Season 2 Mid-Year Recap: Five Questions Our Guests Answered and the Courts Tested

Brandon, we have recorded 10 conversations this season. We've recorded judges, a dean, a chief justice, a attorney CEO, a research scientist, forensic examiners, in-house strategists, etc. And somewhere, Brandon, in the middle of all of that, the courts started answering the exact questions that our guests have been raising.

You're right.

We would ask a guest where the line was on something like prompts, or invalidation, or on authentication, and a few weeks later we'd see a federal judge draw the line in the sand with a specific opinion. So today is not just about a highlight reel. It's really about a reckoning of sorts. We're going to put our guests in conversation with the case law that has landed in the first half of 2026, and there's been 5 questions that have come up throughout every single one of the conversations that we've had, that our guests have answered, and 1 then what the court did with it. So this is going to be not just a mid-year recap, but it's going to be where we're actually looking at for the latter half of 2026.

Let's get into it. Welcome back to another episode of The NeLI Pod, the official podcast of the National eDiscovery Leadership Institute. I'm Daniel Gold, here as always with my co-host, Brandon Mack, and this is a special episode today because we don't actually have a guest today. It's going to be the two of us looking back at all the guests that we have had so far in season 2. And we've picked 5 questions that have come up over and over again this season. For each one, we're going to bring back what our guests told us, word for word, and then set it against the generative AI discussions and decisions the federal courts have made in the first half of 2026. Some of those you've read about, a couple you may not have heard about at all.

And to be clear about the method here, we're going to be pulling out really memorable quotes that we've had from each one of our guests so far, and their words from each of the recordings. In every case that we cite, we pull the opinion, and we are reading the court's language directly because we have seen this correlation between what our guests have said and what the courts have been saying, which just goes to show of the incredible brilliance of the guests that we have here on The NeLI Pod. So we have 5 topics. Let's start where, interestingly enough, each one of our judges started with their discussions. All right, so the first topic we're going to get into is the rules still hold, as Brandon and I are calling it.

So we had the first question here, and that is this: when technology moves faster than the rules, what is the court's job? Do we rewrite the rulebook, or does the existing framework already carry weight? Brandon, I think that our judges were remarkably consistent on this, and they didn't actually confer with each other.

No, they were. Chief Magistrate Judge Young Kim of the Northern District of Illinois put it about as plainly as it can be put. He said, and we quote, "When the technology surpasses the rules, we still have the same role of reviewing the evidence before us and making the best possible decision." And he added a line that has come back to me over and over since we had this interview, saying, "The federal rules of civil procedure have to be our anchor," and I really do mean that. I love that because there is so much in that I think that we were almost debating with the judges as well when we recorded those episodes, about what they really thought about the rules. And when we interviewed Judge Rodriguez of the Western District of Texas, and we talked about the hallucination panic, as we're calling it, he really felt the same way.

He said, "Everyone wants to treat Gen AI as the exotic new threat to the system," and his answer was— I mean, I felt like he was almost impatient, in some respects, with the framing. And he said, "It's not rocket science, folks," those are his words. "You can— you use an AI tool, and all you do is you check whatever citations the AI tool has given you. You want to make sure that those are real cases. " And he made the point that I don't think a lot of people make, or they should be making much more, and that is: the hallucination databases, they only count the failures.

And it reminds me, Brandon, of when we— before Gen AI, we used to talk a lot about the sanction cases. If you don't preserve your cases, there's going to be spoliation. If you don't put down a legal hold, there's going to be spoliation, etc., right? There's going to be all these adverse inferences. Gen AI is just another layer to it. And that's what Judge Rodriguez said. He said, quote, "No one's keeping track of how many times a lawyer has responsibly used one of these AI tools, and it has correctly identified the cases or statutes, and it worked as intended. " And I think the impact of what Judge Rodriguez is saying, Brandon, is that if we're only talking about the hallucinations, then we're never talking about the benefits of using it.

We're only talking about the risks. And if we're only talking about the risks, then we're not encouraging lawyers to actually use AI.

Absolutely. And moving away from the judges, one of our other guests, John Benson, who's a local security and forensics guest that we had here in Kansas City, he took it out of the context of litigation and looked at it more as an evidentiary question. One of the things that John said is, "Where we have problems, it's not with the rules, it's in the application of the rules. " The way he was framing it is the way that every trial lawyer should learn to frame it, but seemed to have learned these concepts and then forgotten them. When you strip away every rule of evidence, there are 3 questions that remain: relevance, reliability, and whether there's unfair prejudice that's in play with the use of this evidence.

