The doctrinal landscape
Three years into widespread legal AI use, the federal circuits have begun to diverge on a question that the drafters of Hickman v. Taylor never contemplated: when an attorney prompts an AI tool to summarize, analyze, or draft, what protections (if any) extend to the resulting output?
Five circuits have now ruled on aspects of this question. Three doctrines have emerged. None of them are stable.
The three positions
The Second and Ninth Circuits have taken what we call the tool view: AI outputs are work product to the same extent that the underlying mental impressions of the lawyer are. The fact that the words were typed by a model is irrelevant.
The Fifth and Eleventh Circuits have adopted what amounts to the transformation view: substantial AI involvement may strip work product protection unless the lawyer can demonstrate independent judgment at the prompt and review stages.
The D.C. Circuit, in a fractured opinion last March, gestured toward a third position: that protection turns on the specific nature of the prompt, with formulaic prompts receiving less protection than ones that reflect litigation strategy.
Where this is going
The Supreme Court denied cert on the issue twice in 2025. It is, in our view, almost certain to grant in the 2026-2027 term, given the now-clear circuit split and the practical importance.
In the meantime, prudent practice is to assume the most restrictive standard: document independent legal judgment at the prompt-design and output-review stages, and treat AI prompts and logs as potentially discoverable until told otherwise.