LLM Privacy & Compliance FAQ

Question 1

What is the minimum viable privacy setup for a startup using LLM APIs?

Accepted Answer

Week 1: API plan (not consumer) with DPA, enable Zero Data Retention, update privacy policy. Week 2: basic PII detection (Presidio is free), data classification guide, logging. Month 1: AI gateway/proxy, consent mechanisms, basic DPIA. Total additional cost: ~$0 beyond API subscription.

Question 2

What is the OWASP Top 10 for LLM Applications?

Accepted Answer

OWASP Top 10 for LLM (2025) identifies critical security risks. Privacy-relevant: Prompt Injection (extract training data), Sensitive Information Disclosure (LLM reveals PII), Excessive Agency (access beyond scope), Vector Weaknesses (RAG cross-tenant leaks), Misinformation (hallucinated PII).

Question 3

How do LLMs memorize personal data?

Accepted Answer

LLMs memorize through parameter encoding during training. Data appearing multiple times, unusual sequences (unique names), data near repetitive patterns (email signatures), and structured formats (addresses, phones) are especially prone to memorization. Larger models memorize more.

Question 4

What is machine unlearning and does it work?

Accepted Answer

Machine unlearning aims to remove specific data from trained models without full retraining. Current state (2026): still experimental, cannot be retrofitted to existing models, effectiveness difficult to verify. Dont rely on it as compliance strategy — prevent data from entering the model in the first place.

Question 5

Do wiretapping lawsuits affect developers using AI chatbots?

Accepted Answer

Yes, if you use third-party AI chatbots or conversation intelligence tools. Legal theory: when AI provider uses customer conversations to improve models, the company that embedded the chatbot may be liable for aiding and abetting wiretapping. Use providers that dont train on data and implement explicit consent.

Question 6

What is the conflict between US court orders and GDPR for AI companies?

Accepted Answer

The OpenAI/NYT lawsuit created direct collision: US court ordered preserve all chat logs (2022-2024) while GDPR requires right to erasure. OpenAI must simultaneously comply with US preservation order and EU deletion rights — which is legally impossible. Any AI company with US operations and EU users faces this conflict.

Question 7

Can a regulator force deletion of an entire AI model?

Accepted Answer

Theoretically, yes. EDPB Opinion 28/2024 states possible measures include erasure of the whole dataset used to develop the AI model and/or the AI model itself. In practice, regulators would first consider fines and corrective measures. No regulator has ordered model deletion yet.

Question 8

What was the OpenAI GDPR fine about?

Accepted Answer

In December 2024, Italy fined OpenAI €15 million — first GenAI GDPR fine. Violations: no legal basis for training on user data, lack of transparency, no age verification for minors, failed to report March 2023 breach, inadequate data protection. OpenAI is appealing.

Question 9

What should my privacy policy say about AI processing?

Accepted Answer

Minimum disclosures: that you use AI/LLM services, which providers, what data is sent, whether data is used for training, how users can opt out, retention practices, DPAs in place, and international data transfers.

Question 10

How do I protect documents in RAG systems from privacy leaks?

Accepted Answer

RAG creates specific risk: retrieved document chunks are sent to LLM as context. Protection layers: redact PII at indexing time, scan retrieved chunks before sending, scan LLM output for leaked PII. Also implement access controls on vector DB to prevent cross-tenant data leakage.

Question 11

When should I use a local model instead of a cloud LLM API?

Accepted Answer

Use local models for: medical records (PHI), financial data (PCI), government/classified data, legal documents, any data you cannot legally send to third parties. Use cloud APIs for: public data, after PII redaction, general knowledge queries.

Question 12

How does PII redaction work in practice for LLM APIs?

Accepted Answer

PII detector identifies personal data (names, emails), replaces with placeholders like [PERSON_1], sends redacted prompt to LLM, then de-redacts response for user. Use ML-based detection (Microsoft Presidio), not just regex — regex misses misspellings and contextual PII.

Question 13

What is an AI Gateway and do I need one?

Accepted Answer

An AI Gateway is a proxy layer between your app and LLM providers — like a firewall for AI traffic. It handles PII scanning, redaction, logging, rate limiting, and policy enforcement. You need one if you process PII through LLMs or need audit trails for compliance.

Question 14

Am I liable if my app sends user data to an LLM API that uses it for training?