He said the question of reliability is harder now when we're talking about what we have with generative AI, but the question itself, and the procedures of answering those questions, hasn't changed at all.

You know, it is an interesting point. And John was actually a very interesting guest that we had on the show, for sure. His insights come at it from a very, very different direction than obviously the judges. And so there's 3 guests, 2 of them on the federal bench, all saying the same thing: the rules are the anchor. So now, here's what makes this season different. A judge in the Southern District of New York said the same exact thing in an opinion that all the law blogs are talking about right now, and that's USV Hepner. And this is the securities fraud case where the defendant, Hepner, used a public version of a Gen AI platform to draft his own defense strategy after he knew he was a target.

And that's an important part, that he knew he was a target, and then shared what he did on a Gen AI platform with his attorney. So the government moved to get those documents, and the judge's framing is the sentence to read to the whole bench. And he said— the judge said, "Generative artificial intelligence presents a new frontier in the ongoing dialogue between technology and the law, but AI's novelty does not mean that its use is not subject to longstanding legal principles." And that's Judge Kim's point from our interview, just placed in a published decision. The novelty of a technology does not create some sort of exemption to the rules or how the rules are applied. Judge Rakoff also acknowledged how early we are, writing that the implications of AI for the law are only beginning to be explored.

We'll get deep into the holding on Hepner in the next segment, because it's a privileged case at its core, so there's more to talk about with it. But as a statement of philosophy, it belongs here in this portion of the discussion. So let me put the closing question to you on this one. If the rules already carry the weight, why did 3 of our guests still sound worried about it?

That's a great question, and I think that's because the anchor holds the ship. It doesn't steer it. Kim, Rodriguez, Benson, all trust the framework, right? What they do not trust, it sounds, is that lawyers will do the work the framework assumes, which is checking citations, laying a foundation, supervising the tool. And so— and it's an important point there, because the rulebook itself has not failed. It's— if we as lawyers stop reading the output, then there's the failure. And so I think that's the through line for the whole episode, and it takes us straight into the prompts.

So let's talk about topic 2. So topic 2 is prompts, privilege, and work product, and we grouped the next guests together very intentionally. So the second question is this, Brandon, and this is the one that gets probably the most heat, which is: when a lawyer types a prompt into a Gen AI tool, what is that prompt? Is it protected work product? By the way, this did come up at last year's NeLI Conference as well. And is it a discoverable document like anything else? So Robert Keeling from Redgrave, he did not hedge for a second on this one, right? On the outward-facing prompts, the ones you use to identify responsive documents, he said unequivocally— he said, "The prompt is clearly work product, and it's not even a closed question." I love that it was just— he did not even pause when we asked the question of him.

And when we pushed him on why he was so certain, what did he say? He said, "Because it reflects the thoughts and mental impressions of counsel, including me when I'm doing it about the case, about both the claims and defenses of the case. It's really just classic opinion work product. " I too love the confidence with which he said that, that it's not even a closed question. But he went on to talk about some other aspects of the question around what these prompts really are, and one of the things he said was it almost sounded like a warning to the plaintiff's bar, or a prediction, in that he said that all it takes is 1 or 2 judges to rule in the opposite way to completely slow down the adoption of generative AI.

Using what he said actually in the interview, he said, "If they are trying to litigate for the prompts, I think all we need is 1 or 2 judges to rule that the prompt is discoverable, and you're going to significantly slow down and retard the process and the use of Gen AI in eDiscovery. " In addition to Robert's warnings, we were talking to someone else, one of our other guests, Steve Davis, who came at it from the forensics side, and he told us the uncomfortable truth, talking about what's available to be discovered from these prompts. He said, "Whatever the doctrine says, the data exists, and it is reachable right now through forensic methods. " One of the things he said specifically was, "There are 3 categories of data you can get from AI platforms.

One is clearly the prompts. Number 2 is the output, of course, the answer that the Gen AI gave you. And then the third one is really the source code of the way the large language model operates. " And when I pressed him further and asked whether we can get to the prompts today, he was direct in saying, "Can you get them today? Yes, you can. With the appropriate credentials or admin permissions, you can get to them legally. These are structured databases, and we're dealing with really textual or linear data, which is pretty easy to manage. It's not a particularly heavy lift. " I really enjoyed hearing him say that, and it was basically foreshadowing, I think, some of the things that we're starting to see.

