The privilege illusion
When a vendor demos an AI tool to in-house counsel, the privilege conversation lasts about ninety seconds. There is a slide. The slide says we do not train on your data. The deal moves forward.
This is the wrong conversation. Privilege is not a checkbox about training data. It is a doctrine, with a hundred years of case law, that turns on the substance and circumstances of communication. And current legal AI tools (even the good ones) are designed in ways that quietly degrade it.
Three failure modes
First: indiscriminate context windows. When a tool ingests a hundred documents to answer a question, it is creating a derivative work that may itself be discoverable. The AI summary is not necessarily privileged just because the underlying documents were.
Second: the consultant problem. If the AI vendor is functionally a consultant, a number of jurisdictions have already held that the privilege analysis is materially different than for a software vendor. Most legal AI contracts do not address this. They should.
Third: prompt logs. Most enterprise legal AI tools log every prompt. Those logs can be (and have been) subpoenaed. The fact that the underlying conversation was privileged does not automatically extend to the prompt history.
What good looks like
A privilege-aware legal AI tool is one that lets a GC define, document by document, what enters the context window. That treats prompt history as privileged work product by default. That structures its data residency so that no third party can ever be compelled to disclose without notice.
We are not there yet. But we should be, by the end of 2027. The firms and vendors that arrive first will define the doctrine for the next twenty years.