Is It Safe to Put Client Confidential Information Into ChatGPT?

You should avoid entering client confidential information into ChatGPT on its default settings. On the Free and Plus plans, your inputs may be used for training, and even with training turned off, sending the material to an outside server can itself breach a non-disclosure agreement. At the same time, refusing to use AI at all isn't realistic, especially for freelancers whose work consists mostly of reading and writing documents supplied by clients. This article sets out a framework for deciding what can go in, split into three parts: the contract, the settings, and the alternatives.

"Used for training" and "leaving your organization" are separate risks

Putting confidential information into ChatGPT carries two risks, not one. Mixing them up leads to bad decisions, so it helps to separate them first.

The first is the training risk. On the Free and Plus plans, with default settings, conversation content can be used to improve the models. You can turn this off under Data Controls in settings, and Temporary Chat is available for conversations you don't want saved to history. In other words, this risk can be reduced by the user. Note that the API and business plans don't use inputs for training by default, so the same ChatGPT can be handled differently depending on how you access it.

The second is the external transmission risk. Turning off training doesn't change the fact that what you type is sent to and processed on the provider's servers. Most NDAs prohibit disclosing confidential information to third parties, and whether entering it into a cloud service counts as "disclosure" depends on how the contract is written and interpreted. Some contracts can reasonably be read as allowing it; others are written to forbid uploading material to any outside server.

You'll often see the claim "it's safe because training is off." That answers only the first risk. For freelancers, the second one is usually the heavier of the two. If a contract breach becomes an issue, "I had training turned off" won't carry much weight against a non-disclosure clause.

What to check in your contract

The starting point is not ChatGPT's terms of service but the contract you signed with your client. These are the sections worth rereading at the start of each project.

First, the definition of confidential information. If it's limited to "information marked as confidential," the scope is relatively narrow. If it covers "all information obtained in the course of the work," it can extend to email bodies and meeting notes. In that case, even summarizing the material instead of pasting it may amount to the same thing.

Next, the clause prohibiting third-party disclosure, along with any limits on copying or removing material. Newer contracts sometimes contain explicit clauses about external cloud services or generative AI. If there is one, follow it; if there isn't, you're in a gray area. The most reliable way to handle a gray area is to ask the client, describing the specific use, and keep a record of their approval. It's an awkward question to ask, but far less awkward than the client discovering later that you used AI without telling them. When the request is narrowly scoped, clients rarely refuse, and many appreciate the care.

If the material includes personal data, such as names of contacts at partner companies or customer lists, data protection laws (such as the GDPR in Europe) come into play separately from the contract. How this applies depends on the nature of the project, so when in doubt, defer to the client's legal or compliance team.

Match the method to how sensitive the material is

Not every task needs the same restrictions. It's more practical to choose a method based on how sensitive the material is and what kind of work you're doing:

Masking is easy but has limits. Replacing a company name with "Company A" doesn't help if the business description, deal structure or distinctive figures still identify them. Masking only protects you when the hidden information can truly be separated from its context.

That's where local LLMs come in. The model runs on your own computer, so the material never leaves the device. They don't match the top cloud models, but for clerical tasks like summarizing, standardizing terminology or tidying up drafts, they're increasingly practical. For more on building a workflow that doesn't rely on sending data to the cloud, see the explanation of privacy and local-first design from osFoundry, a BYOK desktop agent workspace that can also run local models directly.

Once you draw the line, you'll spend almost no time agonizing over it project by project. A simple two-tier rule works well: NDA-covered material itself is handled locally or by hand, and drafting or research that doesn't touch confidential information goes to the cloud. Being able to explain your criteria to a client is itself a source of trust for a freelancer.

FAQ

If I turn off training, can I enter confidential information?

It depends on the contract. Turning off training doesn't change the fact that the data is sent to and processed on an outside server. If the NDA prohibits disclosure to third parties or taking material outside the organization, changing a setting won't solve that.

How should I ask the client?

Avoid a vague "Is it okay to use AI?" Instead, state specifically which materials, which service, which settings and for what purpose. Get the approval in a form that leaves a record, such as email, so it's easy to explain later.

Can a local LLM run on an ordinary computer?

Yes. Even a machine with around 16 GB of memory can run smaller models well enough for tasks like summarizing and reformatting text. For advanced reasoning or highly specialized translation, though, there's still a gap compared with top cloud models.