How to Protect Sensitive Data When Using ChatGPT, Claude, and Gemini (2026 Guide)
Team Sonomos · Sonomos
The Sonomos research team writes about AI privacy, data protection, and how to use generative AI safely at work.
When you paste a contract, a customer email, or a patient note into ChatGPT, Claude, or Gemini, that text leaves your device. Once it leaves, you no longer control where it is logged, who can review it, or whether it ends up shaping a future model. For most professionals, that is the single biggest privacy risk introduced by generative AI — and the one most often overlooked.
This 2026 guide explains, in plain English, how to keep sensitive data out of large language models (LLMs) without giving up the productivity gains AI provides. It covers the policies that actually matter, the settings to change today, and the workflow patterns that prevent leaks before they happen.
What counts as "sensitive data" in an AI context?
Sensitive data is any information that could harm a person, a customer, or your organization if it were exposed. In the context of AI prompts, the most commonly leaked categories are:
- Personally identifiable information (PII): names, addresses, phone numbers, government IDs, dates of birth.
- Protected health information (PHI): anything covered by HIPAA in the United States or equivalent regimes elsewhere.
- Financial data: account numbers, card numbers, balances, tax filings.
- Trade secrets and source code: proprietary algorithms, internal architecture, unreleased product plans.
- Privileged communications: legal advice, M&A documents, HR investigations.
- Authentication material: passwords, API keys, session tokens, recovery phrases.
If a prompt contains any of the above, treat it as a regulated workload — even if the AI tool is "free."
How LLM providers actually use your prompts in 2026
The answer depends on which product you use and which tier you pay for. As of November 2026, the publicly documented defaults look like this:
- OpenAI ChatGPT: Free, Plus, and Pro prompts may be used for model training unless you opt out in Data Controls. ChatGPT Business (formerly Team), Enterprise, Edu, and API traffic are excluded from training by default.
- Anthropic Claude: since September 2025, Free, Pro, and Max accounts carry a model-improvement setting; chats are used for training when it is on, and retained for up to five years (30 days when it is off). Claude for Work and the API are excluded by default. Conversations flagged by safety systems can be retained for up to two years.
- Google Gemini: Free-tier conversations can be reviewed by humans and used to improve the product. Workspace and Vertex AI traffic are governed by enterprise terms.
- Microsoft Copilot: when you sign in with a work or school (Entra) account, enterprise data protection applies and prompts and responses are not used to train foundation models. On a personal Microsoft account, Microsoft says the Copilot app released on August 18, 2026 does not use prompts, responses or files to train foundation models; the older app trained by default unless you opted out.
These policies change. The safer assumption is that anything you send may be retained, may be reviewed by a human, and may be subpoenaed.
Step 1: Change the settings that exist today
Before installing anything new, harden the tools you already use. The settings below take less than five minutes total.
ChatGPT
- Open Settings → Data Controls.
- Turn off Improve the model for everyone.
- Use a Temporary Chat for anything you do not want in your history; Temporary Chats are not used for training and are deleted within 30 days, with a copy reviewable only for abuse in the meantime.
- For Enterprise and Edu workspaces, ask your admin what retention period is set; Business has no custom retention setting, so chats stay until deleted. Training is off by default on all three.
Claude
- Claude.ai now asks every Free, Pro, and Max user to choose whether chats may be used to improve Claude; the choice lives in Settings → Privacy.
- Review the Model Improvement setting there and turn it off if you do not want your chats used for training. With it off, the 30-day retention period applies; with it on, chats can be kept for up to five years.
- For Claude for Work, ask your administrator to confirm the SOC 2 Type II report and data retention window. What each Claude tier does and does not make confidential is set out in Is Claude Confidential?.
Gemini
- In the Gemini app, open Settings → History (formerly Keep Activity).
- Check that the boxes letting Google train on your chat content and on audio and Live recordings are unticked. Google says they are off by default and that History must be on to turn them on.
