You can get real help from ChatGPT, Claude or Gemini on a contract, a ledger or a set of session notes. The question is what leaves your computer when you do. This is a routine for keeping the client's identity out of it.

Know what your AI tool does with your data

Start with the terms of the account you actually use. They differ by plan, and the defaults are not the same for personal and business accounts.

For ChatGPT, OpenAI's help center says that when "Improve the model for everyone" is off, new conversations are not used to train its models and can still appear in your chat history. It also says that content from ChatGPT Business, Enterprise and Edu workspaces is not used for training by default. One catch: if you give thumbs-up or thumbs-down feedback on a response, the whole conversation may be used for training.

For Claude, Anthropic describes consumer chats (Free, Pro, Max) as used to improve Claude only if you choose to allow it in your privacy settings. Its commercial products, such as the API and Claude for Work, are covered by separate terms.

For Gemini, Google's Gemini Apps Privacy Hub says that while Keep Activity is on, some chats are read by human reviewers and used to improve Google's models, and it asks you not to enter confidential information you would not want a reviewer to see. With Keep Activity off, future chats are not used to train its models unless you send feedback, but they are still kept with your account for up to 72 hours. Gemini used with a Google Workspace work or school account is covered by the Workspace terms, under which prompts are not used to train models without the organization's permission.

Whatever the setting says, a conversation is still data held by a third party. If you are a lawyer, that matters for confidentiality. ABA Formal Opinion 512 on generative AI tools says that a client's informed consent is required before you input information relating to the representation into a self-learning tool. The same opinion notes that consent is not needed when you use a tool without putting in information relating to the representation. Accountants, therapists and consultants have their own rules and contracts, but the question is the same: what exactly are you putting in?

Send less than the whole document

Before you worry about names, ask whether the AI needs the whole file. Many tasks need one clause, one table or one page. Paste only that part.

A short test: if you would not read this passage aloud to a stranger on a train, do not paste it as it is. Either cut it or change it.

Replace identifying details first

The most useful habit is to swap identifying details for placeholders before anything is sent. "Maria Lopez signed a lease with Northwind Ltd" becomes "Person_1 signed a lease with Company_1". The AI can still summarize the lease, find the inconsistent clause or draft the reply. It does not need to know who Maria is.

Three rules make this work:

  • Be consistent. The same person gets the same placeholder everywhere. If Maria is Person_1 in one paragraph and Person_4 in another, the AI will treat them as two people.
  • Cover every form of the name. "Maria Lopez", "Ms. Lopez", "M. Lopez" and "Lopez" are one person. Names also appear in email addresses, signatures, headers and file names.
  • Look beyond names. Account numbers, ID numbers, addresses, phone numbers and dates of birth identify a person as surely as a name does. The checklist of what to remove covers the full list.

Doing this by hand with find-and-replace works for a one-page letter. It becomes unreliable with a folder of thirty files, because one missed spelling is enough to expose the client.

Keep the key on your side

Replacing names is only half the job. You also need a record of which placeholder stands for whom, so you can put the real names back into the AI's answer. This record is the key. Keep it on your own machine and never paste it into the chat.

Treat the key as confidential in its own right. Regulators make the same point about pseudonymised data in general: the UK Information Commissioner's Office notes that pseudonymised data remains personal data in the hands of anyone who holds the additional information, which is why that information has to be kept separately and securely.

Common mistakes

  • Cleaning the text but not the file name. If you upload a file, its name travels with it. Rename it first, or paste the text instead.
  • Pasting the key along with the document. A list of "Person_1 = Maria Lopez" in the same chat undoes everything you just did.
  • Trusting a single pass. A name can be spelled three ways in one document, and a signature block or a footer is easy to miss. Always read the safe version before you send it.
  • Treating a privacy setting as a guarantee. A setting that stops training on your conversations is not the same as keeping the content out of a third party's hands. The safest content is content that never held the client's identity.
  • Forgetting the people around the client. Opposing parties, employees and relatives are people too. Their names belong on the list.

A routine you can repeat

  1. Check the account. Confirm which plan you use and what its data settings say. Turn off training on your data where the option exists.
  2. Pick the passage. Select only what the task needs.
  3. Replace the details. Use consistent placeholders for every person, company, address, number and date.
  4. Read it once more. Skim the safe version for anything that still points to the client: a rare job title, a small town plus a date, a nickname.
  5. Ask your question. Say what the placeholders are, for example "Person_1 is the tenant and Company_1 is the landlord".
  6. Restore the names locally. Put the real names back into the answer on your own computer, and check the answer before it goes into a client document.

The AI can make mistakes, and replacing names does not change that. You remain responsible for the result, including the parts it got wrong.

ClientSafe, a Windows app from GagarinSoft, does the replacing, the consistency across files and the restoring on your PC. You review what it found before you copy anything. Read more about ClientSafe.

General information, not legal advice. Check the rules that apply to you.

Sources