By Shoumik Goswami · Last updated October 2026 · Policies last checked 7 October 2026

How to use AI with confidential data, without handing it over

You can use AI on confidential work safely, but not by pasting the data into a consumer chat. Use a business AI plan your company has approved, or keep the sensitive values out of the AI entirely. Swapping names, amounts and dates for placeholders lets the AI do the writing or design while the real data stays on your side.

Is it safe to paste company data into ChatGPT or Claude?

On a personal plan, usually not. By default, ChatGPT Free, Plus and Pro may use your conversations to train OpenAI's models. You can switch this off under Settings → Data controls → "Improve the model for everyone" (OpenAI).

Since 2025, Claude Free, Pro and Max users choose whether their chats train Anthropic's models. If they allow it, data is kept for up to five years; otherwise for 30 days (Anthropic).

Business plans are different. OpenAI does not train on ChatGPT Business or Enterprise content by default (OpenAI). Claude's Team, Enterprise and API plans fall under Anthropic's commercial terms, which the consumer change does not cover (Anthropic).

Plan	Trains on your chats by default?	What it means for confidential data
ChatGPT Free, Plus, Pro	Yes, unless you opt out	Avoid for client or company data
Claude Free, Pro, Max	Your choice	Check your setting; still a personal account
ChatGPT Business, Enterprise	No	Use if your company approved it
Claude Team, Enterprise, API	No (commercial terms)	Use if your company approved it

Training isn't the only risk

Even with training switched off, pasting data into a chat still means:

It leaves your control. The text sits on a provider's servers under their retention rules.
It may break a policy or contract. Many companies forbid sensitive data in unapproved AI tools. Client contracts and NDAs often go further.
It follows you. Personal accounts travel with you when you change jobs.

This isn't hypothetical. In 2023, Samsung restricted staff use of generative AI after engineers pasted internal source code into ChatGPT (Cyber Daily). That March, Cyberhaven reported that 4.2% of the 1.6 million workers at its client companies had tried to paste company data into ChatGPT, and estimated that a company of 100,000 people could be sharing confidential data with OpenAI hundreds of times a week. One example it gave: an executive pasted the firm's 2023 strategy document into ChatGPT and asked it to create a PowerPoint deck (Cyberhaven, March 2023).

Three safe ways to use AI on confidential work

Approach	How it works	Best for	Watch out for
1. Approved business AI	Use your company's ChatGPT Enterprise, Claude Team or similar	Everyday work inside policy	Not every company has one; client contracts may still forbid it
2. Mask it yourself	Replace names and numbers by hand or with a redaction tool, then paste	Short text: emails, summaries	Slow; you rebuild the output with real values yourself
3. Placeholders with automatic fill-in	A tool swaps sensitive values for tokens, the AI works on those, then your device puts the real values back	Long documents and finished outputs like decks and reports	Context can still hint at meaning (see below)

How placeholders work

The AI gets the meaning and structure of your material, but none of the sensitive details.

Your original: "Acme Corp renewed at $2.4M, up 18% on last year."
What the AI sees: "[CLIENT_1] renewed at [MONEY_1], up [PCT_1] on last year."
What the AI writes: a slide titled "[CLIENT_1] renewal grew [PCT_1]" with a chart.
What you get: the same slide with "Acme Corp", "$2.4M" and "18%" filled in on your machine.

What placeholders don't protect

Placeholders hide values, not context. If a document says "[CLIENT_1], the Cupertino phone maker", the AI can guess who it is. Reversible placeholders also count as pseudonymized, not anonymous, so privacy laws like GDPR still apply to the data (PC-Tech).

For highly sensitive material, give the AI only the structure: section lengths, table shapes and trends, with no wording at all.

Where OpenGamma fits

OpenGamma is a private AI presentation generator. It lets Claude, ChatGPT or any MCP-capable AI build on-brand PowerPoint decks from confidential documents, using placeholders instead of the real data, which the user's browser fills back in. OpenGamma uses approach 3 for decks and reports. Your browser encrypts your document and swaps sensitive details for placeholders. You approve exactly what Claude or ChatGPT sees. Your browser then fills the real values into an editable, on-brand PowerPoint. Blind mode gives the AI only structure, for the most sensitive files.

See how the privacy model works

Make a private deck free 30 days of Core free, no card needed

FAQ

Does ChatGPT train on my data?

On Free, Plus and Pro plans, yes by default, unless you turn off "Improve the model for everyone" in Data controls. ChatGPT Business and Enterprise do not train on your content by default.

Does Claude train on my chats?

On Free, Pro and Max plans, only if you allow it in your settings. Claude Team, Enterprise and API usage are covered by commercial terms and are not part of that change.

Is turning off training enough to share client data?

Not always. The data still leaves your control, and your company policy or client contract may forbid it regardless of training settings.

What is the safest way to use AI on confidential documents?

Keep the sensitive values away from the AI. Use placeholders so the AI works on structure and meaning, and restore the real values on your own device.

Can AI work out what the placeholders hide?

Sometimes, from context. Remove identifying descriptions, or share only structure for the most sensitive material.

Does masking make data anonymous under GDPR?

No. Reversible placeholders are pseudonymization, so the data is still personal data under GDPR.

Sources

Policies change often. If something here is out of date, tell us and we'll correct it. This is general information, not legal advice. Home · Privacy model