Who possesses AI data, and what happens when the AI itself is the actor
Free. Watch without an account. Sign in to save progress and take the exam.
Up next
Building effective discovery requests for AI artifacts using interrogatories and Rule 26(f)TranscriptThe player shows YouTube’s automatic captions, which are machine captions we do not review. The full transcript is below.
00:00:00And then I have a truck with the individual directly like, so how do you make dinner plans? And then sort of work backwards from that. Right.
00:00:09I mean, I'm dealing right now with two active cases where they, you know, they they have access to their own mobile devices and they did what they did.
00:00:22Right. And the corporation is trying their best to salvage. I will tell you, it's like a band-aid. So you should always tell the truth and just move forward.
00:00:31But where you can really get dinged is if you hide it. Yeah. Not a good look.
00:00:38Who possesses the data? It's pretty simple. You possess some of it and then it sits everywhere else. Everywhere else. No, I mean, and so the rule 34 question.
00:00:51Of a possession, custody or control. Like I said, we have mobile device that has Claude installed on it.
00:00:57Then the question is, it's the personal mobile phone, but you're using Claude and it's I mean, just complicated. Right.
00:01:04And then, you know, who controls and then you look at all that. And then as the arbitrator, I can't countlessly now entertaining subpoenas for like.
00:01:18Third party subpoenas for records and information more in domestic arbitration is an international arbitration, but nonetheless, they still come.
00:01:27So so when the AI is the actor, so I'd love to hear more of your thoughts as to how this works, because autonomous agents don't leave anybody to depose.
00:01:42Yeah, I was just in a very heated debate with someone about personhood and, you know, where and when we get to a place where we will be deposing agents or attempting to depose agents on what we are not there yet.
00:01:58Well, the design of agents today. So you to query. And the question is, what are you to put like an agent now that you're using changes?
00:02:10But like the rate of change in learning in an agent, depending on how it's built, is so complex that I don't even get even if you didn't smoke the agent, which probably you do anyways.
00:02:23Even if you had the agent, how do you know what state the agent was in at the time of the issue? Right. No, you don't. I think I think over time, I'm hoping that someone creates some mechanism for this to basically improve like audit trails for agents so that, you know, maybe we someday get to a place where agents are being deposed instead of human beings.
00:02:45But I don't think we're there. We're working on it, actually. That's something we're specifically working on is taking the audit logs out of the artifact.
00:02:53Deconstructing the agent to then see how you can create an audit trail to actually understand its actions, not based on the metadata of its use. Right. Not the data, but the metadata of its activity.
00:03:06Yeah. And you can and you can add whether you believe it or not, you can ask these agents very often to explain themselves and their steps and their.
00:03:14No, 100 percent. But the problem is, so let me be clear, you can depose the agent. The problem isn't deposing the agent. The problem is the knowledge the agent accumulates before you're opposing it. The problem isn't that the contract of the posing an agent will be dicey in itself.
00:03:32The problem is an incident happens. The agent likely will have access to the data on your network, realize an incident has happened and self correct.
00:03:41So then the question becomes, and what state and memory retention? Because the real problem the agents contend with is their memory. Right. So the, you know, it's a bunch of goldfish. Right.
00:03:53Yeah. Yeah. Yeah. So it becomes tricky because you could depose the agent, but the answer is giving you will be based on a much more perfect set of information.
00:04:03Yeah. And I think because we're, you know, at least currently living in a world where human beings are still going to be responsible and liable for the acts of agents.
00:04:12The best we can do is give them clear instructions, keep records of those instructions, see what accesses they had to various systems, you know, what tools were in place, what harnesses you had, what sorts of guiding instructions you had at every level.
00:04:26Because I think more and more, we're going to be in a situation where human beings who set up a gentle workflows are going to be in the position of just needing to explain themselves and saying, I took every possible reasonable step that I need to take at this point in time.
00:04:40And let me show you all the stuff that I did.
00:04:44And then I think I'll be like, can I just double click on the one thing you said, Mario, which is date stamp everything. Right. Because the what's going to hallucinate the date stamp, though.
