LIVE
Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|Anthropic says Houthi-linked actors used Claude Code for missile guidance software13/09/26 · Anthropic|Yoshua Bengio examines why AI agents lie, cheat and coordinate13/09/26|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|Anthropic says Houthi-linked actors used Claude Code for missile guidance software13/09/26 · Anthropic|Yoshua Bengio examines why AI agents lie, cheat and coordinate13/09/26|
Beginner🔒

AI and privacy: how to protect your data

Your ChatGPT conversations are not private. Here are the GDPR-friendly best practices to use AI without risking your data: settings, anonymization, local models.

12 min readPublished May 4, 2026· Updated September 17, 2026

In One Sentence

When you use ChatGPT, Claude or any AI, your data leaves your machine: it's processed on remote servers, sometimes stored, sometimes used to train other models. Here's how to stay in control without giving up on AI.

🚪
The analogy that works
Imagine each message sent to an AI goes through a door. That door leads to a large shared office where employees (the servers) process your request, can make a photocopy of it (storage), and sometimes use it to train new employees (model training). You can close certain doors, encrypt others, or bring your own door (local deployment).

🇫🇷 Want AI hosted in Europe?

Discover Mistral Le Chat: free, native French, servers in France, GDPR-friendly.

See European models

What happens when you send a message to an AI?

The lifecycle of a message sent to a chatbot

  1. You write your message

    The text is encoded and sent via HTTPS to the provider's servers.

  2. The model processes

    Your data passes through the LLM, generates a response.

  3. Temporary storage

    The conversation is generally kept for at least 30 days (logs, support, moderation).

  4. Possibility of training

    Depending on settings, may be used to improve future models.

The critical point: between steps 2 and 4, your data is readable by certain people (moderators, security teams, sometimes subcontractors).

The 3 risk levels

Which uses, which risks?

 Low risk🔴High risk
Brainstorming generic ideasOK
Writing a job application emailOK
Translating public textOK
Learning a technical conceptOK
Pasting a confidential client contractTo avoid
Asking advice on a medical fileTo avoid
Pasting personal data (customers)To avoid
Submitting sensitive proprietary codeTo avoid

GDPR and AI: what you need to know

If you're in Europe (or your users are), GDPR applies. Here are the 3 key principles:

You need a legal basis to process personal data with an AI. Common bases:

  • Explicit consent from the person
  • Legitimate interest (e.g. analysing your own emails)
  • Contract (e.g. commissioned customer support)

2. Transfers outside the EU

ChatGPT, Claude, Gemini: their servers are mainly in the US. EU→US transfers must be regulated (Data Privacy Framework, standard contractual clauses).

To stay in the EU: Mistral (French), Aleph Alpha (German), or self-hosted deployment.

3. Individual rights

If you use AI to process other people's data (customers, employees), they have rights:

  • Information: tell them their data goes through an AI
  • Access: let them see what's being processed
  • Deletion: be able to erase their data
The classic corporate trap
Pasting a PDF "List of my customers with their emails and phone numbers" into ChatGPT to "summarise it" is an unregulated personal data transfer. Real GDPR risk: possible sanction from the data protection authority up to 4% of turnover.

How to protect yourself: the 5 essential reflexes

📚Best practices in detail

1. Disable training in settings

ChatGPT: Settings → Data Controls → "Improve the model for everyone" → OFF

Claude: You can ask Anthropic via the support form, or directly use Team/Enterprise versions (which never use your data).

Gemini: myactivity.google.com → activate "Pause" on Gemini Apps Activity.

Mistral: Le Chat doesn't train on your conversations by default.

2. Anonymise before submitting

If you need to process text containing names/emails/phone numbers, replace them with placeholders:

  • "Jean Dupont" → "[CLIENT_A]"
  • "jdupont@gmail.com" → "[EMAIL]"
  • "06 12 34 56 78" → "[TEL]"

The AI will process the text normally. Then you reform it with a simple find/replace.

3. Use temporary conversations

ChatGPT and Claude offer a "temporary conversation" mode (the special icon in the chat) that doesn't save anything.

Use for: anything containing slightly sensitive info, even if not confidential.

