What You Should Not Share With ChatGPT, Gemini or Claude: 12 Types of Data Safer Kept Out of Prompts

What You Should Not Share With ChatGPT, Gemini or Claude: 12 Types of Data Safer Kept Out of Prompts

Rating
—/5 0 ratings
Follow us on Google News
Technology15 min read

What information should you not give ChatGPT, Gemini or Claude? The safest rule is simple: do not enter into an AI chatbot information that you would not knowingly send to an online service, especially if it could give access to your money, accounts, identity or confidential documents. Passwords, authentication codes, cryptocurrency seed phrases and full card details should not be entered in a prompt. For other sensitive information — medical documents, identity documents, contracts or work files — the right approach is to minimise and anonymise the data, rather than uploading it in full “just in case”.

In 2026, ChatGPT, Gemini and Claude are no longer merely boxes into which we type questions. They can analyse files, images, spreadsheets, emails and documents, and some features can connect to other apps. That is why privacy no longer depends only on the text of a prompt, but also on uploaded files, conversation history, model-improvement settings, connected apps and the type of account being used.

This InfoBuz guide was documented on 1 November 2026, based on official privacy and data-control pages published by OpenAI, Google and Anthropic. Settings and retention periods may change, while business, enterprise, education or organisation-managed accounts may be governed by different rules from personal accounts.

For context on the data apps may already learn from your phone, see Your phone knows more about you than you think: what apps can learn from location, the microphone and digital habits. For the direction in which app-connected assistants are developing, see also After the smartphone comes the era of AI assistants.

What you should not give ChatGPT, Gemini or Claude: the quick answer

Type of information Main risk What to do instead
Passwords and PINs Direct access to accounts Describe the problem without the real password
2FA / OTP codes Hijacking a session or transaction Do not share them; enter them only in the service that legitimately requests them
Crypto seed phrases / private keys Irreversible loss of funds Keep them offline and never copy them into chat
Full card details Financial fraud Use fictional values or, where needed, the last four digits
Full copies of identity documents Identity theft Redact the Romanian personal numeric code (CNP), document series, number, address and photo if they are not needed
Workplace secrets Breach of confidentiality / NDA Use a company-approved environment or anonymise the document
Identifiable medical data Exposure of health information Remove names, CNP, address, codes and irrelevant dates
Children’s data Unnecessary exposure of a minor Use an approximate age and only strictly necessary information
Other people’s data Privacy breach Anonymise names, phone numbers, emails and identifiers
Confidential legal documents Professional privilege / strategy Remove identities and protected information before analysis
Exact location and personal schedule Physical-security risk Use the city or area rather than the full address where that is enough
API keys and tokens Access to systems and unauthorised costs Replace them with values such as YOUR_API_KEY

1. Passwords, PINs and authentication codes are not “context” for AI

A chatbot does not need your real password to explain why you cannot sign in, how to reset an account or how two-factor authentication works. If a technical example requires a password, use an invented value. The same principle applies to an OTP code received by SMS, email or an authenticator app: that code exists precisely to prove that the person entering it controls the device or account.

OpenAI explicitly advises users of agentic features not to enter passwords and private information directly in messages. The reason goes beyond model training: an assistant connected to the web or to apps may have access to more context, and a malicious webpage may attempt to manipulate an agent through a prompt-injection attack. See OpenAI’s official guide to ChatGPT agent and data safety.

2. Cryptocurrency seed phrases and private keys: never

Some information should not be “slightly anonymised”; it should not be shared at all. A 12- or 24-word seed phrase and a private key can give control over a crypto wallet. Whoever holds them can, in many systems, move funds without a conventional process for reversing the transaction.

If you need help setting up a wallet, you can provide the name of the app, the network in use and the error message, but not the seed phrase. In a code example, replace the real key with YOUR_PRIVATE_KEY or a fictional string.

