Why ChatGPT's Privacy Settings Don't Fully Protect You

Before building the alternative, it's worth understanding what you're actually opting out of — and what you aren't.

OpenAI offers "Temporary Chat" mode and a setting to disable conversation history. These prevent prompts from being saved to your account. They do not prevent OpenAI from logging queries for safety monitoring, abuse detection, or legal compliance. Enterprise plans add stricter contractual guarantees around data handling, but they start at $30 per user per month — most developers aren't paying for Enterprise.

If you're on a standard ChatGPT Plus account, the practical reality is: OpenAI has technical access to every prompt you've ever sent, regardless of your training data preferences.

For most casual use cases, that's an acceptable trade-off. For developers handling client code under NDA, security researchers, or anyone working with proprietary data, it isn't.

The Private AI Stack: Four Layers, Four Jobs

The stack below replaces ChatGPT's core functions with tools that either process data locally or operate under significantly stricter data policies:

|-------|----------|------|------|

Total: roughly $35–$45/month versus ChatGPT Plus at $20/month. You're paying a bit more — but you get full local AI capability, private web research, zero-knowledge file storage, and encrypted email. The capability jump is substantial.

Layer 1: Ollama + Open WebUI for Daily AI Work

For the 80% of tasks people use ChatGPT for — writing, code review, summarization, brainstorming — a local LLM is faster, free, and completely private. Your prompts never leave your machine.

What you need:

A machine with at least 8GB RAM (16GB recommended for larger models)
Ollama — open source local LLM runtime
Open WebUI — a ChatGPT-style browser interface for local models

Setup in under 15 minutes:

```bash

Install Ollama

curl -fsSL https://ollama.ai/install.sh | sh

Pull a model (Llama 3.1 8B — 4.7GB, runs on 8GB RAM)

ollama pull llama3.1:8b

Launch Open WebUI via Docker

docker run -d -p 3000:80 --add-host=host.docker.internal:host-gateway \

-v open-webui:/app/backend/data \

ghcr.io/open-webui/open-webui:main

```

Open localhost:3000 and you have a functional ChatGPT interface backed by a model that runs entirely on your hardware.

Which model to run:

Llama 3.1 8B — Best all-around for 8GB RAM. Strong at writing, coding, and summarization.
Qwen2.5 14B — Step up in quality if you have 16GB RAM. Excellent at structured data and reasoning tasks.
DeepSeek-Coder-V2 — Purpose-built for code. Competitive with GPT-4o on many benchmark coding tasks.

The limitation local LLMs can't solve: they don't know what happened after their training cutoff. For anything requiring current information, that's what Layer 2 is for.

Layer 2: Perplexity for Web-Aware Research

When you need current information — a newly published CVE, today's framework release notes, recent case law — a local LLM will either hallucinate or admit it doesn't know. Neither is useful.

Perplexity Pro is a research-focused AI that retrieves and synthesizes information from the live web, then cites every source. You get a direct answer with verifiable links rather than ten blue links to sort through yourself.

From a privacy standpoint, Perplexity is meaningfully better than using Google plus ChatGPT together:

No advertising business model. Perplexity doesn't sell your queries to ad networks. Their revenue comes from subscriptions, not behavioral targeting.
Source transparency. Every answer includes citations, so you can verify claims rather than trusting a black box.
Focus modes. You can restrict searches to academic papers, Reddit, GitHub, YouTube, or news sources, which reduces noise and keeps queries more contained.

Perplexity does log queries for abuse prevention — it's not zero-knowledge like a local LLM. But for research into technical topics, it's far more contained than pasting your questions into Google's search bar, where they feed directly into your advertising profile.

Use Perplexity for: real-time technical research, security vulnerability checks, scanning industry news, and any task that requires knowing what happened in the last 90 days.

Use your local LLM for: everything that doesn't require current events.

Affiliate Disclosure: This article may contain affiliate links. If you make a purchase through these links, we may earn a small commission at no extra cost to you. We only recommend products we genuinely believe in. This helps support our work and allows us to continue providing free content.

Layer 3: Tresorit for Storing Sensitive AI Outputs

Here's a privacy failure mode most people miss entirely: you use a local LLM to process a sensitive document, generate a clean output, and then save that output to Google Drive. Your source document never left your machine — but your AI-generated analysis did.

Tresorit closes this gap. It's a zero-knowledge encrypted cloud storage service incorporated in Switzerland, meaning Tresorit cannot read your files even if compelled by a court order. Files are encrypted client-side before they leave your device. Only you hold the decryption keys.

