Legal Cyber Academy

Privilege, work product protection, and recent AI discovery case law

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

Up next

Sign in to take the examPass it for a verifiable Certificate of Completion

Then

TranscriptThe player shows YouTube’s automatic captions, which are machine captions we do not review. The full transcript is below.

00:00:00Point us to something or we can't help you. Privilege. This is something that the law has made pretty clear. Well, I'll let Mari you can weigh in here on the privilege slide here because I mean.

00:00:15Yeah, just because somebody is asking a question that should be asked to a lawyer or has to do with a legal issue in a tool does not make it privileged.

00:00:25And so, you know, being really careful and conscientious about when folks are using a tool to prompt or ask questions or research something related to a litigation has to be very well monitored.

00:00:42I think there's a little bit more leeway when it comes to work product, but you want to make sure that people are being clear when they prompt about what they're doing, that they are doing something at the direction of counsel in connection with a litigation or that what they're doing or saying reflects legal advice from a lawyer.

00:00:59And in all instances, you know, people should be using enterprise tools when they undertake these kinds of things or else there's a significant risk of way.

00:01:09So I agree with you 100%. Not only do I tell them to use enterprise tools, but what I tell them to do for experts particularly is create a version of the tool that's called a work product tool.

00:01:22So they're never ever going to have a question about it, right? Because then you have your normal tools that you use and then you have your tools that you're using specifically at the instruction of counsel.

00:01:33One of the problems with experts is that they co-mingle their use of tools and they have no way to prove that this one was used for just this, this one was used for just that.

00:01:42So you want to sort of have two totally separate things because otherwise all of your work product can be unfortunately discovered and co-mingled and they're not clean about it.

00:01:53And I mean, it can be a real...

00:01:56Yeah, that's totally right. And I love that tip. That's really good.

00:02:00And even if an expert is using a single tool, it's also important to think about sort of what steps they are taking.

00:02:11No, you should create an instance of that tool called work product because then you can turn on the logging for that tool and then you will have a log that they can provide to counsel that will show them it was done at the direction of counsel.

00:02:25Correct. Yeah, the only caveat though for experts is I would say that if it has to do with their methodology...

00:02:32Or then your troubles.

00:02:33Or if the tool provides something that an expert actually reviews or relies on for a report, that can be discoverable.

00:02:44100%. But they shouldn't do that anyways.

00:02:47Yes, fair.

00:02:49That's the whole point of having the expert, then you don't need an expert, you just use the AI.

00:02:55And there are also limitations on these tools that don't support more than one email at a time. I mean, like, this is...

00:03:02Right? I mean, realistically, right? These tools are not really built for...

00:03:08We spin up instances for every single case. We don't use it for any opinion ever.

00:03:17Then we just use deterministic AI. I don't even look to AI to do anything.

00:03:22I feed it in the brief, the opinion, and then I say, go find this.

00:03:26And it goes and it finds all of the forensic data and then we verify and validate all of the findings and data.

00:03:32So there's never any reliance on anything the AI does.

00:03:35Because let's be honest, we've just seen so many situations where the harness...

00:03:41It just doesn't work perfectly and there's no room to error.

00:03:46You can't error, right? Because then your entire case falls apart.

00:03:51Right. I mean, so a lot of recent decisions, I can't comment on either one of these.

00:03:56Just so happens that I touched them in some fashion.

00:04:00So I'll turn it over to Mari or Karen, if you want to touch on either one of these.

00:04:06Mari, go for it.

00:04:08Yeah. I mean, I think the big picture takeaway from these is one, talking about sort of burden and proportionality,

00:04:14is that in some cases where huge amounts of data actually are relevant, parties have been required to produce it.

00:04:21So that's the open AI copyright case. They had to produce millions and millions and millions of logs,

00:04:28even after making all of their burden arguments.

00:04:31And then second goes to the point we were just talking about expert witnesses

00:04:35and basically the requirement to produce prompts where an expert was using an AI tool.

00:04:42This is just not talking about either case.

00:04:45We can talk about one of these, but one of them I can't.

00:04:48But with all of the cases, one of the interesting issues is the benefit argument.

