AI Safety: What Data Should You Never Share with AI?

What should you never paste into AI?

Do not enter the following categories of information into publicly available AI models:

• personal data – your own or other people’s names, addresses, national ID or Social Security numbers, phone numbers and email addresses;
• medical records – medical histories, test results and diagnoses, which are subject to strict legal protections such as the GDPR and HIPAA;
• confidential business information – unpublished financial reports, marketing strategies, customer databases, and plans for mergers, acquisitions or layoffs;
• trade secrets – software source code, production formulas, proprietary algorithms and internal technical documentation;
• passwords and credentials – API keys, login details, authentication tokens, payment card details and encryption keys of any kind;
• copyright-protected material – complete books, paywalled research articles or scripts that you are not licensed to share;
• illegal content – material promoting violence or hatred, or instructions for unlawful activity.

What happens to data entered into AI?

Data sent to cloud service providers can go through several stages of processing and analysis. Its lifecycle does not necessarily end when the tool generates a response.

Model training

Information entered into standard consumer versions of generative AI tools may be used to improve their models. Depending on the provider and your settings, submitted information could influence a model’s parameters and, in theory, appear in a response to another user in the future.

Human review and moderation

Alongside automated processing, AI systems use safety filters. If a query is flagged as suspicious, such as a potential breach of the terms of service, the conversation may be sent for human review. Provider staff or external reviewers, sometimes called data labelers, could then see the submitted text while working to improve moderation.

Data storage and backups

AI providers store conversation records on their servers to maintain services, diagnose problems and retain context for future sessions. Deleting a chat through the user interface does not necessarily remove it from the provider’s servers immediately. Information may remain in backups for a period of time, increasing the risk that confidential material could be exposed if the provider’s cloud infrastructure is breached.

Personal versus business accounts: How does protection differ?

Free and standard paid personal accounts on many popular services may use submitted information for model improvement by default. If you share confidential material, you risk losing control over where it goes.

Business and enterprise environments – such as Enterprise accounts or services deployed in private cloud environments through APIs from providers including Microsoft Azure, AWS and Google Cloud – typically offer stricter protections. Providers may commit to security standards such as ISO 27001 and SOC 2, support GDPR compliance, and state in their agreements (SLAs and DPAs) that customer data will not be used to train public foundation models. These arrangements may allow greater flexibility over what you enter, but they still require risk management policies and access controls.

What can you safely paste into AI?

Despite these restrictions, many types of information can generally be shared with AI tools, including:

• publicly available material – open-access articles, blog posts, public market reports, legislation and website content, such as Wikipedia pages;
• non-confidential information – general instructions, questions about theoretical concepts, grammar and spelling rules, mathematical equations and dictionary queries;
• anonymized excerpts from documents – letter templates, sample contracts or code snippets with all identifying variables, keys, client names and details of internal architecture removed;
• your own creative drafts – draft emails, social media posts, presentation outlines or articles that contain no sensitive business information;
• open datasets – synthetic data or statistics from public repositories used to practise analysis and formatting.

How can you anonymize messages and data before sending them to AI?

Removing identifiers and replacing them with placeholders

One approach is to replace sensitive details with artificial placeholders. For example, change the name “Alex Smith” to “[Client_1]” or the company name “Acme Ltd” to “[Organization_A]”. This preserves the document’s structure, syntax and context for the language model while helping to prevent identification of the original person or organization.

Masking sensitive information

Masking hides critical strings, often by replacing characters with symbols, as in the card number 4532 **** **** ****. Another useful technique is generalization: instead of giving an exact figure, such as a monthly salary of $4,500, use a range, such as “monthly salary of $4,000–5,000”. This reduces the risk of re-identification.

Automating the process with DLP tools and RegEx

Removing sensitive information manually from long documents makes omissions easy. Professional workflows may use regular-expression (RegEx) scripts or DLP (Data Loss Prevention) tools. These can scan a prompt before it is sent, then flag, block or replace strings that match formats such as national ID numbers, tax or VAT numbers, email addresses and bank account numbers.

How can you stop AI providers from training on your data?

Configuring the tool itself can also reduce the risks associated with sharing information. Most providers offer a way to opt out of using your content for model training.

• ChatGPT (OpenAI) – under Settings → Data controls, turn off “Improve the model for everyone”. This stops eligible future conversations from being used to improve models without removing them from your visible chat history. OpenAI may still retain data for a limited period for safety and abuse monitoring; check its current retention policy for the account and feature you use.
• Gemini (Google) – turn off “Gemini Apps Activity” in your privacy settings. New conversations will not be saved to that activity or used to train models under this setting. Google may still retain them for up to 72 hours to maintain service safety.
• Claude (Anthropic) – check the privacy settings for your consumer account and turn off “Help improve Claude” if you do not want eligible conversations used for model improvement. Anthropic’s default data-use settings have changed over time, so review the current terms for your account.

Remember that privacy-setting changes and training opt-outs generally apply to future data use. They may not reverse the use of information already incorporated into a model’s training.