How to Research Sensitive Topics with AI Without Building a Profile That Follows You

The bottom line up front: Use Perplexity with privacy settings for web-sourced sensitive research, run Ollama locally for queries you would never say out loud, and store your notes in end-to-end encrypted storage. That three-layer stack keeps your sensitive research from becoming a data profile that follows you into insurance decisions, targeted ads, or a leaked breach.

Here is why this matters, and exactly how to set it up.

The Problem: AI Is a Perfect Research Partner With a Perfect Memory

If you have a new symptom, a legal dispute with a landlord, a question about a DUI from your past, or you are quietly researching whether to leave your job — AI is genuinely useful. It explains things clearly, does not judge you, and does not make you wade through SEO-poisoned search results.

But every one of those queries is stored somewhere.

ChatGPT logs your conversation history and uses it to improve future models unless you explicitly opt out — and even then, OpenAI retains logs for safety monitoring. Google Gemini ties queries to your Google account, which is already connected to your email, location history, and purchases. Claude keeps conversation history for 30 days by default.

The companies themselves are not the only risk. The real threat model for sensitive research is:

Data broker pipelines. AI companies share aggregate behavioral data with data brokers, who correlate it with your device fingerprint and sell profiles that inform insurance underwriting, employer background checks, and targeted advertising.
Breach exposure. Sensitive queries stored in plaintext on a server become sensitive queries in a breach. The health queries you made in a HIPAA-exempt consumer AI app are not protected the way your doctor's records are.
Legal discovery. Subpoenas reach cloud providers. Divorce attorneys and plaintiff lawyers know this. If your queries about "hiding assets," "criminal defense," or "leaving a company" are stored on a server, they can be compelled.
Profile-based discrimination. Insurance pricing algorithms ingest behavioral signals. A cluster of searches about chest pain, high-stress jobs, and medication costs does not have to be explicitly labeled "cardiac risk" to raise your rate.

None of this requires a conspiracy. It is just how data pipelines work in 2026.

What "Sensitive Research" Actually Covers

Sensitive research is anything you would hesitate to Google from your work laptop or say in front of a stranger. In practice, for tech workers this includes:

Medical and mental health: symptoms, diagnoses, medications, therapy options, addiction, fertility
Legal: criminal history, immigration status, workplace disputes, divorce proceedings, contracts
Financial: bankruptcy options, debt situations, hidden savings strategies, employment contract clauses
Relationship and identity: anything personal you are not ready to disclose publicly
Professional: job search while employed, salary negotiation, whistleblowing, IP ownership questions

The irony is that AI is most useful for exactly these categories, because they are the ones where you want clear explanations without paying $400/hour for a professional consultation. The goal is not to avoid using AI for this research — it is to use the right AI tool for the sensitivity level of the query.

Tier 1: The Easy Win — A Less-Tracking Search Layer

For general sensitive research that involves pulling from the web — understanding a legal concept, looking up drug interactions, researching a diagnosis — you do not need a local model. You need a research tool that does not build a persistent profile around your queries.

Perplexity AI takes a different approach from general-purpose AI chatbots. It is designed as a research and search tool rather than a persistent assistant. Queries are used to generate answers, not to build a long-term user model tied to your identity. The paid Pro tier adds enhanced privacy controls and does not use your conversations to train future models.

For sensitive research, Perplexity Pro's workflow advantage is real: you get cited, source-linked answers you can verify, without the query history accumulating in a profile the way it does in ChatGPT. Use it with a Proton Mail address (not your real email), pay with a privacy-preserving payment method, and you have substantially reduced your exposure for web-sourced research.

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.

The practical workflow: run your sensitive research session in Ollama or Perplexity, paste your notes into a local text editor, then upload to Proton Drive. If you are on macOS, the Proton Drive desktop app mounts as a local folder so you can save directly to it. Notes stay encrypted at rest and in transit.

Proton also bundles encrypted email (Proton Mail) and a calendar, which matters if your sensitive research connects to scheduled appointments or correspondence — therapy sessions, legal consultations, medical appointments. Keeping that thread end-to-end encrypted closes a significant surveillance gap.

Sharing Research With Professionals — Without Email Attachments

If your research ends with "I need to share this with my attorney / doctor / accountant," the standard move is email attachment. Email attachments are not end-to-end encrypted. They travel through servers you do not control, get stored in inboxes, and can be forwarded or accessed in a breach.

Tresorit solves this for professional document sharing. It is zero-knowledge encrypted cloud storage with a built-in share-link feature. You upload a document, generate a share link, set an expiration and password, and send just the link. The recipient downloads the document directly — no copy lives in your email thread, no copy lives on Tresorit's servers in readable form.

For sensitive documents — a medical history you are sharing with a specialist, a financial statement you are sending to a bankruptcy attorney, an employment contract you are reviewing with counsel — Tresorit's share links are meaningfully more private than "attach to email and send."

Tresorit also offers team workspaces, which matter if you are doing this research in the context of a business: a startup with a co-founder, a family business with a shared accountant, or a small legal or medical practice where multiple people need access to the same sensitive documents without exposing them to Google or Microsoft.

The Full Stack, Summarized

| Sensitivity Level | Tool | Why |

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

| General sensitive research (web sources needed) | Perplexity Pro + Proton email | Less profiling than Google/ChatGPT, cited answers |

| High-sensitivity queries (no cloud at all) | Ollama + local model | Nothing leaves your machine |

| Storing research notes | Proton Drive | Zero-knowledge encryption, can't be read by provider |

| Sharing with professionals | Tresorit share links | No unencrypted email attachments |

Set up Ollama in an afternoon. Start a Proton account (free tier covers most users). Add Perplexity Pro if you regularly do web-sourced research. The entire stack costs roughly $25/month — less than a single hour with any professional whose help you might be trying to avoid needing.

A Note on Threat Modeling

This setup is not bulletproof. Local models can be observed if someone has access to your device. Proton is subject to Swiss legal orders. Perplexity's privacy controls are a policy, not a technical guarantee.

But privacy is not binary. The relevant question is whether your sensitive research creates a persistent, accessible profile that can follow you through insurance underwriting, employer background checks, targeted advertising, or legal discovery. This stack dramatically reduces that risk compared to ChatGPT logged into your Google account on your work browser.

Do not let perfect be the enemy of private enough.

Have a sensitive research workflow that works for you? The strategies above are starting points — the right balance of convenience and privacy depends on your specific threat model. If you want a weekly dispatch on the tools and settings that actually protect your data without requiring a security engineering degree, the newsletter below is where that goes.

Get the PrivateAI Weekly →

One practical privacy technique, every Thursday. No fluff, no product spam.

Last updated: 2026-06-26