00:04:58No, no, you check it. But but I guess my point is that if you're litigating over something that happened in 2024, what's reasonable to expect, you know, if you're litigating in 2028 about something that happened,
00:05:12in 2024, remembering what was reasonable to expect in 2024 becomes very important. Right. Because by 2028, Daniel's company will have solved this problem. Right.
00:05:24Just that solution. But that's but that solution just won't have been available at the time.
00:05:31I mean, it's right. And it's so complicated because.
00:05:35And then the date stamp thing, what we're finding is that the metadata that's embedded in an AI generated document hallucinated.
Automatically generated; it may contain errors.
Study materials
Notes, flashcards and practice questions for this lesson, reviewed before publication.
Study notes
Possession, custody, and control of AI data under Rule 34
When someone uses an AI tool such as Claude installed on a personal mobile device for work, the discussion frames this as a Rule 34 problem: is the resulting data in that person's possession, custody, or control, and is it also within the corporation's control because of how the tool was used? The point made is that the data "sits everywhere else" as well as with the individual - it is not neatly held in one place. Working backward from a simple question like "how do you make dinner plans" was offered as a way to trace how access and control actually work in practice.
Corporate response when employees used their own devices
In active matters discussed, individuals had access to their own mobile devices and used them as they saw fit, leaving the corporation trying to manage the aftermath. The guidance given was straightforward: tell the truth and move forward, because hiding the issue is what causes real damage ("not a good look").
Arbitration subpoenas involving AI-related records
Speaking from the arbitrator's seat, the discussion notes that arbitrators are still fielding third-party subpoenas for records and information connected to these issues. This happens in domestic arbitration and, even more so, in international arbitration - the subpoenas keep arriving regardless of forum.
Deposing an AI agent: the personhood debate
When the AI itself is the actor, autonomous agents leave no person to depose. The lesson describes an ongoing, "heated" debate about personhood - whether and when the law will reach a point of deposing, or attempting to depose, agents directly. The consensus in the discussion is that we are not there yet.
Part of the difficulty is how agents are designed and how quickly they change. The rate of learning and change in an agent, depending on how it is built, is described as so complex that even preserving the agent itself does not tell you what state it was in at the time of the issue in question. Having the agent is not the same as knowing what it "knew" or how it behaved at a specific past moment.
Agent memory and audit trails as the real discovery problem
The hope expressed is that someone will eventually create a mechanism to improve audit trails for agents, so that agents might someday be deposed instead of (or alongside) human beings - but that capability does not yet exist. One effort described is taking audit logs out of the artifact itself and deconstructing the agent to build an audit trail based on the metadata of its activity, not merely the data it produced. Agents can also often be asked directly to explain themselves and their steps.
Even so, the deeper problem is not the mechanics of deposing an agent - that part can be done. The real problem is the knowledge the agent accumulates before it is deposed. If an incident happens, an agent with access to network data may recognize that an incident occurred and self-correct, changing its own state before anyone examines it. This connects to the agent's memory and retention: agent memory is described as being like "a bunch of goldfish" - limited and prone to not holding onto the full picture. As a result, an agent's answers under deposition would reflect a more complete or "perfect" set of information than it actually had at the relevant time, which undermines the reliability of the testimony.
Human responsibility for agentic workflows
Because human beings are still the ones held responsible and liable for the acts of their agents, the recommended approach is to give agents clear instructions and keep thorough records: the instructions themselves, what systems and data the agent could access, what tools were in place, what harnesses were used, and what guiding instructions existed at every level. The expectation is that humans who set up agentic workflows will increasingly need to explain themselves - showing that every reasonable step available at the time was taken.
Date-stamping, metadata, and reasonableness over time
Mari Grace raised the point of date-stamping everything as part of this recordkeeping. A caution followed: date stamps themselves need to be checked, because they can be wrong rather than simply trusted at face value.
The discussion also stresses that reasonableness must be judged against the time of the conduct, not the time of the dispute. If a matter from 2024 is litigated in 2028, what was reasonable to expect in 2024 is what matters - even if, by 2028, better tools exist to solve the same problem. A solution that exists later does not mean it was available or expected at the time in question.
Finally, the metadata embedded in AI-generated documents was flagged as a practical problem: in some cases, that metadata itself has been found to be hallucinated, meaning it cannot automatically be relied on as accurate.
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