Privacy-First AI: How to Use ChatGPT, Claude, and Gemini Without Sharing Your Data

The privacy problem with AI chat tools

Every time you type a message into ChatGPT, Claude, or Gemini through their web apps, that conversation is stored on their servers. Your prompts, the AI's responses, any files you upload — all of it lives on infrastructure you don't control.

For many users, this is a real concern:

• Developers paste proprietary source code into AI chats for debugging help
• Lawyers and consultants discuss confidential client matters
• Writers share unpublished manuscripts and ideas
• Researchers work with sensitive datasets and findings
• Business teams discuss internal strategy, financials, and product plans

Each of these use cases involves data that probably shouldn't live on a third-party server indefinitely.

What happens to your data in traditional AI tools

Here's how the major AI platforms handle your conversation data:

OpenAI (ChatGPT)

By default, OpenAI stores your conversations and may use them to improve their models. You can opt out of training data usage in settings, but your conversations are still stored on their servers for up to 30 days (or longer for abuse monitoring).

Anthropic (Claude)

Anthropic states they don't use consumer conversations for training by default. Conversations are stored on their servers and retained according to their data retention policy.

Google (Gemini)

Google stores Gemini conversations and may use them for product improvement. Conversations may be reviewed by human annotators. You can manage and delete conversation history, but it's stored on Google's infrastructure.

The common thread

In every case, your data leaves your device, travels across the internet, and lands on servers managed by someone else. Even with good privacy policies, you're trusting that:

• The company won't change their policy
• Their servers won't be breached
• Employees with access won't misuse data
• Government requests won't expose your conversations

What local-first AI means

Local-first is an architecture principle: your data is stored on your device first, and only leaves when you explicitly send it somewhere.

In a local-first AI workspace like NovaKit, this means:

• Conversations are stored in your browser's IndexedDB — a database on your machine
• API keys are encrypted and stored locally — never sent to NovaKit's servers
• Settings, prompts, and knowledge base documents stay on your device
• The only outbound requests are the AI API calls you initiate — and those go directly to the provider

How NovaKit protects your data

Client-side encryption for API keys

Your API keys are the most sensitive data in any AI tool. In NovaKit, they're encrypted with AES-256-GCM before being written to storage.

If you choose to set an optional passphrase, the encryption key is derived from your passphrase using PBKDF2 with 100,000 iterations. This means:

• Without your passphrase, stored keys cannot be decrypted
• NovaKit never sees your passphrase or your unencrypted keys
• Even if someone accesses your browser's storage, the keys are encrypted ciphertext

Direct API calls — no proxy server

When you send a message in NovaKit, here's the data flow:

1. Your browser constructs the API request (message + model selection + any context)
2. Your browser sends the request directly to the AI provider's API endpoint
3. The AI provider sends the response directly back to your browser
4. The conversation is stored locally in your browser's IndexedDB

NovaKit's servers are never in the middle. We don't proxy, log, cache, or inspect your requests. We literally can't see your conversations because they never touch our infrastructure.

Zero telemetry on usage

NovaKit doesn't track:

• What models you use or how often
• How many messages you send
• What topics you discuss
• How long your conversations are
• What files you upload or attach

We don't know these things because we never collect them.

Browser-level data isolation

Each browser profile has its own isolated IndexedDB storage. This means:

• Your work browser and personal browser have separate NovaKit data
• Different browser profiles can have different API keys and conversations
• Incognito/private windows don't have access to your stored data
• Clearing browser data removes all NovaKit data from that profile

The trade-offs of local-first AI

Local-first architecture is a deliberate choice with trade-offs. It's important to understand what you gain and what you give up:

What you gain

• Complete data ownership — Your conversations are yours, period
• No server breaches — Can't breach data that doesn't exist on a server
• No policy changes — Your privacy doesn't depend on a company's terms of service
• Offline access — Past conversations are available without an internet connection
• No account required — Use NovaKit without creating an account or providing an email

What you give up

• No cross-device sync (currently) — Your data lives in one browser on one device. We plan to add optional end-to-end encrypted sync in the future.
• Browser data vulnerability — If you clear your browser data or the browser corrupts its storage, your data is gone. Regular exports are recommended.
• No server-side search — Search happens locally, which is fast for most users but doesn't scale to thousands of conversations as well as server-side search.

Who needs privacy-first AI?

Software developers

If you paste code into AI for debugging, refactoring, or code review, that code is now on someone else's server. For proprietary codebases, this can violate NDAs, employment agreements, or compliance requirements.

With local-first AI, the code goes directly to the API provider and is stored only on your machine.

Professionals handling confidential data

Lawyers discussing case details, consultants working on client strategy, healthcare professionals referencing patient scenarios — all of these involve data that has regulatory or ethical constraints on where it can be stored.

Anyone who values digital autonomy

You don't need a specific reason to want privacy. The principle that your conversations are your business — not a training dataset for an AI company — is reason enough.

How to evaluate AI privacy claims

When evaluating any AI tool's privacy claims, ask these questions:

1. Where is my data stored? On their server, or on my device?
2. Can they read my conversations? Even if they say they won't, can they?
3. What happens if they're acquired or shut down? Where does my data go?
4. Can I delete my data completely? Or does it persist in backups and logs?
5. Do they use my data for training? Even "anonymized" data can be problematic.

The strongest privacy guarantee is architectural: if the tool never has your data, it can't misuse, lose, or expose it.

Getting started with private AI

1. Open NovaKit — No account required for the free tier
2. Add your API keys — Encrypted locally, never sent to our servers
3. Start chatting — Conversations stored in your browser only
4. Export periodically — Back up your conversations as JSON files

Your AI conversations should be as private as the thoughts that prompt them.