4. Prefer Team/Enterprise offerings for professional use

Professional offerings (ChatGPT Team, Claude for Work, etc.) have contractual commitments:

  • No training on your data
  • Time-limited storage
  • GDPR compliance (DPA signed)
  • Security audit (SOC 2, ISO 27001)

It's ~£25-30/month/user, but it's mandatory for serious professional use.

5. For critical cases: local execution

You can run LLMs directly on your computer: your data never leaves.

Tools:

  • Ollama (ollama.com), install Mistral, Llama, DeepSeek in 1 command
  • LM Studio (lmstudio.ai), simple graphical interface
  • Jan (jan.ai), open-source alternative

Performance: a MacBook M2/M3 runs 7B-13B models very well. Not as powerful as GPT-5, but more than sufficient for 80% of uses.

The decision table: which tool for which use?

Choosing according to data sensitivity

 📝Low-sensitivity data🔒Sensitive data
Recommended toolChatGPT/Claude/Gemini freeMistral, Team/Enterprise, or local
SetupDisable trainingDPA + GDPR mandatory
OK usesBrainstorm, personal, publicCustomer data, medical, proprietary code
CostFree£25-30/month/user OR local

Providers compared on privacy

Privacy score (our assessment)

Ollama (local)100/100
Mistral Le Chat85/100
Claude (Anthropic)80/100
ChatGPT65/100
Gemini (Google)55/100

Methodology: score based on EU/US transfers, training opt-out, storage duration, GDPR certifications.

Case studies

📚5 common situations and the right answer

1. "I want to summarise a PDF received from a client"

If the PDF contains personal data (names, addresses, contacts) or trade secrets:

  • ❌ Not in free ChatGPT
  • ✅ ChatGPT Team/Claude for Work with signed DPA
  • ✅ Or locally (Ollama + Mistral)

If the PDF is public (market report, white paper):

  • ✅ Any tool

2. "I want ChatGPT to correct my CV"

The CV contains your name, email, background = your own data, but public (LinkedIn).

  • ✅ OK in any chatbot
  • ⚠️ Disable training anyway

3. "I want to ask a medical question"

  • ❌ Never paste a real medical report
  • ✅ Rephrase generically: "Someone with X symptoms, what might they have?"
  • ⚠️ And always consult a real doctor, AI is not qualified

4. "I want AI to help with my code"

If the code is proprietary (company):

  • ❌ Not in GitHub Copilot personal without employer agreement
  • ✅ GitHub Copilot Business (no-training commitment)
  • ✅ Or Cursor/Claude Code Enterprise

If the code is open source or personal:

  • ✅ All tools OK

5. "I want to scrape a website and summarise with AI"

  • ⚠️ Check the site's ToS (often prohibited)
  • ⚠️ Don't use scraped personal data
  • ✅ For truly public data (Wikipedia, institutional sites), it's OK

The classic mistake: the "it stays between us"

No, it doesn't stay between you
Many think a conversation with a chatbot is "like talking to a therapist": confidential by default. This is false. Without precautions: - The conversation can be read by moderators - Stored for at least 30 days - Used to train future models - Available to authorities on legal request (subpoena) Golden rule: everything you write in a chatbot, treat it as if you were saying it aloud in a crowded café. If you'd be embarrassed for someone to hear it, don't type it.

The metaphor that sums it all up

📮
The digital postcard
Sending a message to a chatbot is like sending a postcard. It's not private: the postman can read it, the sorting centre can photocopy it, the archivist can study it 30 years later. If you want privacy, you need a sealed letter (end-to-end encryption, rare in AI), a trusted courier (Enterprise offering with DPA), or keep the message at home (local execution).

Essential takeaways

  • Disable training in the settings of all your chatbots
  • Anonymise names/emails/phone numbers before submitting
  • For professional use: Team/Enterprise offerings mandatory
  • For critical use: local execution (Ollama, LM Studio)
  • Never: passwords, medical data, industrial secrets in clear text

AI is a fantastic productivity tool, but not a confessional. Learn the right reflexes, and you can enjoy it with peace of mind.

🧠 Quiz
Question 1 of 3

What happens with your conversation by default on free ChatGPT?

Going further

Tags
ConfidentialitéRGPDDonnéesSécurité

Read next