And when we talked to Mike at JS Held, he gave us a different view of this, and he said that it makes the privilege question make much more sense. He said, "Treat the tool like a person you are responsible for." So his words exactly on the episode were, "Think of it as bringing on a really talented young lawyer who went to a good law school and passed the bar, but they don't know anything yet. " And, you know, Brandon, you and I have talked about this as well. This is, you know, "Treat Gen AI like a junior associate. " And I think that was really Mike's view. And then, critically, on the ethics piece, Mike reminded us that the responsibility does not move. And he said, specifically, "The duty to supervise still exists.

Whatever output that individual has, you're ultimately responsible for." That's right. So those are the practitioner views. Keeling says the prompt is opinion work product. Davis says it's sitting there, available, and databases waiting to be produced. And Gaudet says you own the output either way. So this seems to be, in the first 6 months of 2026, where the court's did most of their heavy lifting in that time period. It's true, and, you know, I think that's right. And they also split. So start with the case that we opened with, Hepner. So the judge there held that the defendant's exchanges with the AI tool were not privileged and not work product, because he did it on his own, on a public platform, and then shared it afterward.

The waiver language is the part that we need to really focus in on. So in that opinion, it states, quote, "Even if certain information that Hepner input into Claude was privileged, he waived the privilege by sharing that information with Claude and Anthropic just as if he had shared it with any other third party. " Brandon, that's— that's wild, isn't it?

It is. And it's pretty definitive the way it's stated there in the opinion. But here's the split, and it's a pretty clean and bright split when you look at it, because 1 week before Hepner, a magistrate judge in the Eastern District of Michigan went the other way on different facts for a pro se plaintiff. In Warner versus Gilbarco, the court held that the plaintiff's AI-assisted materials were protected work product, and it rejected the idea that using a public tool waives that protection. Its reasoning cuts against the third-party theory in Hepner. The court held that generative AI programs are— quoting from the court— "tools, not persons, even if they may have administrators somewhere in the background," and that work product waiver requires disclosure to an adversary or in a way likely to get into an adversary's hand.

Okay, so now let's take that case, right? Let's take the Hepner case. And then what happened in March? The District of Colorado went deeper on the same logic. And so this is the Morgan v. V2X matter. So the Morgan court distinguished Hepner on a point that matters for every pro se document in the country. So the court there said, quote, "A pro se litigant is simultaneously the party and the advocate. " And then it reached back to Fourth Amendment law, to Carpenter v. United States, to say that the routing information through a third party does not automatically extinguish your privacy in it. The court wrote that AI systems collecting user data, quote, "does not eliminate all expectations of privacy or automatically waive protections. " And that brings us back to Robert Keeling's core claim, the idea that the attorney's prompts are opinion work product.

And while it was his claim on our interview, it's not just his opinion anymore. The Northern District of California said the same thing in the Concord Music litigation against Anthropic. Those courts treated attorney prompts and outputs as opinion work product, described as virtually undiscoverable under Rule 26. But— and Keeling would appreciate this— waivable when a party puts the substance of its investigation at issue in the matter itself. It's a sword and shield issue. You can't rely on the prompts to build your claim and then hide the rest of it. So our closing question on this one, Daniel: Keeling said 1 or 2 judges ruling the wrong way could freeze adoption. We now have judges on both sides. Do you think his fear is coming true, or has come true?

I mean, partly, and in maybe a more interesting way than even Keeling sort of predicted, let's call it, right? So the prompts themselves are holding up as work product when a lawyer drives them, and that is the Concord line, and it protects the Keeling workflow. What is not holding up is the client freelancing on a public tool and expecting privilege to attach later. That's Hepner. So the doctrine is drawing exactly the line our guests drew, right?

So this is so interesting. So lawyer-directed use inside a controlled process does, in fact, get the appropriate protection. So on a target of an investigation, talking to a consumer chatbot at midnight obviously does not. So the tool, again, does not change the rule, right? The relationship does. And I think that's really what we really have to focus in on, and what our guests really focused in on, is that the tool is just another tool, and it is ultimately, at the end of the day, the responsibility of the attorney to supervise the tool and to make sure that the tool is working correctly before submitting it to the courts. So that leads us into a very, very interesting segue, I think, which is our next few guests that we're going to talk about, and really topic 3, which is validation, defensibility, and the open versus closed line.

