🔒 AI Security, Privacy & GDPR

This guide helps you use AI tools responsibly – with a focus on data protection, anonymization, and legal frameworks.

1 Why Does This Matter?

Every input into an AI tool is potentially:

  • Processed on external servers (OpenAI, Google, Anthropic, etc.)
  • Used for training (depending on the provider and settings)
  • Stored and logged (even after deletion)
flowchart LR
    A[Your Prompt] --> B[AI Server]
    B --> C{Processing}
    C --> D[Response to you]
    C --> E[Logging]
    C --> F[Training?]
    
    style E fill:#fff3e0
    style F fill:#ffebee

Rule of thumb: Treat every input like an email to a stranger. If you wouldn’t send it to an unknown person, don’t put it into AI.

2 What Does NOT Belong in AI Tools?

Red Line – Never Enter

Category Examples Risk
Personal data Names, addresses, birth dates, phone numbers GDPR violation
Health data Diagnoses, medications, medical records Art. 9 GDPR – special categories
Financial data Bank details, salaries, credit card numbers Identity theft
Credentials Passwords, API keys, tokens, SSH keys Security incident
Trade secrets Unreleased products, strategies, M&A plans Competitive disadvantage
Internal communications Confidential emails, Slack messages Breach of trust
Customer data Contracts, orders with real names GDPR + breach of contract

Yellow Zone – Only With Caution

Category When OK?
Anonymized data When all identifiers are truly removed
Public information When it’s freely accessible anyway
Own code (open source) When no secrets are included
Generic business processes Without reference to specific people/projects

Green Zone – Safe to Use

Category Examples
General questions “Explain concept X to me”
Public knowledge Wikipedia content, tutorials
Fictional sample data Made-up names and numbers
Your own creative texts Blog drafts, ideas without personal references

3 GDPR Basics for AI Usage

What Is the GDPR?

The General Data Protection Regulation (GDPR / DSGVO) governs the handling of personal data in the EU. It applies to every processing operation – even if the AI is hosted in the US.

Key Principles

Principle Meaning for AI Usage
Purpose limitation Use data only for the original purpose
Data minimization Enter only what’s necessary – no “context overload”
Storage limitation No permanent personal data in AI chats
Integrity & confidentiality Maintain technical safeguards
Accountability You must be able to prove GDPR compliance

To enter personal data into AI, you need one of these:

  1. Consent of the data subject
  2. Contract performance (processing is necessary for the contract)
  3. Legitimate interest (balanced against data subject rights)
  4. Legal obligation

In most cases: None of these legal bases allow you to simply enter personal data into ChatGPT.

flowchart TD
    A[Want to enter data into AI?] --> B{Personal data?}
    B -- No --> C[OK - not a GDPR issue]
    B -- Yes --> D{Legal basis available?}
    D -- No --> E[STOP - Do not enter]
    D -- Yes --> F{Data processing agreement in place?}
    F -- No --> E
    F -- Yes --> G{Data minimized?}
    G -- No --> H[Anonymize first]
    H --> I[Then enter]
    G -- Yes --> I
    
    style E fill:#ffcdd2
    style C fill:#c8e6c9
    style I fill:#c8e6c9

4 Anonymizing Data – Practical Guide

Why Anonymize?

Anonymized data is not subject to the GDPR. If you modify data so that no personal reference can be established, you can safely use it in AI tools.

Techniques Overview

Technique Description Example
Replacing Swap real names for fictional ones Max Müller → Person A
Generalizing Make specific values more general 28 years → 25-30 years
Removing Delete unnecessary fields Remove phone number
Masking Partially obscure data max@company.com → m**@c***.com
Aggregating Combine individual data into groups 5 salaries → Average

Step-by-Step: Clean Data Before AI Input

Checklist before every AI use:

Before pasting into an AI tool, check:

1. Does the text contain real names? → Replace with Person A, Person B
2. Does the text contain email addresses? → Replace with example@test.com
3. Does the text contain phone numbers? → Remove or replace with 555-XXXXXXX
4. Does the text contain addresses? → Replace with "Sample City, Region X"
5. Does the text contain company names? → Replace with "Company A", "Company B"
6. Does the text contain financial data? → Replace with realistic fictional numbers
7. Does the text contain credentials? → NEVER enter – remove completely

Before / After – Example

Before (DO NOT enter into AI):

Dr. Thomas Becker (born 14 March 1978) from Munich signed
contract #V-2025-4471 with Mueller GmbH on 12 January 2025.
Contact: t.becker@mueller-gmbh.de, +49 171 5559876.
Salary: 95,000 EUR/year.

After (anonymized – AI-ready):

Person A (age: 45-50) from [city in southern Germany] signed
a service contract with Company B on [date in Q1 2025].
Salary: between 90,000 and 100,000 EUR/year.

Prompt: AI as Anonymization Helper

You can use AI itself to anonymize texts – but only if the source text doesn’t contain highly sensitive data:

Anonymize the following text according to GDPR guidelines:

Rules:
- Replace all person names with Person A, Person B, etc.
- Replace company names with Company A, Company B, etc.
- Replace email addresses and phone numbers with placeholders
- Replace exact dates with time ranges (e.g., "Q1 2025")
- Replace exact amounts with ranges
- Remove all other identifying characteristics

Text:
[PASTE HERE]

Output the anonymized text and a table of replacements.

5 AI Providers and Their Privacy Settings

Overview of Major Providers

Provider Trains on data? Opt-out available? EU servers? DPA available?
OpenAI (ChatGPT) Yes (Free), No (Team/Enterprise) Yes (Settings) No (default) Yes (Enterprise)
Microsoft Copilot No (M365 Business) — Yes (EU tenant) Yes
Google Gemini Yes (Free), No (Workspace) Yes Partially Yes (Workspace)
Anthropic (Claude) No (API), Yes (Free) Yes No Yes (API)
Local models No — Yes (own server) Not needed

DPA = Data Processing Agreement (Art. 28 GDPR) – required when personal data is processed.

For every AI tool, verify:

1. Disable chat history / training (if possible)
2. Use business version (not the free tier)
3. Sign DPA with provider (for companies)
4. Enable EU data residency (if available)
5. Define team policies: What may be entered?

Decision Guide: Which Tool for Which Data? 

flowchart TD
    A[What data?] --> B{Sensitivity level}
    
    B -- Public --> C[Any AI tool OK]
    B -- Internal --> D{Personal data?}
    B -- Confidential --> E[Only local models<br/>or Enterprise AI]
    B -- Top secret --> F[Do not use AI]
    
    D -- No --> G[Business AI with<br/>training disabled]
    D -- Yes --> H[Anonymize first,<br/>then Business AI]
    
    style C fill:#c8e6c9
    style G fill:#fff9c4
    style H fill:#fff3e0
    style E fill:#ffecb3
    style F fill:#ffcdd2

6 Company Policies – Template

AI Usage Policy (Template) 

AI Usage Policy for [Company]

Effective: [Date]
Scope: All employees

1. PERMITTED USAGE
   - General research and knowledge questions
   - Text creation without personal references
   - Code assistance (without secrets)
   - General text translations

2. RESTRICTED USAGE (only after anonymization)
   - Analysis of internal documents
   - Summarizing meeting minutes
   - Processing customer communications

3. PROHIBITED USAGE
   - Entering passwords, API keys, tokens
   - Entering personnel files or job applications
   - Entering patient records or health data
   - Entering confidential strategy documents
   - Using personal AI accounts for company data

4. TECHNICAL MEASURES
   - Use approved AI tool: [Tool]
   - Chat history disabled by default
   - DPA signed with provider
   - Regular training (at least once per year)

5. VIOLATIONS
   - Report to: [Data Protection Officer]
   - Document the incident
   - Measures per Art. 33/34 GDPR in case of data breach

7 Checklist: Safe AI Usage

Before Every Input

  • Does the text contain personal data? → Anonymize
  • Does the text contain trade secrets? → Do not enter
  • Does the text contain credentials? → NEVER enter
  • Am I using the right tool (Business vs. Free)?
  • Is chat history / training disabled?