Accepted Answer

Potentially, yes. Recent cases establish liability: Jones v. Peloton (2024) found company liable for third-party AI chatbot using data for training. Ambriz v. Google (2025) ruled AI a third party eavesdropper. You need DPA, Zero Data Retention, PII redaction.

Question 15

Can I use AI tools and still be GDPR compliant?

Accepted Answer

Yes, but it requires active effort: enterprise plans with DPA, established legal basis, updated privacy policy, DPIA conducted, user opt-out mechanism, data minimization, and configured retention policies.

Question 16

How do I create an AI usage policy for my company?

Accepted Answer

Essential elements: data classification (public/internal/confidential), approved tools list, prohibited actions (passwords, API keys, PII), approval process for new tools, incident response procedures, mandatory training, and usage monitoring.

Question 17

What is the difference between enterprise and personal AI plans for privacy?

Accepted Answer

Enterprise plans: no training on data, configurable retention, DPA available, SSO, audit logs, EU data residency, compliance certs. Personal plans: usually training enabled, provider-set retention, no DPA, no audit logs. The price difference pays for legal protection.

Question 18

What is Shadow AI and why should enterprises care?

Accepted Answer

Shadow AI is employees using personal AI accounts for work without IT approval. 47% of enterprise GenAI users use personal accounts, 77% paste corporate data. Samsung engineers pasted proprietary source code into ChatGPT. Data becomes subject to training policies with no audit trail.

Question 19

Is Microsoft 365 Copilot the safest enterprise AI option?

Accepted Answer

Among the safest for organizations in Microsoft ecosystem. Data never used for training, stays within M365 boundary, inherits existing security (Purview, Defender). But Microsoft can still access data, and third-party plugins may create new data flows.

Question 20

Why is Geminis 3-year data retention concerning?

Accepted Answer

Google retains reviewed Gemini conversations for up to 3 years — the longest among major providers. Combined with human review, your data could exist on Googles servers for 3 years and be reviewed by employees. Deletion requests take up to 72 hours.

Question 21

What changed in Claudes privacy policy in August 2025?

Accepted Answer

Anthropic made training opt-in (users must actively allow it). But if you opt-in, retention increased dramatically from 30 days to up to 5 years. If you opt-out, 30-day retention continues.

Question 22

How do I disable training on my data in ChatGPT?

Accepted Answer

Settings > Data Controls > Toggle OFF Improve the model for everyone. This prevents future conversations from training but doesnt affect past data. Data is still retained for 30 days and human review for safety is still possible.

Question 23

Does using voice mode or image uploads change privacy?

Accepted Answer

Yes, it adds more data types. Voice may be processed by third-party speech services. Images may contain EXIF metadata (location, device). Files like PDFs are processed in their entirety and may be stored longer.

Question 24

Can other users see my AI conversations?

Accepted Answer

Not intentionally, but bugs happen. March 2023: ChatGPT bug exposed chat titles and payment info. July-August 2025: Over 4,500 ChatGPT share links were indexed by Google. No provider guarantees bugs wont expose your data.

Question 25

What happens if I use Temporary Chat (ChatGPT) or Incognito mode (Claude)?

Accepted Answer

Conversations are not saved to history and not used for training. However, data is still retained on servers for ~30 days for abuse monitoring, and human review for safety is still possible.

Question 26

What is the safest AI assistant for privacy?

Accepted Answer

No cloud-based LLM is fully private. Claude (opt-in training) and M365 Copilot (enterprise, no training) are most private cloud options. Actually safest: self-hosted models via Ollama with Llama, Mistral, or Qwen.

Question 27

Do employees at OpenAI / Anthropic / Google read my conversations?

Accepted Answer

Potentially, yes. All major providers have some form of human review for quality, safety, and policy violations. Data is typically anonymized before review, but not always.

Question 28

Is there a difference between deleting a conversation and removing data from training?

Accepted Answer

Yes, completely different. Deleting conversation removes it from chat history but not from training. Disabling training prevents future use but doesnt affect past data. GDPR deletion removes stored data but not model weights.

Question 29

Can I delete my data from an AI model after it was used for training?

Accepted Answer

No. Once data is incorporated into model parameters during training, removing it is technically near-impossible. There is no undo button. Machine unlearning techniques are still experimental.

Question 30

Does ChatGPT / Claude / Gemini use my conversations to train models?

Accepted Answer

Yes, most do by default. ChatGPT, Gemini, and Perplexity use your data for training. Claude (Anthropic) switched to opt-in model in August 2025. Enterprise plans for all providers do not use data for training.