This is not legal advice and not an opinion about your obligations. The operative rules are your state’s, and many state bars have issued their own guidance on AI. It varies.
The short answer
Often you can, with care. The duty is to make reasonable efforts to prevent disclosure, and the ABA’s guidance asks for the client’s informed consent before confidential information goes into a third-party AI tool where that would disclose it. So the questions are what the tool does with what you type, whether that is a disclosure, and whether the client has agreed. A tool that runs on your own machine and sends nothing avoids the disclosure question on that point.
What the rules ask
- Rule 1.6(c): reasonable efforts. Not a guarantee that nothing ever leaves the office. Comment [18] lists the factors: how sensitive the information is, how likely disclosure is without more safeguards, what those safeguards cost and how hard they are, and whether they get in the way of representing the client.
- ABA Formal Opinion 512 (July 2024). The ABA’s first ethics opinion on generative AI. It asks lawyers to keep confidential everything relating to a representation, whatever its source, and to get the client’s informed consent before using a third-party generative AI program where that would disclose confidential information.
- Rule 1.1, Comment [8]: competence in technology. Understanding enough about a tool to judge the risk is part of the job.
- Rules 5.1 and 5.3: supervision. The output needs the review any junior work would get.
Questions to ask about any AI tool
- Where does the model run, and does what I type leave my computer?
- Is it kept? For how long, including backups, and who can read it?
- Is it used to train models, by default, and can that be switched off?
- Can a court or a third party reach it? Logs that exist can be ordered preserved, and in 2025 a court did exactly that to an AI company in copyright litigation.
- What contract, certification or data-processing terms come with it, if your clients’ guidelines require them?
- Does it check what the model says, or leave that entirely to you?
The options
- Keep confidential material out and use AI for the parts of the work that don’t need it.
- Remove the identifying details first. That helps, but a set of facts can identify a client on its own.
- An enterprise tool with contractual terms: no training, limited retention, and the certifications some clients require. Your words still go to the vendor.
- A tool that runs on your own machine, where the words are not transmitted at all, so there is nothing for a vendor to retain or produce.
Where Zero Witness fits
In browser mode the model runs on your own computer and nothing you type or attach is sent, so no third-party program receives it. On our hardware your words are encrypted in your browser, read in memory on hardware we own, not written to disk and not trained on, and that one is a disclosure to us. The page always shows which you are in, and a matter can be set so its work never leaves your device. Checking a citation sends its volume, reporter and page, and a search sends its words, and both are named on screen before they go.
What it lacks matters as much. There is no audit trail and none can be added, which rules it out for anyone under a records-retention duty. There is no SOC 2 report, BAA or data-processing agreement. And the model can be wrong: citations and quotations are checked against real law, its reasoning is not.
The technical brief for lawyers goes through each rule in more detail, including recording and transcription guidance and Illinois and California.
More: the technical brief for lawyers. Related guides: How to catch a fabricated citation before you file · How to build a table of authorities without uploading your brief · How to tell whether an AI chat is really private · How to get a deposition summary from AI that you can check. All guides.