- Turn History off, or leave it on and set auto-delete to 3 months (the shortest option; the default is 18). With it off, chats are still held for around two months, according to Google's Privacy Hub as updated in November 2026.
- Delete past activity if it contains sensitive prompts.
- For Workspace accounts, the no-training commitment is part of Google's Workspace terms; confirm with your admin that you are using a work account, not a personal one.
These changes reduce — but do not eliminate — the surface area. They cannot help if the prompt itself contains regulated data.
Step 2: Redact sensitive content before it leaves your device
The most reliable defense is to never send the sensitive bits in the first place. There are three ways to do this:
- Manual redaction. Replace names, IDs, and identifiers with placeholders before pasting. Reliable but slow and easy to forget under deadline pressure.
- Synthetic substitution. Swap real values with realistic but fake ones (e.g., "Jane Doe" → "Mira Patel"). Preserves the model's ability to reason about the structure of the document.
- Local-first redaction tools. Software that runs on your device, detects sensitive entities before submission, and masks them automatically.
Sonomos Locke 1.0 is in the third category: it runs on your device as a desktop app, with a companion browser extension that screens AI websites and a local proxy that screens command-line, script, and SDK traffic. It identifies categories like names, financial accounts, and health terms and replaces them with one-way placeholders so the model still receives surrounding context. Because detection happens before the request is sent, the detected values do not reach OpenAI, Anthropic, Google, or Sonomos; only settings and licence data leave the device. Coverage depends on the app, platform, and settings. Reversible tokenization is planned shortly after 1.0.
Step 3: Adopt a "minimum necessary" prompting habit
Borrowing from HIPAA's Minimum Necessary Standard, ask yourself before every prompt:
- Does the model need this name to answer the question? Usually no — "the customer" works.
- Does it need the full document, or only the paragraph in question?
- Does it need the real numbers, or rounded approximations?
Most useful AI tasks — drafting, summarizing, brainstorming, refactoring — work just as well on redacted or synthetic input. Reserve raw data for tools you control end-to-end.
Step 4: Separate what AI sees from what AI returns
A surprising amount of leakage happens on the way out of the model. If you ask ChatGPT to "rewrite this email," and you pasted the customer's medical history into the original, the medical history is now in the model's response window — and any browser extension, screen recorder, or screen-share session can capture it.
Two practices help:
- One-way redaction in 1.0. Mask data locally before it leaves the device. Locke 1.0 does not unmask it on the way out; reversible tokenization is planned shortly after 1.0.
- Output review. Read the response before you copy it back into a system of record. AI assistants regularly hallucinate identifiers; copying without review can introduce new PII problems.
Frequently asked questions
Is it safe to paste my own personal data into ChatGPT?
Pasting your own data is your decision to make, but understand the trade-off: free-tier prompts can be retained, reviewed, and (unless you opt out) used to refine future models. For one-off tasks involving low-risk personal information, that may be acceptable. For anything you would not want to read on the front page of a newspaper, redact first.
Does using the API instead of the web interface protect me?
Partially. Most major providers exclude API traffic from training by default and offer enterprise-grade data processing agreements. The API does not, however, prevent the contents of your prompts from being logged for abuse monitoring, typically for up to 30 days; zero-data-retention arrangements exist but are subject to the provider's approval. If those logs would expose regulated data, redact before sending.
What about "incognito" or "temporary" chat modes?
Temporary Chat in ChatGPT, Incognito chats in Claude, Temporary Chats in Gemini, and similar modes keep the conversation out of your history and (per the providers' terms) out of training. They do not prevent transit to the provider's servers, and a copy may still be held for a period — up to 30 days in ChatGPT and Claude, around two months in Gemini (what a Gemini temporary chat keeps) — for safety review. Treat them as "do-not-store" rather than "do-not-send."
Can my employer see my AI prompts?