For a private AI workflow, Tresorit serves two roles:

Storing AI outputs and working documents. When your local LLM processes client contracts, code reviews, or internal research memos, save those outputs to a Tresorit folder instead of Google Drive or iCloud. The content is encrypted before it uploads.

Syncing your AI workspace across devices. Open WebUI conversation exports and Ollama model configurations can be stored in Tresorit and synced between your work machine and personal laptop without routing through a server that can read your files.

Tresorit is SOC 2 Type II certified and GDPR-compliant. For developers handling client data under NDA, it provides a defensible audit trail: your storage provider genuinely cannot access your files, which holds up in contract disputes and data handling reviews.

The free tier gives you 10GB — enough for a year of AI output files. The Business plan (~$15/month) expands to 1TB and adds folder-level access controls for team sharing.

Layer 4: Proton for Project Communications

The final piece is communications. If you collaborate with clients or teammates on AI-assisted work, email is a significant exposure surface. Standard Gmail routes through servers Google can read and uses your email content to inform its advertising and AI training systems.

Proton Mail provides end-to-end encrypted email hosted under Swiss privacy law. Proton cannot read your emails — messages between Proton users are automatically end-to-end encrypted. Messages to non-Proton users are encrypted at rest on Proton's servers and can be sent with optional password protection.

For this stack, Proton's role is:

Sending AI-generated deliverables to clients without those emails routing through Google's infrastructure
Using Proton Drive (included with paid plans) as a secondary file exchange when clients can't use Tresorit
Keeping AI project communications separate from your personal or employer-monitored email account

Proton's free tier covers basic email. Proton Plus at $9.99/month adds 15GB storage, custom domains, and Proton Drive — a reasonable option if you prefer consolidating storage and email with one privacy-focused provider. For heavier file sync needs, the combination of Proton Mail and Tresorit is more capable than either alone.

What a Typical Workday Looks Like With This Stack

Here's a concrete workflow for a developer handling client code under NDA:

Morning: Use Perplexity to check for new vulnerabilities in the libraries your client's project depends on. No query history stored in your ad profile. No Google autocomplete learning your client's tech stack.

Code review session: Paste client code into Open WebUI, running against your local DeepSeek-Coder model. The code never leaves your machine. Quality is comparable to ChatGPT-4o for standard review tasks.

Document output: Ask the local LLM to draft a technical summary from your notes. Save to a Tresorit "Client Work" folder. The file is encrypted before it leaves your drive.

Client delivery: Send the summary via Proton Mail with a Tresorit shared folder link. Your client downloads an encrypted file. You have a verifiable chain of custody if questions arise later about data handling.

Zero prompts logged by OpenAI. Zero working files on Google's servers.

The One Gap: Multimodal and Real-Time Voice

Local LLMs are closing the gap with proprietary models quickly. Two areas where ChatGPT still leads: high-quality image generation and deep voice integration.

If your work relies heavily on DALL-E or ChatGPT's voice mode, you'll still need a ChatGPT account for those specific sessions. The mitigation: use Temporary Chat mode, keep a throwaway account for those tasks, or evaluate whether those specific features are actually load-bearing in your workflow.

For the majority of developers and privacy-focused professionals, this stack handles 90–95% of real ChatGPT use cases — and the 5–10% that's missing is almost entirely in creative multimedia tasks, not daily knowledge work.

Building the Stack This Weekend

Total setup time is roughly 90 minutes:

Install Ollama and Open WebUI — 30 minutes including model download
Start a Perplexity Pro trial — 5 minutes
Create a Tresorit account and configure a sync folder — 20 minutes
Set up Proton Mail for project communications — 30 minutes

After that, the workflow is lower-friction than it sounds. Open WebUI looks and feels like ChatGPT. Tresorit works like Dropbox. Proton Mail works like Gmail. You're replacing the interface with private equivalents, not rebuilding your entire workflow from scratch.

The question isn't whether you need AI tools — you do. The question is whether you're comfortable with where your prompts go after you hit send.

Last updated: 2026-05-31

Want to lock down your AI outputs? Start with a free Tresorit account — zero-knowledge file storage that closes the most common gap in a local AI setup.

Get the PrivateAI Weekly

One actionable privacy setup guide per week. No noise, no surveillance capitalism pitch.

Get the Private AI Toolkit

Weekly: one setup guide, one tool review, zero telemetry lectures.