3. Bank cards and financial data: the difference between “useful context” and direct access

To analyse a bank statement or a budget, AI does not need to receive the full card number, CVV, PIN, internet-banking passwords or authorisation codes. If you want to classify spending, remove unnecessary identifiers and retain only the date, merchant, category and amount. If you want to understand a charge on a statement, the last four digits of a payment instrument are generally enough for your own internal reference.

The right principle is data minimisation: provide only the information needed for the task. An AI model does not become better at explaining an interest rate because you also give it your IBAN, full address and card number.

4. Identity documents: a photo may contain far more than you need

A Romanian identity card, passport, driving licence and other official documents can combine a full name, date of birth, photograph, address, document series and other identifiers in a single image. If you want to ask where information appears on a form or what a field means, a complete image of the document is rarely needed.

Use a redacted copy: cover identifiers, codes, the signature, photo and any information not needed for the question. Importantly, an area that has been lightly blurred can sometimes be less secure than an opaque cover applied in an editor and checked in the final file.

5. Work documents: the issue is not only AI, but also your duty of confidentiality

An internal report, commercial offer, contract, technical drawing, customer database or internal conversation may be protected by an employer’s policy, an NDA, contractual obligations or legislation. Even where an AI service offers strong privacy controls, the user may not have the right to upload the document to a personal account.

The type of account also matters. OpenAI says that, by default, content from ChatGPT Business, Enterprise and Edu is not used to train its models, whereas personal accounts have their own controls for “Improve the model for everyone”. OpenAI’s Data Controls documentation explains the distinction. That does not replace company policy, however: if an organisation requires the use of an approved platform or prohibits the upload of certain documents, its internal rule takes priority for that material.

6. Medical data: you can ask for explanations without sending the whole file

AI can help explain medical terms, the structure of a test result or questions you want to prepare for a doctor. But a lab PDF or medical letter may contain the name, CNP, address, case number, doctor, healthcare provider and other data that are not needed to interpret a marker.

For a general question, retain only the relevant values, units, reference ranges and the context you need. In addition, an AI answer should not be treated as a diagnosis or replace a medical professional where there are significant symptoms, abnormal results or treatment decisions.

7. Children’s and other people’s data: having it does not mean you should upload it

A parent may legitimately have access to a child’s documents, and an employee may see colleagues’ or clients’ data. Still, an AI question usually does not require that person’s full name, exact date of birth, photo, address, school, phone number or identifiers.

Instead of “Maria Popescu, born on date X, a pupil at school Y”, “a child aged about six” is often enough. Instead of a client’s name, use “Client A”. Anonymisation does not eliminate every risk, but it substantially reduces the volume of personal information shared unnecessarily.

A contract can be analysed for structure, unusual clauses or difficult wording, but it is prudent to remove names, addresses, bank accounts, signatures and irrelevant annexes. A dispute or lawyer-client relationship may entail additional obligations around confidentiality and professional privilege.

If a document has significant legal consequences, use AI to organise your questions, not as a substitute for qualified advice. A model may miss an exception, cite an outdated rule or misinterpret the jurisdiction.

9. Exact location, when your home is empty and travel plans can become physical-security data

For a restaurant recommendation, a neighbourhood or city is often enough. For a weather forecast, the locality may suffice. A full address, building access code, periods when a home will be empty, or the combination of an address and schedule are more sensitive details.

That does not mean every location service is dangerous. It means there should be a clear reason for the granularity of data you provide.

10. API keys, tokens and configuration files: a common trap for developers

A developer asking for help with a script may accidentally paste an API key, OAuth token, URL containing credentials, a .env file or a log excerpt that includes secrets. Replace these values before sending them — for example, OPENAI_API_KEY=YOUR_KEY_HERE.

If a secret has already been exposed somewhere it should not have been, the prudent approach is to revoke and rotate it, rather than merely delete the message and assume the problem has gone away.

11. Photos and screenshots: check the background, notifications and visible metadata

A screenshot may show an email address, phone number, company name, account balance, private notifications or a QR code. A photograph of a document may include other pages in the background. Before uploading, inspect the image at full size and crop anything unrelated to the question.

