Can You Upload That File to AI? A Privacy and Secrets Checklist

The short answer: if you cannot answer the data questions, do not upload the original

Identifiable personal data, company secrets and unpublished government documents should not be pasted directly into unapproved public AI tools. That applies even if you only want a summary, translation, rewrite or debugging help. If the input could reveal a person, customer, internal decision, credential or protected information, first remove or mask sensitive fields, turn the material into a non-identifying summary, or use an approved controlled environment.

The safer test is not the name on the AI product. It is whether four things are clear:

• what kind of data you are submitting;
• how the service keeps, uses or deletes it;
• whether your organisation explicitly allows that use;
• whether the activity can be monitored, traced and handled if something goes wrong.

NIST includes data protection, data retention, monitoring, incident response, opt-outs and risk-based controls in generative AI governance. If those points are unanswered, do not upload the original file.

Personal data, business secrets and government files: how to think about the risk

Five questions to ask before uploading anything

If you cannot answer even one of these, do not paste the original content into a public AI tool.

1. Does the content include personal data or sensitive information? If the material could identify someone or create a privacy risk, do not paste the original. The EDPB document is specifically concerned with privacy risks and mitigations in LLM-based systems.
2. Will the service retain the input or output, and for how long? NIST lists data retention as a generative AI risk-management issue.
3. Can the data be used commercially, further processed or used to improve the service? Is there an opt-out? NIST includes commercial use, data protection, data retention and opt-outs among governance topics.
4. Who can use the tool, and can use be traced? NIST refers to user credentials and qualifications, discouraging anonymous use and monitoring. In practice, an organisation needs to know who used the tool, for what purpose and with what kind of data.
5. Has the organisation planned for impact assessment, incident response and risk-based controls? These are also part of NIST’s generative AI risk-management guidance.

A sentence in your prompt saying “keep this confidential” is not a security control. What matters is whether the service retains the data, who can access it, whether reuse can be refused, who handles an incident and whether your organisation permits the upload.

Green, yellow and red: a practical upload guide

The following list turns data protection, retention and risk-based control principles into everyday decisions. It is not legal advice, and your own organisation’s security, privacy, legal and records-management rules should come first.

Green: usually lower risk, but still check the terms

You can consider using AI when the material is:

• already public, low sensitivity and something you are allowed to use;
• de-identified, stripped of sensitive fields or rewritten so it cannot reasonably point back to a person, customer, case or internal secret;
• reduced to the minimum background needed for the task, rather than a full contract, government file, customer spreadsheet or codebase.

Public does not always mean risk-free. If public material still contains personal or sensitive information, treat it under privacy and data-protection rules.

Yellow: rewrite, mask or get approval first

Pause before uploading material that includes:

• information about customers, employees, suppliers, case parties or members of the public;
• draft contracts, financial data, internal slide decks, meeting notes, legal opinions or policy drafts;
• source code, technical documentation or architecture diagrams, especially anything that may include keys, credentials or vulnerability details; NIST includes secure software development and risk-based controls in generative AI governance;
• internal government documents, unpublished official correspondence, approval notes, evaluation material or cross-agency working files; public-sector generative AI use still has to manage personal and sensitive-information risks.

These materials are not necessarily impossible to process with AI. But they should not be dropped into an unapproved public tool without retention rules, monitoring, incident response and an approval path.

Red: do not upload to a public AI tool

Do not upload:

• material that law or internal policy says must not be disclosed;
• classified documents or material involving national security, law enforcement, investigations, procurement evaluation or other highly sensitive work;
• passwords, API keys, private keys, certificates, access tokens or anything that can be used to enter a system;
• data for which you cannot confirm the source, permission to use it, retention rules, deletion rules or reuse conditions.

De-identification is more than deleting names

Removing a name is not enough if other details still point to the person or case. Identification numbers, phone numbers, email addresses, addresses, account numbers, case numbers, rare job titles, or unusual combinations of dates and locations can still create privacy risk. Because LLM privacy risks and mitigations are a central concern in the EDPB material, upload preparation should remove or rewrite identifying details, linkable clues and unnecessary fields together.

Safer options include:

• replacing real names and organisation names with labels such as Person A or Company B;
• providing only the minimum excerpt needed;
• rewriting the original file into an abstract scenario;
• aggregating lists, logs or tables before sharing them;
• using an approved organisational tool and process if the task genuinely requires original text.

Government documents: public data and internal files are not the same thing

Public-sector use of generative AI is not simply a yes-or-no question. The European Commission’s Joint Research Centre treats public-sector use as a specific area of discussion in its Generative AI Outlook report. A European Parliament annexed case summary also refers to use of official Bundestag data, while avoiding personal or sensitive information.

What may be suitable for AI is usually already public, low-sensitivity official information that can lawfully be reused. What needs a much more conservative approach includes unpublished correspondence, internal approvals, policy drafts, investigation material, enforcement material, procurement evaluation records and any document containing personal or sensitive information. The former still requires checking the conditions of use; the latter should not be pasted directly into a public AI service.

The simplest rule

If a leak would harm a person, an organisation, the public interest or legal compliance, do not give the original data to a public AI tool. Minimise it, mask it or summarise it first. If the original is truly necessary, use an approved controlled tool and confirm data protection, retention, access rights, monitoring and incident-response arrangements.