Data Privacy in AI Systems: What Gets Sent Where

Every word you type into ChatGPT, every document you upload to Claude, every voice command to Siri — it all goes somewhere.

And somewhere is not a vague concept. It's a specific server. In a specific country. Run by a specific company. Subject to specific laws.

Most people treat AI like a private conversation. It's not. It's a data transmission. And if you're building AI products, you need to understand exactly what that means.

The Data Journey: From Your Keyboard to a GPU

Let's trace what happens when you type "Write me a resignation letter" into an AI app:

Step 1: Your Device

The text sits in your browser or app. So far, so private. If you're using a web app, it's just local JavaScript.

Step 2: TLS Encryption in Transit

Your data travels over HTTPS. TLS 1.3 encrypts it from your device to the server. Nobody in between can read it — not your ISP, not the coffee shop Wi-Fi owner.

Step 3: The API Provider's Servers

Your data hits OpenAI's, Anthropic's, or Google's servers. Here's where it gets interesting:

Provider Data Storage Retention Training Use

Step 4: The Model Processes It

Your data goes into GPU memory. It's processed, the response is generated, and then... what happens to the input?

API vs Consumer: Two Different Privacy Worlds

This is where people get confused. There are TWO completely different privacy regimes:

Consumer Apps (ChatGPT, Claude.ai, Bard)

Data may be used for training

OpenAI and Anthropic use consumer data to improve models

You can opt out, but it's not the default

Free tier users have less control

API Access (What Developers Use)

Data is NOT used for training by default

Enterprise agreements guarantee this

Data retention is limited (usually 30 days for abuse monitoring)

You pay, so you get privacy

The rule: If you're not paying money for API calls, your data is probably helping train next year's model.

🔧

PII Detection: Finding the Needles in the Haystack

PII = Personally Identifiable Information. Names, emails, phone numbers, Aadhaar numbers, credit cards, medical records.

When you're building AI apps, PII leaks in two directions:

Inbound: User Sends PII to AI

Users paste everything into chat boxes. Resumes, medical reports, legal documents. They don't think about where it goes.

Detection approaches:

python

Option 1: Regex patterns (fast, brittle)

import re

aadhaar_pattern = r'\d{4}\s?\d{4}\s?\d{4}'

Option 2: NER models (smarter)

from presidio_analyzer import AnalyzerEngine

analyzer = AnalyzerEngine()
results = analyzer.analyze(text=text, language='en')

Option 3: LLM-based detection (most accurate, slowest)

detection_prompt = """

Identify all PII in this text. Return as JSON.

Categories: name, email, phone, address, credit_card, ssn, medical_info

"""

Outbound: AI Returns PII from Training Data

Models can memorize training data. Rare, but documented. GPT-4 has been shown to regurgitate exact email addresses and phone numbers from its training set.

Mitigation:

Never train on raw PII

Use differential privacy for fine-tuning

Implement output filters

📊

Data Residency: Where Your Data Sleeps

Data residency laws are exploding. Here's what matters:

Regulation Region Key Requirement

GDPR EU Data can leave EU only to "adequate" countries

DPDP Act India Sensitive personal data must stay in India

PIPEDA Canada Consent required for cross-border transfer

China PIPL China Critical data must stay in China

What This Means for AI Apps

If you're serving European users:

Use Azure OpenAI with EU regions

Or use European providers (Mistral, Aleph Alpha)

Sign Data Processing Agreements (DPAs)

If you're serving Indian users:

Store data in Indian data centers

AWS Mumbai, Azure India Central, GCP Mumbai

Get explicit consent before sending to foreign AI providers

Compliance Checklist for AI Builders

Building an AI product that handles user data? Here's your baseline:

1. Privacy Policy (Required)

What data you collect

Where it goes (specific providers)

How long you keep it

Whether humans review it

1. Data Processing Agreement (DPA)

Required for GDPR, DPDP. Your agreement with OpenAI/Anthropic/etc that specifies:

They act as data processor

You are data controller

Subprocessors are listed

1. User Consent Mechanism

Explicit opt-in before AI processing

Granular consent (what specifically are they agreeing to?)

Easy withdrawal

1. Data Retention Limits

Delete user data after X days

Delete AI conversation logs

Anonymize analytics data

1. Security Measures

Encryption at rest and in transit

Access logging

Regular security audits

Red Flags: When to Worry

🚩 "We use AI to improve our services" — Vague. Ask specifics.

🚩 Free AI features with no privacy policy — You're the product.

🚩 Sending data to multiple providers without disclosure — Compliance violation.

🚩 Storing full conversation logs indefinitely — Liability nightmare.

🚩 No DPA with your AI provider — GDPR/DPDP violation waiting to happen.

🔬

Practical Takeaways

Consumer AI apps train on your data — API access does not, by default

PII detection should happen BEFORE sending to AI — Don't rely on the model to be careful

Data residency laws are real — EU data in EU, Indian data in India, or face fines

DPAs are non-negotiable for business use — Get them from every AI provider

Retention limits protect everyone — Set auto-deletion for conversation data

Output can leak training data — Implement post-processing filters for PII

🛡️

What's Next?

Episode 55: Cost Optimization — Caching, model selection, batch processing, prompt compression, and token reduction. Because privacy is important, but so is not burning cash on unnecessary inference.

← Previous

Ep 53: Sandboxing AI Agents

Next →

Ep 55: Cost Optimization

Next: Episode 55 — Cost Optimization

A side project costs ₹500/month. At scale, ₹5,00,000/month. The code doesn't change. The usage does. Here's how to fix that.

This is part of a 98-episode series covering AI engineering from tokens to production deployment.