For more information about ChatGPT, see ChatGPT designated a VLOSE: the EU does not review every answer, but risks at scale, which provides complementary details on this subject.

This matters increasingly as AI assistants can directly understand images and screens, rather than only manually entered text.

12. An “anonymous” prompt can become identifiable through a combination of details

Removing a name does not guarantee anonymity. “The 47-year-old finance director at the only factory X in town Y” may be identifiable even without a name. The same can happen with a very rare medical case, an unusual transaction or an incident described with a date, place, role and company.

When anonymising, ask whether the remaining information could be combined to identify the person. If so, generalise the details: use an age range rather than a date of birth, an industry rather than a company, a region rather than an address, and approximate values where precision is unnecessary.

Do ChatGPT, Gemini and Claude retain data in the same way? No

The three platforms have different policies and controls, and the rules change over time. The table below summarises official information available on 1 November 2026 for personal use. It should not be read as a universal promise for every plan or region.

Service Relevant user control What to remember
ChatGPT “Improve the model for everyone” and Temporary Chat On eligible personal accounts, you can switch off the use of conversations for model improvement. Temporary Chat does not appear in history, does not create memories and is not used for model improvement; OpenAI says it may be retained for up to 30 days for safety.
Gemini Keep Activity and Temporary Chat Google says Keep Activity is on by default for users aged 18 and over. Temporary chats and chats created with Keep Activity switched off are retained for 72 hours and are not used to train AI models unless you submit feedback. Some human-reviewed conversations may be retained for up to three years.
Claude Model-training setting in Privacy Settings Anthropic gives Free, Pro and Max users the option to allow conversations to be used for model improvement. For participants, the company announced a retention period of up to five years for new or resumed conversations; if you do not enable this option, its documentation indicates a standard 30-day period.

Official sources: OpenAI – Data Controls in ChatGPT, Google – Gemini Apps privacy help and Anthropic – Updates to Consumer Terms and Privacy Policy.

ChatGPT: what Temporary Chat does — and does not — do

According to OpenAI, a Temporary Chat does not appear in conversation history, does not create or update memories, and is not used for model improvement while it remains temporary. OpenAI says, however, that such conversations may be retained for up to 30 days for safety reasons. “Temporary” should therefore not be understood to mean “the data never reaches a server” or “it disappears instantly”.

On personal Free, Plus or Pro accounts, the “Improve the model for everyone” control can be switched off under Settings → Data controls. OpenAI says switching it off does not automatically delete existing conversations; history and model-improvement settings are separate controls.

Gemini: why “Keep Activity” is an important setting

Google’s Gemini documentation, updated in October 2026, states that Keep Activity is on by default for users aged 18 or over. Activity is automatically deleted after 18 months by default, with options for 3 months, 36 months or retention without automatic deletion.

If you switch off Keep Activity or use Temporary Chat, Google says new conversations are not used to train its AI models unless you submit feedback. They are nevertheless retained in the account for up to 72 hours for service operation and protection. Google also warns that some data may be reviewed by human reviewers, and explicitly says not to enter confidential information that you would not want a reviewer to see or to be used to improve services.

Gemini can also use connected apps. Depending on the features enabled, this may bring emails, files, events, photos, contacts or other information into context. Connecting an app should be treated as expanding the data surface, not as a purely cosmetic feature.

Claude: the model-training choice and the difference between consumer and business services

Anthropic introduced an option for Claude Free, Pro and Max users to allow conversations and coding sessions to be used for model improvement. The company says this choice can be changed in Privacy Settings. Where a user permits data to be used for model training, Anthropic announced a five-year retention period for new or resumed conversations and feedback; without this option, the company indicates a standard 30-day period.

The rules are not identical for services covered by Commercial Terms, such as Claude for Work or the API. That is why claims such as “Claude retains data for X days” are incomplete unless they specify the plan and the user’s setting.

What does “not used for training” mean? It does not mean “not processed”

This is one of the most important distinctions. A service has to process a prompt in order to answer it. There may also be processing for security, abuse prevention, system operation, legal compliance or feedback. The option not to use a conversation to train future models is therefore important, but it does not turn a chatbot into a vault for passwords and secrets.