If you use a managed browser, a corporate VPN, or a workspace account, assume yes. DLP (data loss prevention) products can inspect AI-bound traffic. This is one more reason to redact at the source: the cleaner your prompts, the smaller the audit surface. For the Google Workspace specifics, see Can My Employer See My Gemini Chats?.
Do I need a different tool for each AI?
Usually not, but coverage is not universal. Locke runs on your device as a desktop app. Its companion browser extension screens AI websites such as ChatGPT, Claude, Gemini, and Perplexity; a local proxy screens command-line, script, and SDK traffic; desktop wrapper apps are covered subject to platform and settings. Microsoft Copilot's browser chat is not screened by the extension. Coverage depends on the app, platform, and settings, so check the current list before relying on it for a given tool.
The bottom line
Protecting sensitive data in 2026 is not about avoiding AI — the productivity gains are too large to ignore. It is about giving each model the minimum necessary information and keeping the rest under your control. Tighten the provider settings you already have, build a redaction habit, and use a local-first tool for the moments when habit slips.
Your thoughts, your clients' data, and your company's secrets do not have to leave your device for AI to be useful. Treat that as the new default.
Sources and primary references
Reviewed November 2, 2026. Regulations are cited to their official text. Vendor statements are cited to the vendor's own documentation as it read on that date; vendors revise these pages, so confirm before relying on them.
- Data controls in ChatGPT — training is on by default for Free, Plus and Pro, and turning it off keeps chat history.
- How your data is used to improve model performance — Business, Enterprise, Edu and API content are not used for training by default.
- Temporary chat in ChatGPT — Temporary Chats are not used for training and a copy may be kept up to 30 days for safety.
- OpenAI API: your data — abuse-monitoring logs retained up to 30 days; zero data retention subject to approval.
- OpenAI privacy policy — content may be retained longer under a lawful subpoena.
- Anthropic: updates to consumer terms — Free, Pro and Max users choose a training setting; five-year retention if on, 30 days if off; commercial products excluded.
- Is my data used for model training? — the Model Improvement setting and the Incognito-chat exclusion.
- Incognito chats in Claude — not saved to history, not used for training, retained 30 days by default.
- What certifications has Anthropic obtained? — SOC 2 Type II, ISO 27001 and the HIPAA-ready configuration with BAA.
- Gemini Apps Privacy Hub — human review, training use, the History setting, auto-delete options and the two-month retention of temporary chats and chats with History off.
- Generative AI in Google Workspace Privacy Hub — Workspace prompts are customer data under the Cloud Data Processing Addendum and are not used for training without permission.
- Microsoft Copilot Chat FAQ — enterprise data protection for Entra sign-ins; prompts not used to train foundation models.
- Privacy FAQ for Microsoft Copilot (consumer) — now limited to the older Copilot app; there, consumer conversations are used for training unless the user opts out or falls in an excluded category.
- 45 CFR 164.502(b), minimum necessary — the HIPAA standard the post borrows.
Sonomos publishes this blog and makes Locke, which detects and masks sensitive information on your device before it reaches an AI model. How Locke works
This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.
Reference pages for this post
Related Articles
Best Privacy-Focused AI in 2026: Chatbots That Keep Your Data Private
The most private AI runs locally, via Ollama or LM Studio. Lumo, Brave Leo and Venice say they keep no chat logs; Duck.ai hides your IP. Compared, Nov 2026.
Can My Employer See My AI Chats? ChatGPT, Claude, Gemini and Copilot
Yes, if your employer controls the account, the computer or the network. Work accounts on ChatGPT, Claude, Gemini and Copilot let it read or keep chats. A side-by-side table of what each vendor documents, as of November 2026.
Can My Employer See My Copilot Chats? Work vs Personal Accounts
Copilot chats on a work account sit in your mailbox, where your employer can search, hold and export them with Purview. A personal account sits outside the tenant, but a work laptop can still capture it. What Microsoft documents, as of November 2026.