Legal Cyber Academy

Building effective discovery requests for AI artifacts using interrogatories and Rule 26(f)

Free. Watch without an account. Sign in to save progress and take the exam.

Up next

Privilege, work product protection, and recent AI discovery case law
TranscriptThe player shows YouTube’s automatic captions, which are machine captions we do not review. The full transcript is below.

00:00:00But the timestamp of the document itself being generated by the system isn't hallucinated. Right.

00:00:06So the source of truth is different than the document itself, because it's not actually associated with the document. It's attached to the documents generation.

00:00:13But it's not it's not embedded in it. And so you have to embed that in the document as well.

00:00:20But anyways, these are all things that will and then, oh, yeah, rerunning an agent has improved because the agent learns every time.

00:00:29So it's, you know, like a kid, you know, you point it eventually the kid realizes to brush their teeth after they go to the dentist enough of the time.

00:00:40Discovery, interrogatories, all that my way in here, because I feel like this is like, I'm going to give you 10 interrogatories and I get 10 interrogatories with 20 parts each.

00:00:52Yes. Under the federal subparts count as parts. Keep that in mind.

00:00:56I know. And you have to tell them this. And I'm like, nobody reads the rules.

00:01:00Yeah. I mean, I think it's a it's a great point. What's here is like to up front serve interrogatories that are going to give you the baseline information.

00:01:08That way, you know how to make narrower, more discreet, more targeted document requests and deposition topics for a corporate deposition witness.

00:01:18Because again, courts are still trying to figure this out. But at least in some recent cases, you know, judges have been saying, you need to explain to me why this particular piece of evidence that you're requesting actually ties to your claims.

00:01:32And so asking for all documents related to X, I think, is not going to cut it in many cases. And so serving interrogatories up front that help you get a lay of the land will help you narrow those requests so that, you know, if you ever get to a motion to compel a judge can easily say, yes, you've articulated why what you're asking for is actually connected to these claims or defenses or vice versa.

00:01:54You know, you just ask for everything under the sun. And so, you know, the other side has provided a very good explanation for why, you know, 75 percent of what you're asking for is irrelevant.

00:02:03So this is a good way to sort of hone things down and filter out what's just noise at an earlier stage of the case.

00:02:13I think the thing where you should really make an effort is make an effort to understand the product, the artifact name, like prompts, output, system prompts that relate because nothing's more irritating than you getting it wrong.

00:02:26And they're like, we don't have that. But I can tell you as a special master that's hearing it. I know that they have it, but I know you asked it the wrong way.

00:02:34It's not my job to help you get what you want. It's your job to help yourself. You help yourself by learning how the system actually refers to things.

00:02:42And if you don't call it the right thing, then you don't you won't get it. It's just that simple.

00:02:47And I mean, it sounds trivial, but a lot of people don't take the actual time to understand because even, you know, and the time periods and all the things and metadata and all these things.

00:02:59And third party possession is something you want to get up front in your interrogatory so you can adjust that quicker rather sooner rather than later because data is ephemeral.

00:03:09So let's be clear. If you're suing in 2028, you're saying your litigation hold notice in 2025, incident happened in 2024, right?

00:03:20And it's a third party that you'll be lucky that company is still around, right? So you really need to figure that out really quickly in the process.

00:03:3126F, this is common sense. I mean, I get this is just annoying to me. Just be prepared.

00:03:38Like just do your work. Like it's not a one-up. The whole point of 26F is for you to streamline how discovery is supposed to work.

00:03:48Not win a case. Don't litigate the 26F. I don't know, Mari, if there's anything you want to...

00:03:57Yeah, I agree with you wholeheartedly.

00:04:01Yeah. Oh, yeah. Here's the thing. If you're like, there's only a handful of cases and I'm involved with several of them where they're about model versions and the impact and the marketplace.

00:04:11That is like a very like two, four cases in the country that actually have to do with a particular model and a particular tort.

00:04:20Most cases, the model version frees, you better know exactly what you're asking and be very specific about it. Because if you're not specific about your requests, not all models are the same.

