🧠 Artificial intelligence

Privacy and AI: what you should never paste into a chatbot

Which data not to share with chatbots, how to read the settings on conversation retention and which habits protect your privacy when you use AI services.

Typing to an artificial intelligence assistant gives a feeling of privacy: there is just a text box, no visible person on the other end. In reality every message is sent to a provider's servers, processed and, depending on the service and its settings, kept for a certain period. That is not necessarily a problem, but it is worth knowing before you paste anything.

What happens to your messages

When you use an online chatbot, this is generally what happens:

  1. the text you type and the files you attach are transmitted to the provider;
  2. the model processes them to generate the answer;
  3. the conversation may be saved in your account so you can find it again;
  4. depending on the terms, the provider may keep it for security, abuse prevention or model improvement purposes;
  5. in some cases, a portion of conversations may be reviewed by authorised staff.

Rules vary greatly from service to service and between free, paid and business versions. That is why the first useful habit is to read the privacy settings of the service you use, looking in particular at whether data is used for training and whether you can switch that off.

What never to paste (or almost never)

Some categories of information deserve particular caution.

  • Credentials. Passwords, access codes, API keys, recovery codes. There is no reason to enter them into a chatbot.
  • Identity documents and banking data. ID cards, tax codes combined with other data, IBANs, card numbers.
  • Health data, yours or other people's: test results, diagnoses, treatments.
  • Third-party data. Customer emails, candidates' CVs, information about colleagues. Sharing it may breach confidentiality obligations and data protection law.
  • Business secrets. Non-public contracts, confidential price lists, proprietary source code, strategic plans, unless you are using tools with adequate contractual guarantees approved by your organisation.
  • Information about proceedings or investigations you are aware of for professional reasons.

How to get help without exposing data

In most cases you can achieve the same result by removing identifying information.

Anonymise or pseudonymise

Instead of pasting a customer's email as it is, replace the data: "Customer A, who owns a shop in [city], complains about a delay in delivering order no. [X]". The model can help you reply without knowing who the customer is.

Use dummy data

If you need help with a spreadsheet formula, create a few invented example rows with the same structure as the real data.

Describe instead of copying

For a technical problem it is often enough to describe the behaviour ("the function throws an error when the list is empty") and paste only the relevant part, stripped of internal addresses, server names and credentials.

Instead of... You can write...
A real full name "Person A", "the employee"
A full address "an address in Milan"
A real case number "case no. 123" (invented)
An entire log file The essential lines, without IPs and usernames

Settings to check

In the most widely used services you will usually find options such as:

  • opting out of the use of conversations for training;
  • temporary chats that are not saved to your history;
  • deleting your history and, in some cases, your whole account;
  • memory management, if the service remembers information between conversations: it is worth checking periodically what it has stored.

Remember that deleting a conversation from your interface does not always mean it disappears immediately from the provider's systems: the terms of service state the retention periods.

Your rights

If you live in the European Union, the GDPR also applies to AI services that process your personal data. Among other things, you have the right of access to data concerning you and, where applicable, the right to erasure (Article 17). Providers must explain in their privacy notice how to exercise these rights.

Use at work

In a work context, the choice is not only personal. An employee who pastes a client's contract into a free service is effectively transferring data to a third party on the company's behalf. That is why it helps if the organisation:

  • states which tools are authorised;
  • prefers business plans with data processing agreements;
  • explains with concrete examples which information must not be entered;
  • offers an approved alternative, because banning without offering anything pushes people to use tools on the sly.

Attachments and integrations

Many assistants now read files and access email, calendars or shared documents. Every connection widens what the service can see. Before authorising one, ask yourself whether it is really needed and grant the minimum access required, for example to a single folder rather than your entire archive.

In short

  • What you write to a chatbot is sent to the provider and may be retained.
  • Do not enter credentials, banking data, health data, third-party data or business secrets.
  • Anonymise, use dummy data or describe the problem instead of copying real data.
  • Check the settings on training, history and memory.
  • At work you need authorised tools and clear rules, not just bans.
MM
Marco Marzaduri

Marco Marzaduri works in cybersecurity, digital forensics and web development. A computer science graduate, he combines rigorous technical analysis with attention to the people who use technology.

Read also