Confidentiality duties, privilege-waiver risk, court disclosure orders, and the sanctions risk from fabricated citations, in plain terms.
Lawyers ask me a version of the same question in almost every conversation about AI: does using this tool put my client's confidentiality at risk. The honest answer is that it can, depending on which tool, which data, and which settings. The duty has not changed. What changed is how easy it now is to violate it by accident.
The duty of confidentiality generally covers information relating to the representation of a client, regardless of its source, and it exists independently of the evidentiary attorney-client privilege. Pasting matter material into a third-party tool can implicate both. On the confidentiality side, the question is whether you have taken reasonable steps to protect client information from unauthorized access or disclosure, and whether the tool's handling of that data is consistent with those steps. On the privilege side, the concern is disclosure to a third party outside the privileged relationship, which can weaken a claim of privilege over that material if the matter is ever challenged.
Whether a specific AI tool's use counts as a disclosure that risks waiver is not settled the same way everywhere, and it depends on the tool's terms, how the data is used and retained, and the jurisdiction. That uncertainty is itself the reason to be conservative rather than to treat the question as resolved in either direction.
A growing number of courts have issued standing orders addressing the use of generative AI in filings, some requiring disclosure that AI was used, some requiring certification that citations were verified, and some restricting AI use for drafting altogether. These orders vary by court and change over time. The practical implication is that you cannot rely on a firm-wide AI policy alone. Check the standing orders for each court and each judge on a matter before using AI in anything that will be filed, and confirm before every filing rather than assuming the rule you checked last year still applies.
Generative AI tools can produce citations to cases, statutes, or quotations that do not exist, presented with the same confidence as real ones. This has already led to sanctions against attorneys who filed briefs containing fabricated citations without independently verifying them. The failure in those instances was not that AI was used. It was that the output was filed without a human checking every citation against the actual source. That verification step is not optional, and it is not satisfied by asking the same AI tool to check its own work.
Before any client or matter material goes into an AI tool, I ask lawyers to check four things: whether the tool's terms permit use with confidential client data, whether there is a signed agreement covering data handling and retention rather than just a checkbox in a settings menu, whether the client has given informed consent where that is required or advisable, and whether opposing counsel, the court, or a regulator would view the specific use as a disclosure. If any of those four is unclear, treat the material as off-limits for that tool until it is resolved.
Draft language, outline arguments, and summarize public or already-disclosed material with AI freely. Do not put non-public client facts, matter strategy, or privileged communications into a tool without a signed data handling agreement in place and a clear answer to the four questions above. And never file anything with an AI-generated citation you have not personally verified against the source.
If you want to work through where your firm's current AI use stands against these questions, that is a conversation I have regularly with firm leadership. See the contact page to set up a call.
This article is general information, not legal advice. Talk to your firm's general counsel or ethics counsel about your specific situation.