00:04:34Clod, Anthropic, they have, I don't even know how many, and then you have MPC server. Like there's a linear moving part.

00:04:42So you really need to understand what you're asking for because they're going to come back and rightfully so argue that it's unduly burdensome and overly broad to say freeze the entire model. Like it's just not practical or pragmatic.

00:04:55Yeah. And maybe one other point about sort of fronting issues at the beginning.

00:05:02I used to hate getting those like pre discovery requests, preservation letters, like before a party would actually serve requests, but they would send you a letter saying preserve these five million things.

00:05:13And I would usually either not respond or say, yeah, I know what my obligations are. Leave me alone.

00:05:19But if there is something very particular, like very, very particular, like like this example that Daniel just gave of, you know, model versions that need to be frozen.

00:05:29Find a way to front just those discrete issues right away.

00:05:32If you actually are concerned that there's going to be material evidence that's lost because, you know, someone getting one of those letters is not going to know what to focus on. But if you do anything with it.

00:05:41Yeah, what's that?

00:05:42They don't do anything with it. I mean, they get like big corporations get them all the time. Like there's a whole department's dedicated to the process.

00:05:52Eventually, they'll change their processes and protocols, large corporations to address the issue. But it's like turning a ship. It takes time.

00:06:02Totally. But like the actual player in a litigation, if they reach out to outside counsel and say, like, hey, this whole litigation is about this one issue.

00:06:11And, you know, please make sure you're preserving that. I think that actually will go a long way in at least having someone take steps to preserve it.

00:06:19Sorry, Karen, I cut you off.

00:06:21No, I was going to say just from the other perspective, when you're preparing counterclaims or anything like that, how do you what are some tips for people who to narrow that down if you don't know.

00:06:34Right.

00:06:35So I actually just dealt with this as an expert. This is a fascinating problem because the A.I. is a black box. So this is where it gets real tricky and depending on the type of A.I.

00:06:47So what you would want to do is say, so this is the whole point of the 26 F being a productive process. So what I tell them, and there's no easy way to do this because of the adversarial nature of the entire process.

00:07:00But what I tell them is, look, go to the other side and say, look, we don't understand how this works. This is what we're looking for.

00:07:05Can you give us information so we can understand. So we're not going to ask for everything and help focus the conversation.

00:07:12So you shift it back to them. So then if they say, oh, definitely not or whatever else, then you can at least point to the court that you made a good faith effort to say, hey, look, we actually made it a constructive attempt, but they just refused to provide us anything.

00:07:26So all we had to work with. Right. Sometimes you can query the system, read some papers or whatever else, but it's truly not as nearly as effective as them getting the engineer talking to the lawyer for 30 minutes and then I'm telling you the answer.

00:07:42And but then it transfers the risk back to them. Right. Because then it's a real bad look. If you end up destroying whatever was sought, even though they came to you and ask you for the specifics. Right.

00:07:53So by shifting that sort of cover, it's about controlling a narrative. Right. And so that's why having a more cooperative ends up being more effective tool for the litigation downstream.

00:08:06That's the general way I tell people to approach it because you're 100 percent right. Right. Most people have no idea. So the only way to shift that is to ask for the information or say, look, otherwise we're going to have to ask for everything.

Automatically generated; it may contain errors.

Study materials

Notes, flashcards and practice questions for this lesson, reviewed before publication.

Study notes

Embedding timestamps so AI-generated document metadata is not lost

A generated document's own creation timestamp is not hallucinated, but it is attached to the generation event rather than embedded in the document itself. Practitioners are told this timestamp must be deliberately embedded in the document, or the source of truth for when it was created becomes separated from the artifact. The lesson also notes that rerunning an agent can produce improved results over time because the agent learns from repetition, compared to a child who eventually learns to brush their teeth after enough visits to the dentist.

Why federal subpart counting changes how interrogatories should be drafted

Under the federal rules, subparts count as separate interrogatory parts, so a request structured as ten interrogatories with twenty parts each will be treated as far more than ten interrogatories. A speaker stresses that this limit must be explained to clients and colleagues because many people do not read the applicable rules closely enough to account for it.

