Legal Cyber Academy

Why AI evidence disappears before you ever request it in discovery

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

Up next

Who possesses AI data, and what happens when the AI itself is the actor
TranscriptThe player shows YouTube’s automatic captions, which are machine captions we do not review. The full transcript is below.

00:00:00I think it's really important, but I mean, well, but as I agree with that, but as your governance therefore gets operational and technical, that's a whole new category of people to to educate about what.

00:00:14You have to the implementation of governance into an engineering tech stack having just went through this building this front of this software, but it's totally different.

00:00:24Like you have to recreate and relearn and then teach all of your engineers a whole new lingua frank. It's quite an exhausting process.

00:00:34But once you get it, it works, but it takes a good three or four. It takes some time.

00:00:40Why AI evidence disappears before you ask for it? Not because the AI is devious. I just tell people.

00:00:48AI is efficient. So if you say be efficient in your use of space, getting rid of logs is probably an efficient use of space to store that out there.

00:01:00But retention windows are measured in days, not months, not years. I mean, days is being generous. Sometimes like one day, two day.

00:01:10Like these are huge amounts of data, right? Like.

00:01:15You know, people think like Verizon, like your cell records are kept in the phone number style.

00:01:21You're just it's a magnitude more of data. Like it just doesn't work.

00:01:26Here's the real kicker. Your customers, the corporations don't actually have the data and their contracts don't usually give them access to that data.

00:01:37Just throwing that out there. It's a real. So it's not that it's necessarily some vindictive thing. It's that when you contract with a global AI company, they usually don't agree to give you this information, like we'll stop.

00:01:52So and I don't know, I guess, Mari or Karen, you would know the contracts require preservation of this information.

00:02:01I can tell you. I've seen spoliation motions on it, but I don't know about the contracts themselves.

00:02:10And the deprecation as well of models are just done at the vendor's discretion, right? I mean, it's not.

00:02:20And when you and just so you clear, when you deprecate it, just your smoke, your meaning no longer supported next thing, you know, so it's important there.

00:02:30And consumer accounts are always the last thing people think about.

00:02:36I don't know, Mari or Karen, anything you want to add?

00:02:40Maybe just two quick things. The first is given how quickly some of this data can disappear, figuring out what is retained where before there's a preservation obligation is critical because, you know, getting a complaint or getting a demand letter and then spending two weeks trying to figure out where all this stuff lives is just not enough time.

00:03:03And some of it could be gone. So that's the first thing is like figure it out when you're not under the gun.

00:03:08And then the second is on the consumer accounts.

00:03:12Going back to what we were talking about with respect to cell phone data and messaging platform data from several years back, I think consumer A.I. accounts are sort of the next wave of that.

00:03:24And, you know, I think we will probably get to a place where the expectation will be that employers are supposed to ask their employees if they've been using consumer accounts for for work related tasks and then needing to go and collect that and potentially produce it.

00:03:40You know, even if it's a corporation, it's an actual named party and a litigation. So I think that's that's something for everyone to be aware of just because your employees are using consumer tools doesn't mean it's not your problem.

00:03:51Amen. And they're doing it on mobile. They're doing I mean, they're doing it wherever they are in the world. Right. So that's the worst part is that they have mobile versions of cloth.

00:04:00It is literally like from a forensic perspective.

00:04:07It's a lot. It's a lot. And there's still such a range of. So it's equally plausible that the senior executive you're seeking documents from.

00:04:20Has never used A.I. as it is that they use it to figure out where they want to go to dinner tomorrow. Right.

00:04:28There's no customary usage pattern yet that you want better. Like they're connecting these private these consumer grade personal assistance.

00:04:41They figured out how to connect them into the MDM solution to give them authorization to access your schedule, your whatsapp, your calendar, which then feeds into the A.I.

00:04:55Which then feeds into their model, which then gives them access, which, you know, it's fine. Like Google has all your data, but in litigation, it gets a little more tricky.

00:05:05And all these all these all these sort of agent sort of agenda chiefs of staff. But but I'm just what I'm saying.

00:05:12But it's equally plausible that somebody in the same role in another company just has never turned the machine on.

00:05:21We have people that don't even check email. They literally have someone. I have two classes right now where they literally don't have a computer, don't check email every day.

00:05:31Their admin brings a printed stack of emails that they need to respond to and they dictate response. So so from a discovery perspective, like what does that mean?