The 60-second method before sending a document to AI

  1. Ask whether AI needs the whole document. A snippet is often enough.
  2. Remove passwords, codes, API keys and full financial data.
  3. Remove personal identifiers: names, CNP, address, phone number, email and document series.
  4. Remove other people’s data unless it is essential.
  5. Check company policy before uploading work material.
  6. Check the platform’s settings: training/model improvement, history, Temporary Chat and connected apps.
  7. Read the prompt again as though it would become visible to someone else. If that would create a serious problem, redact it again.

What to do if you have already sent sensitive information

The action depends on the type of information. If you sent a password, change it. If you sent an API key, revoke it and generate another. If you exposed an access token, invalidate the session. If you sent a card number together with authentication elements, contact the card issuer and follow its security procedure. If it was a confidential work document, notify the responsible person under your internal policy.

Deleting the conversation is a reasonable hygiene measure, but it does not replace rotating a compromised secret. A secret such as a password or API key should be treated as exposed if it has reached a place where it should not have been.

You can use AI safely without giving up useful features

The right message is not “never upload anything to AI”. These tools are useful precisely because they can work with information. The difference lies between providing the necessary context and providing everything at hand. An anonymised contract can be analysed. A transaction table without account numbers can be classified. A medical result without identifiers can be explained. A code snippet without the API key can be debugged.

The best habit is the same one long used in information security: minimum necessary. Give the system only the data it needs for the specific task, and nothing more.

Frequently asked questions about data sent to ChatGPT, Gemini and Claude

Can ChatGPT see my password if I type it in a conversation?

If you enter a password in a message, it becomes part of the content the service has to process in order to respond. There is no normal reason to send a real password to get help. Use a fictional value, and change the password immediately if you exposed it by accident.

Does Temporary Chat mean the conversation is not retained at all?

No. OpenAI says Temporary Chat may be retained for up to 30 days for safety. Google says Gemini Temporary Chat is retained for up to 72 hours. In both cases, the feature reduces persistence and use for model improvement, but should not be confused with zero processing.

Can I upload an employment contract to an AI chatbot?

Technically, the platform may allow the upload, but first you should check your organisation’s policy and remove personal data or confidential clauses that are not necessary. For an important legal decision, have the result checked by a qualified professional.

Is it safe to send medical test results to AI?

For a general explanation, you can reduce the risk by removing the name, CNP, address and other identifiers, and sending only the relevant values. AI should not be used as a replacement for medical diagnosis or treatment.

If I turn off training, can I put any information in chat?

No. Switching off the use of data for model improvement is an important privacy control, but the service must still process the message and retention periods or security uses may still apply. Passwords, 2FA codes, seed phrases and private keys should not be entered regardless of the setting.

Conclusion: treat the chatbot as a powerful tool, not a vault

What information should you not give ChatGPT, Gemini or Claude? First and foremost, any secret that gives direct access to accounts or money: passwords, PINs, OTPs, seed phrases, private keys and tokens. Then, any personal, medical, legal or professional data that can be removed without undermining the question should be redacted before upload.

ChatGPT, Gemini and Claude offer real controls for history, model training and temporary conversations, but the settings are not identical and do not remove the need for caution. The safest question before pressing “Send” remains very simple: does the AI really need this information to answer me?

Etichete: ChatGPTClaudeAI privacyPersonal DataGDPRGeminiArtificial intelligencePasswords
Share this article
COMUNITATEA INFOBUZ

Discuția cititorilor

Ai o opinie sau o întrebare despre acest articol? Spune-ne ce crezi.

0 comentarii

Fii primul care deschide discuția.

Lasă un comentariu

Email address este vizibilă doar administratorilor și nu va fi publicată. Câmpurile marcate cu * sunt obligatorii.

Te rugăm să păstrezi un limbaj civilizat și să discuți la subiect. Comentariile spam sunt filtrate automat.