Is ChatGPT safe for confidential business information?

The short answer depends on the account

The question "is ChatGPT safe" is really three questions. Which plan is the person using? How is it configured? And what kind of information is going in? A staff member pasting a client's financial statements into a personal free account is a very different situation from the same task inside a company-managed business workspace.

According to OpenAI's own help documentation, on personal workspaces with ChatGPT Free, Plus or Pro, data sharing for model improvement is enabled by default, and users can turn off the setting called "Improve the model for everyone." OpenAI also states that by default it does not use content from ChatGPT Business, Enterprise, Edu or Healthcare workspaces, or from its API, to train its models.

So the useful answer for an owner is conditional. With a business plan, sensible settings, a clear data rule and staff who know it, many firms can use ChatGPT and similar tools with confidential material. Without those, the risk sits mostly with whichever employee happens to be using a personal account that day.

What a business plan changes, and what it doesn't

A managed business plan gives you control that personal accounts do not: the company owns the workspace, can remove departed staff, and sets organization-level policies. Similar commitments exist from other vendors. Anthropic states it does not use inputs or outputs from its commercial products, such as Claude for Work, to train its models by default. Microsoft states that prompts, responses and data accessed through Microsoft Graph in its Copilot for organizations are not used to train foundation models.

But "not used for training" is not the same as "no risk." Business plans still store conversations, can be accessed by anyone who gets into the account, and may keep data for a retention period. Feedback buttons can change how a conversation is handled; OpenAI and Anthropic both note that submitting feedback may allow the related conversation to be used. Read the vendor's current terms rather than relying on a summary, including this one.

Permissions inside your own systems matter too. Assistants connected to SharePoint, Google Drive or a shared mailbox can surface any document the user already has access to. If your file permissions are loose, an AI assistant makes that visible very quickly, so tidy access before you connect it.

Risks that have nothing to do with training

Most real incidents come from ordinary security and process gaps. A widely reported 2023 example: Samsung restricted generative AI tools on company devices after engineers pasted sensitive source code and meeting notes into ChatGPT. Patterns to plan for:

• Personal accounts used for client work, which leave the firm with no control or record when someone leaves.
• Weak logins. Turn on single sign-on or multi-factor authentication for every AI account.
• Shared links to conversations that are forwarded beyond the intended audience.
• Browser extensions and free "AI" apps with unclear data terms.
• Uploading whole folders when only a paragraph was needed.

Professional and regulatory obligations

Your own obligations may be stricter than the vendor's defaults. Some examples, in general terms:

Opinion 512 goes further than many owners expect: where informed client consent is needed, it says boilerplate language in an engagement letter is not enough. It also expects lawyers to review a tool's terms of use and privacy policy, or have someone qualified do it, before putting client information in.

• Law firms: ABA Formal Opinion 512 says lawyers using generative AI must consider their duties of competence, confidentiality and communication, which can include client consent in some situations.
• Health care: protected health information should only go into a service that will sign a HIPAA business associate agreement. Without one, keep patient data out.
• Tax and accounting: the FTC Safeguards Rule requires a written information security program, and IRS Publication 4557 explains how preparers should protect taxpayer data. AI tools belong in that plan.
• Client contracts: some engagement letters, NDAs and master service agreements restrict where data can be processed. Check before you paste.

A practical data rule for your team

Write the rule in terms people can apply in the moment. One version that works for many firms:

Never in any AI tool: passwords, bank or card numbers, Social Security numbers, and health information unless a business associate agreement is in place. Company business accounts only: client names, project details, contracts, financials, employee information. Anywhere: public information and generic writing tasks.

Add two habits. Share the minimum needed, for example the paragraph you want rewritten rather than the whole file. And remove names and account numbers when they are not needed for the task.

Give examples from your own work, such as which documents from a typical client file fall into each tier.

Questions to ask before approving any AI tool

The same checks apply to ChatGPT, Claude, Copilot, Gemini or an AI feature inside your practice software. Write the answers down and keep them with your approved tool list.

• Is this a business plan with a company-owned workspace, or an individual's account?
• Does the vendor state that business inputs and outputs are not used for training by default, and where is that stated?
• How long are conversations and files retained, and can an administrator change that or delete them?
• Who at the vendor can access content, and under what circumstances?
• Does it support single sign-on or multi-factor authentication, and can we remove a user immediately?
• Will the vendor sign the agreements our obligations require, such as a HIPAA business associate agreement or a data processing agreement?
• What happens to connected data, such as SharePoint, Google Drive or email, and does the tool respect existing permissions?
• Can we export or delete our data if we leave?

Setting it up properly

If your team will use ChatGPT, or any similar assistant, for client work, a short setup checklist covers most of the risk: buy a business plan for regular users, connect it to your company login with multi-factor authentication, review the retention and sharing settings, disable or limit features you do not need, write the data rule into your AI policy, and stop client work on personal accounts.

Revisit the settings every six months, since plans and terms change. If you want help with the setup or the policy wording, InTheory covers this in a typical AI policy engagement, and your counsel should review anything that touches privileged or regulated data.