Okay, so if we are going to trust these tools to decide what is responsive and what is privileged, then the question becomes: how do we prove that the process worked? How do we say that Gen AI review, right, is something new or something we already know how to validate? And Brandon, Mike, our friend Dr. Jeremy Pickens, I think, gave the best reframe for us, not just of Season 2, but of Season 1 as well. Brandon, walk us through.

Yeah, I have to admit this was one of my favorite interviews of our season so far, being the kind of torn analytics geek that I am. Getting to talk to Jeremy Pickens and pick his brain, as it were, about this new development in the use of AI in eDiscovery was really fascinating. And when he was talking about it, one of the things he did, one of the things that he said, is something that I think will make a lot of vendors and a lot of software developers in the industry rather uncomfortable to hear his opinion. But his position is that generative AI in the use of generative AI in review is not really a break from TAR. In his opinion, it is TAR. In his own words, he said, "It is a TAR process because TAR is a process and not a predictive model.

" From that, he went on to draw the conclusion that you could say is actually a warning, that if Gen AI review is a TAR process, then all the validation discipline that we spent 15 years learning still applies to the way we use Gen AI. And in his opinion, he feels like the industry is kind of throwing it away, looking at Gen AI as a new tool that doesn't require all the things that we used to do for TAR. And this is the line of all the things that he said that I would put up on the screen. It's his statement that we've forgotten what control sets are, and we're now like surgeons who go into surgery without washing our hands first. Well, this is different from what it used to be.

I write prompts, and so germ theory doesn't exist. He said, "No, germ theory is still there. " The idea that just because we've advanced along the line of how we use AI, that therefore the foundational principles we built on how we validate AI no longer apply, he clearly doesn't believe that's the case for what we're looking at with Gen AI.

Well, and you said it yourself, Brandon. You are a TAR and analytics geek, right? So you grew up in this, like I did as well, with TAR 1. 0. Dr. Pickens has doubled down on this, right, about the whole idea that Gen AI is a TAR 1. 0 workflow. Do you agree with that?

I do agree. I mean, when you think about the differentiation between TAR 1. 0 and TAR 2. 0, it's how often are you sort of updating the model, is the term we used to use for it, or the phrasing we used to use for it. With Gen AI, you're walking into a model that's already done. You're not using something that you're training additionally. You're just giving it the documents on which you want it to act. That's the same thing you're doing once you establish and stabilize a TAR 1. 0 model. So his analogy that what we're looking at with Gen AI is really TAR 1. 0, yet it's still an advancement from what we had with TAR 2. 0, I think is absolutely correct. So theoretically, if you have a bad prompt and you rely upon the output of said bad prompt, then you are essentially rerunning everything off of the output of the bad prompt, and so you're going to get the bad output.

Absolutely. Yeah. That's interesting. And Jeremy also named a specific failure he's watching vendors commit right now, which is he says when you build a prompt against the same documents and how you measure it, you have contaminated it. And so this goes right to what you were just talking about, Brandon. He says the problem is that you are training on your test set. So Brandon, I mean, you've basically nailed it here, is that you've got a— that is the germ theory, right there. That is 100% the germ theory. And Robert Keeling, from the practitioner side, agreed on the substance while lowering the temperature a little bit from Dr. Pickens. And so he does not think Gen AI needs some exotic new validation regime. His words were, "There's nothing unique or special with Gen AI that calls for any unique or special validation protocols to be applied to it, the same validation that we've been using with respect to TAR." And you— that's exactly what you're thinking.

And you are, Brandon, an analytics and TAR whiz, so this falls right within your, you know, your wheelhouse. Do you agree with this?

Absolutely. I absolutely agree. I think that the things that we've learned to do over the last decade, decade and a half, in validating TAR, in monitoring how TAR's outputs are really coming to us, they still apply. There's no reason why we need to change that. And so maybe the output needs to be that we need to retrain some of the classic TAR lessons when utilizing Gen AI tools today.

Absolutely. And, you know, on that, sort of bolstering what both Dr. Pickens and what Robert Keeling were saying, we also interviewed Amy Amberson at United Lex, and she talked about what does it look like to really adopt these tools when you're doing it at a large-scale organization that is executing at scale with these tools. And she said, for them specifically, you know, "We have a steering committee. We test all of those things. We vet them. " So it's providing some of the proof behind it. And that's really the way we need to look at approaching the use of these tools. It's not that we're changing things in how we validate. It's not that we're doing anything foundationally different from what we used to do. We did the same thing with TAR tools a decade ago.

