
The Prompt Is the New Search Term. But Not Every Prompt Is.
If you use a prompt to find or cull documents in a disclosure exercise, is that prompt a search term? I think it is, and I think that is going to matter. I also think the answer is not the same for every prompt, and that the difference is where the real work lies.
This is my opinion, and I should say plainly what it rests on. I work in this field and I talk to people who use these tools every day, but I am not a lawyer, and I have not found a ruling that settles the point. Everything below that is a claim about the rules is sourced. Everything else is me.
Why a prompt looks like a search term
A keyword search is a way of telling a system which documents to look at and which to leave alone. A prompt that asks a model to identify the relevant documents, or to sort a collection into responsive and not responsive, does the same job in plain English. The other side has always been entitled to ask how you narrowed the data down. If the method has changed, I do not see why the question should.
The rules already point this way. PD 57AD, which has governed disclosure in the Business and Property Courts since 1 October 2022, expects parties to cooperate. Through the Disclosure Review Document, it expects them to discuss and seek to agree how they will search, including keyword searches and other automated searches, and the use of technology assisted review. You can read the practice direction and the Disclosure Review Document form yourself. The wording does not mention prompts. Reading a culling prompt into it is my interpretation, not something the form says.
There is also a precedent for how this tends to go. The first English decision approving predictive coding, Pyrrho Investments v MWB Property in 2016, was largely agreed between the parties, so the court was approving a method both sides had already accepted. I would expect agreement between the parties to do most of the work here too.
Why not every prompt is
A team might use one prompt to cull a data set and another to understand a case, test a theory or plan an investigation. I have heard it argued in the industry that these are different things: the first shapes what gets disclosed, and the second is the team's own thinking. I find that argument persuasive, with one condition. You can only make it if you can show which prompt was which.
The alternative is hard to defend. In a large matter, many lawyers may each run many prompts over months. Disclosing every one of them could expose strategy, would be hard to review, and may tell the other side little about how the documents were actually found. A rule that treats every prompt as a search term does not look sensible to me, and I doubt a court would want one.
What the courts have said so far
In the searching I did, I did not find an English decision on whether GenAI prompts must be disclosed. The judiciary's AI guidance is written for judges. The Civil Justice Council's consultation, as summarised by Herbert Smith Freehills Kramer, is mainly about AI in preparing court documents. It does say that guidance on AI in disclosure may be required in due course, that parties are anecdotally cooperating on AI-assisted disclosure, and that there does not appear to be a pressing case to make disclosure lists or statements address AI use. I have not checked whether the Council has reported since. If you know of a ruling, I would like to hear about it.
So this is unsettled, and it may stay quiet for a while. That is a reason to prepare, not a reason to wait.
What I would do about it
Keep your prompts. Version them. Record which ones were used to cull, which were used to test and validate, and which were part of understanding the case, and record it at the time, not after someone asks. If you are heading to a case management conference (CMC), raise it yourself and propose how you will handle it. The cooperative approach PD 57AD asks for is the sensible way in.
If you have seen a prompt disclosed, challenged or refused, tell me. I would rather be corrected than be early and wrong.
#FromTheBunker #RolfBerryman #eDiscovery #GenAI #LegalTechnology #PD57AD