not turn a password, someone else\u2019s medical record, or a trade secret into appropriate chat content. Check the current policy instead of relying on a remembered setting or retention period; product names and terms can change.<\/span><\/p>\n

Personal, temporary, and managed AI accounts compared<\/b><\/h2>\n

A personal account provides consumer privacy and data controls, but you still decide what to submit. A temporary chat can limit history and model-improvement use under the provider\u2019s stated conditions. It does not mean the service performs no processing or that every category of information is acceptable.<\/span><\/p>\n

A business, enterprise, or school workspace can have different contractual protections and controls. It can also be managed by an organization that may audit, retain, export, or delete content according to its configuration and policies.Google describes privacy protections for Workspace with Gemini and the controls governing access to Workspace data.<\/span><\/p>\n

Ask more than \u201cIs this used for training?\u201d Ask whether the tool is approved, whether real information is necessary, whether you have permission to use another person\u2019s data, who manages the account, and whether the task can be completed with less content.<\/span><\/p>\n

How to redact a document before uploading it<\/b><\/h2>\n

Redaction is not just deleting a name from the title. Review the material around the specific task:<\/span><\/p>\n

    \n
  1. Define the goal.<\/b> If you want help rewriting one clause, do not upload the entire contract.<\/span><\/li>\n
  2. Remove direct identifiers.<\/b> Replace names, SSNs, license and passport numbers, addresses, account details, and patient identifiers with placeholders.<\/span><\/li>\n
  3. Look for indirect identifiers.<\/b> A rare job title, exact date, small town, family detail, or distinctive event can reveal a person when combined.<\/span><\/li>\n
  4. Inspect every format.<\/b> Crop unrelated notifications from screenshots, remove unnecessary pages, and review comments, images, audio, and metadata.<\/span><\/li>\n
  5. Read the reduced version again.<\/b> Ask whether a person, company, or case is still identifiable and whether the assistant truly needs each remaining detail.<\/span><\/li>\n<\/ol>\n

    Use labels such as \u201cClient A,\u201d \u201cCity X,\u201d and clearly fictional amounts. Redaction can reduce exposure, but it does not replace consent, an organization\u2019s policy, or a valid reason to process someone else\u2019s information.<\/span><\/p>\n

    What to do after sharing sensitive information<\/b><\/h2>\n

    Start by identifying the type of information. The right response is different for each category:<\/span><\/p>\n