Regular Reviews

  • Is the AI usage policy known across the team?
  • Is the DPA with the AI provider current?
  • Have new employees been trained?
  • Has a Data Protection Impact Assessment been conducted (Art. 35 GDPR)?
  • Has the record of processing activities been updated (Art. 30 GDPR)?

In Case of an Incident

  • What data was entered?
  • With which provider?
  • Inform the Data Protection Officer
  • Report to supervisory authority within 72 hours (if required)
  • Notify affected individuals (if required)

8 Prompts for Data Protection Tasks

Create a Data Protection Impact Assessment (DPIA) 

You are an experienced data protection consultant.

Create a Data Protection Impact Assessment (DPIA) per Art. 35 GDPR
for the following initiative:

Initiative: [description of the AI use case]
Data categories processed: [e.g., customer data, employee data]
Tool used: [e.g., ChatGPT Enterprise]

DPIA structure:
1. Description of the processing
2. Assessment of necessity and proportionality
3. Risk assessment for data subjects
4. Measures to mitigate risks
5. Residual risk assessment
6. Recommendation (approve / adjust / reject)

Create a Record of Processing Activities Entry 

Create an entry for the record of processing activities per Art. 30 GDPR:

Processing activity: Use of AI tool [name] for [purpose]
Controller: [company]
Data subjects: [e.g., customers, employees]
Data categories: [e.g., names, email addresses]
Recipients: [AI provider]
Third-country transfer: [Yes/No – where?]

Format: Tabular, GDPR-compliant

Check Text for Privacy Issues 

Review the following text for data protection issues.

Identify:
- Personal data (names, addresses, etc.)
- Special categories (health, religion, etc.)
- Credentials or secrets
- Company names or project identifiers

Output a table:
| Location | Data type | Risk | Recommendation |
|----------|-----------|------|----------------|

Text:
[PASTE HERE]

9 Special Topic: Local AI Models

When data protection is the highest priority, locally installed models can be the solution:

Advantages

  • No data transfer to third parties
  • Full control over processing
  • No training on your data
  • GDPR-compliant without a DPA
Tool Description Hardware Requirements
Ollama Simple local model management 8+ GB RAM, optional GPU
LM Studio Desktop app with GUI 16+ GB RAM recommended
GPT4All Offline chat application 8+ GB RAM
llama.cpp Minimalistic C++ backend 4+ GB RAM

Quick Start with Ollama 

# Installation (macOS)
brew install ollama

# Download and run a model
ollama pull llama3.1
ollama run llama3.1

# Enter prompts – everything stays local

Local models are less powerful than GPT-4 or Claude, but sufficient for many tasks (summaries, anonymization, text review).

10 Summary 

flowchart TD
    A[AI Usage in Business] --> B[Create policy]
    A --> C[Train employees]
    A --> D[Technical measures]
    
    B --> B1[Permitted / Restricted / Prohibited]
    C --> C1[Learn anonymization]
    C --> C2[Know red lines]
    D --> D1[Use business tools]
    D --> D2[Disable training]
    D --> D3[Sign DPA]
    
    B1 & C1 & C2 & D1 & D2 & D3 --> E[Safe AI Usage]
    
    style E fill:#c8e6c9
    style A fill:#e3f2fd

The 5 Golden Rules

# Rule
1 No personal data without anonymization
2 No passwords, keys, or secrets – ever
3 Use business versions of AI tools
4 Disable chat history and training
5 When in doubt: Don’t enter it

⚠️ Disclaimer: This guide does not constitute legal advice. Consult your Data Protection Officer or a specialized law firm if in doubt.

Back to overview: 🏠 Home · Grundlagen (DE) · Fundamentals (EN)

Author: Justin Szczepaniak · LinkedIn

ProPrompt © 2026 – MIT License