00:05:40I'm going to turn it around and ask you guys a question. Like, how do you structure validate that you're getting what you deserve? Right.

00:05:52Where do you set expectations and what how do you even begin to approach that?

00:05:58It's still an outline that somebody wouldn't use email. So I mean, so for me, it works with I, I usually just have talk with the IT people.

Automatically generated; it may contain errors.

Study materials

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

Study notes

Why AI Evidence Disappears Before Discovery Requests Are Made

AI-generated data often vanishes long before a litigant ever sends a preservation letter or discovery request. The lesson stresses that this disappearance is not the result of intentional wrongdoing by the AI system. Instead, AI systems are built to be efficient, and deleting logs is simply an efficient use of storage space. Because of this design philosophy, retention windows for AI-related data are measured in days, not months or years. In some cases the window can be as short as one or two days. The volume of data generated by AI interactions is described as a magnitude greater than people assume, far beyond something like traditional phone records, which are commonly (and incorrectly) used as a mental benchmark for how long data might be kept.

Retention Windows and Vendor Contracts as Barriers to Discovery

A key point is that corporate customers of AI vendors often do not actually possess the underlying data themselves, and their contracts with AI vendors typically do not guarantee access to that data. This means that even when a company wants to preserve or produce AI-related information, it may not have the contractual right to obtain it from the vendor. The lesson frames this as a structural problem rather than a vindictive one: global AI companies generally do not agree to provide this kind of data access as part of standard contracts. There was uncertainty among the speakers about whether contracts require preservation of this information, though one speaker noted having seen spoliation motions concerning this issue, without having direct knowledge of the contract terms themselves.

Model Deprecation and Loss of Access

AI model deprecation is handled entirely at the vendor's discretion. When a model is deprecated, it means the model is no longer supported going forward. This vendor-controlled process adds another layer of risk: even if data or outputs exist in some form, the model that generated them may cease to be supported, affecting access to and interpretation of that information. Consumer accounts are described as consistently being the last consideration in discussions about AI data retention and access, making them an overlooked risk area.

Mapping Data Locations Before a Preservation Obligation Arises

Given how quickly AI-related data can disappear, the lesson emphasizes that organizations need to determine what data is retained and where it is stored before any preservation obligation attaches. Waiting until a complaint or demand letter arrives and then spending weeks trying to locate this information is not workable, because by that point the data may already be gone. The practical guidance is to do this mapping work proactively, while not under time pressure from active litigation.

Consumer AI Accounts Used for Work-Related Tasks

Consumer AI accounts are identified as the next wave of a problem previously seen with personal cell phone data and messaging platforms. The expectation is likely to develop that employers must ask employees whether they have used consumer AI accounts for work-related tasks, and then collect and potentially produce that data if the employer is a named party in litigation. The lesson stresses that the fact employees are using consumer-grade tools rather than company-sanctioned systems does not make the resulting data someone else's problem; it remains the employer's responsibility to account for it.

Mobile Use and the Forensic Difficulty of Tracking AI Activity

Employees use AI tools on mobile devices wherever they happen to be, which adds forensic complexity. From a forensic perspective, tracking this kind of distributed, mobile AI usage is described as a significant challenge.

Unpredictable AI Usage Patterns Among Senior Executives

There is no reliable, customary pattern for how senior executives use AI. It is equally plausible that a senior executive sought in discovery has never used AI at all, or has used it only for something trivial like picking a restaurant. Some executives connect consumer-grade personal AI assistants into their organization's mobile device management (MDM) solution, giving the AI access to calendars, messaging platforms, and schedules, which then feeds into the AI model itself. This arrangement may be unremarkable in ordinary use but becomes legally complicated once litigation arises. The lesson also notes that some executives do not use email or computers directly at all; instead, staff print out emails for them to review and dictate responses, which has direct implications for how discovery requests directed at such individuals should be framed and validated.

Open Questions Raised About Structuring Discovery Expectations

The lesson closes by raising, rather than resolving, practical questions for discovery practice: how to structure and validate that a party is receiving what it is entitled to in discovery, how to set realistic expectations given the variability in individual AI usage, and how to even begin approaching verification when usage patterns are so inconsistent across individuals and organizations.

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 7Why does the lesson say AI evidence often disappears before a discovery request is even made?
Choose an answer to check it.

Slides & materials