00:04:54And I think where plaintiff's lawyers are doing a really good job and defense counsel should take some notes

00:04:59is understanding that your clients benefit by a lot of the data that you're saying is unduly burdensome to produce.

00:05:07You have huge data lakes that you are monetizing and using.

00:05:11And yet you're asserting to the court that, so you have to do something beyond the fact that you say it's too expensive

00:05:16and there's too much of it.

00:05:18The better approach is to start thinking about it, about saying we'll use statistical modeling,

00:05:24we'll use all sorts of other ways to get at it, or use an iterative approach, right?

00:05:28And I think you might find a better outcome or more.

Automatically generated; it may contain errors.

Study materials

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

Study notes

Why asking an AI tool a legal question does not create privilege automatically

Mari Grace explained that simply because a question put to an AI tool is the type of question that could be asked of a lawyer, or touches on a legal issue, that does not make the interaction privileged. Organizations need to be careful and conscientious about how people use AI tools to prompt or research anything connected to litigation, because that use has to be monitored closely.

Mari Grace noted there is somewhat more leeway for work product protection than for privilege. To preserve that protection, people using the tool should be clear, at the time they prompt, that they are acting at the direction of counsel in connection with litigation, or that what they are doing or saying reflects legal advice received from a lawyer.

Why enterprise tools matter for privilege and work product

Mari Grace stressed that in every instance, people should be using enterprise versions of AI tools when undertaking this kind of work. Without an enterprise tool, there is a significant risk of waiver.

Keeping AI-assisted expert work separate through dedicated work-product tool instances

A speaker who agreed with Mari Grace's point described telling experts to create a separate instance of a tool specifically labeled as a work product tool, so there is never any ambiguity about why it was used. The idea is to have one set of tools an expert uses normally and an entirely separate tool used only at the instruction of counsel.

The problem identified is that experts often co-mingle their use of tools, with no way to prove that one instance was used only for one purpose and another instance only for a different purpose. Without that separation, work product can end up discoverable because the uses are co-mingled and not kept clean.

Creating a dedicated work-product instance also allows counsel to turn on logging for that tool, producing a log that can be given to counsel showing the work was done at counsel's direction.

When an AI tool's output becomes discoverable despite work-product efforts

Even with a dedicated tool instance, there is a caveat for experts: if the AI tool's output has to do with the expert's methodology, or if the tool provides something the expert actually reviews or relies on for a report, that material can be discoverable. The participants agreed that experts generally should not be relying on AI output in that way in the first place, since doing so would undercut the purpose of having an expert at all.

Using deterministic AI to avoid reliance risk in expert work

One speaker described spinning up a separate instance of a tool for every single case and never using AI for opinions, relying instead on deterministic AI. The approach described is to feed the AI a brief and an opinion and direct it to go find specific forensic data, with all findings and data then verified and validated independently, so there is never any reliance on anything the AI itself produces. The speaker framed this as necessary because there is no room for error, since an error in that process could cause an entire case to fall apart.

Recent case law on producing AI logs and prompts in discovery

Mari Grace discussed recent decisions on burden and proportionality in AI discovery. In one case, the OpenAI copyright case, a party was required to produce millions and millions of logs even after raising burden arguments. A second decision addressed expert witnesses and the requirement to produce prompts where an expert had used an AI tool.

The benefit argument in burden and proportionality disputes over AI data

The panel identified the "benefit" argument as a recurring, interesting issue across these cases. Plaintiffs' lawyers have been pressing the point that a client benefits from the very data it claims is unduly burdensome to produce, for example large data lakes the client is monetizing and using. The point made is that a party resisting production has to do something beyond saying the data is too expensive to produce and there is too much of it.

The panel suggested that a better approach for parties resisting production is to propose alternatives such as statistical modeling, other analytical methods, or an iterative approach to get at the relevant information, which may produce a better outcome than simply arguing burden and expense.

Flashcards

13 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 13

Practice quiz

Practice — not graded, not the final exam

Question 1 of 6Per Mari Grace, what is required for an AI tool interaction touching a legal issue to be treated as privileged?
Choose an answer to check it.

Slides & materials