Using interrogatories up front to narrow later discovery requests

Serving interrogatories early is described as a way to get baseline information before sending document requests or scheduling a corporate deposition witness. This baseline helps a party draft narrower, more targeted document requests and deposition topics. Courts are still working out standards in this area, but recent cases show judges asking parties to explain why a specific piece of evidence actually ties to their claims. A blanket request for "all documents related to X" is described as unlikely to succeed. Interrogatories served up front help a party later show, on a motion to compel, that what is being requested connects to the claims or defenses, rather than seeking everything under the sun and prompting the other side to explain why most of it is irrelevant. This approach filters out noise earlier in the case.

Learning the correct artifact names before requesting AI system data

A speaker who has served as a special master explains that the most important preparation step is learning how a system actually refers to its own artifacts, such as prompts, outputs, and system prompts. If a request uses the wrong terminology, the responding party may truthfully say "we don't have that," even though the special master hearing the dispute may know the data exists but was simply asked for incorrectly. The speaker states plainly that it is not a special master's job to help a party get what it wants; it is the party's job to help itself by learning the right terms, time periods, and metadata before serving requests.

Addressing third-party possession and data ephemerality early in a case

Third-party possession of relevant data should be identified early in interrogatories because data is ephemeral. The lesson gives a timeline example: litigation filed in 2028, a litigation hold notice sent in 2025, and an incident that occurred in 2024, with a third party involved. In that scenario, a party may be fortunate if the third-party company still exists by the time discovery occurs, so identifying third-party possession quickly in the process is important.

Treating Rule 26(f) conferences as preparation rather than a competitive event

The Rule 26(f) conference is described as existing to streamline how discovery is supposed to work, not to win a case. A speaker calls it common sense that parties should simply be prepared and do their homework rather than treat the conference as an opportunity to gain a one-up advantage. Mari Grace agrees with this point.

Why freezing an entire AI model version is usually unduly burdensome

Only a small number of cases nationally actually concern a specific AI model version and its impact in the marketplace tied to a particular tort claim. In most cases, a party must be very specific about what it is asking for regarding model versions, because models differ significantly, for example Claude and Anthropic offerings, in addition to components like an MCP server. Asking to "freeze the entire model" without specificity invites a legitimate objection that the request is unduly burdensome and overly broad, since it is not practical or pragmatic.

Fronting discrete preservation issues instead of broad preservation letters

Broad pre-discovery preservation letters asking a party to preserve a large, generic list of items are often ignored or met with a reply that the recipient already knows its preservation obligations. Large corporations receive many such letters and have entire departments dedicated to handling them, and changing their processes takes time, like turning a ship. However, if there is one very particular issue, such as needing to preserve specific model versions, fronting that discrete issue directly with outside counsel and explaining that the litigation turns on it is more likely to prompt someone to actually take preservation steps.

Narrowing discovery when the AI system is a black box

Karen Silverman raises the problem of narrowing discovery requests, including for counterclaims, when a party does not yet understand the AI system involved. One speaker responds that this is a genuinely difficult problem because AI can be a black box, and the difficulty depends on the type of AI at issue. The recommended approach, framed as part of the productive purpose of the Rule 26(f) process, is to tell the other side directly that you do not understand how the system works and ask them to provide information so you can focus your request rather than demanding everything. If the other side refuses, that refusal can be pointed to later as evidence of a good-faith effort that the court can weigh. Querying the system or reading papers is described as far less effective than having an engineer speak directly with a lawyer for about thirty minutes. This approach also shifts risk back to the other side: if they are asked for specifics and later destroy the sought material anyway, that becomes a serious problem for them and strengthens the requesting party's position, making cooperation a more effective tool for downstream litigation.

Flashcards

14 cards. Test your recall, then flip to check.

Select the card, or press Space, to flip it. Left and right arrow keys move between cards.

1 of 14

Practice quiz

Practice — not graded, not the final exam

Question 1 of 5Per the lesson, why does a document's generation timestamp create a metadata problem?
Choose an answer to check it.

Slides & materials