You would test different tools against each other. You would test them against your processes to figure out what worked. And she's pushing for the same sort of slow and considered approach to adopting these generative AI tools. And one of the things that specifically people talk about with generative AI, because the model is pre-built, it's not something that you are building as you're going, is that the model may have some bias in it that then interacts with your prompts in a way that gets you results that you don't expect. Her response to that, or on that point, was, "As a human, I'm biased. So if we're concerned that the machines are going to be biased, we're dealing with that all the time. What is the bias that might exist, and how can we counter for it or account for it?

" And the thing that I loved about this answer is, when you're taking that slow and considered approach and testing and vetting things, one of the things that you're doing is recognizing where the weaknesses are and figuring out how to build processes that will deal with those weaknesses in a way that still gets you the output and the result that you want. It's not saying, "This is a magic easy button, and we don't have to look at it." But it's also not saying that, "Oh, this is just completely flawed, and we can't get around this. " It is considering this, vetting it, and figuring out how you build processes around it that make it work the way you need it to.

Well, that just bolsters and reinforces the point that the quality control check of the output of Gen AI is so important. If you are worried, as, you know, Amy's talked about, if you are worried about bias, then that furthers the need for human oversight, human review. It's why we have the rules of professional responsibility that we do, right?

Absolutely. Because it says that we have to— it's no different than an associate who drafts this for a partner. The associate drafts it for the partner, and the associate has some sort of inherent bias about what she thinks, about what the case direction goes. That bias is going to come through in whatever the draft is that she hands over to the partner. Then it's the partner's responsibility then to say yes or no.

Absolutely. Right?

You know, Judge Kim told us exactly what persuades him when a process is defensible or not defensible. And what he said was, "It has to be plain English information that the expert can tell me that I feel good about the fact that this process is indeed reliable." And with that, you know, when Judge Kim said that, what it made me feel like is that if you can't explain the AI output to the bench, or if your answer about your AI usage to the bench sounds like AI, then you have literally no idea what you're talking about, right? I mean, he's saying it's got to be in plain English. Explain exactly what it was. Okay, so this is the prompt that I used, and this is the output that I got.

This is the refinement that I used. And walk through that workflow. And if you can walk through that workflow in plain English and defend the use of it, not in a, like, "I have to defend it," but, you know, "I have to defend your use of it and how you utilized it," and it's in plain English, that's what Judge Kim is saying.

Right. So our guests on this question seem to all be aligned. Validate it like it's TAR, because it is. Test before you trust, and explain it in plain English. So, Daniel, the courts added another dimension to this that our guests only touched on, which is, where does the data that's involved in this process actually physically go?

Okay, so the narrowest finding. I'm so happy we get to talk about this one. This is the Kansas decision. Judge Mitchell, who she has been at many, many NELLI conferences before and is a great speaker on the panels that we have at the NELLI conference. And I was so happy to see this come out of the state of Kansas, because that's exactly where we are, Brandon. And it's one where I think every eDiscovery team needs to read, okay? This is, in case we didn't say it clearly, the District Court of Kansas. So it's a federal decision, and it's about protective orders. This is the Jeffries versus Harco's Chemicals matter. And this is a putative class action by residents living near a chemical facility, and the defendants ask the court to bar parties from uploading discovery materials into public open-loop AI tools.

So I want to reinforce that. That means, like, a consumer, not enterprise, a consumer version of the Gen AI tool. And here's what surprised people. The court extended that bar to non-confidential material too. The court found good cause and granted it in the reasoning, in the court's own words, "The wholesale submission of discovery materials to an open AI tool for these eDiscovery tasks could expose massive amounts of data. " Brandon, get into this one for me.

The practical concern that the court flagged is the one that Jeremy and Steve Davis would both recognize from the conversations we had with them. Once the data's in the model, you can't get it back. The court found that clawing back or deleting data submitted to an open AI tool is, quote, "impossible as a practical matter because such data is used to continually develop and improve the model. " So the line that the court's drawing here is open versus closed. You know, open, not speaking specifically about any company being an open AI company, but whether or not the system itself is closed to the rest of the public. You can use a closed secure tool that is contractually barred from training on your inputs and can delete on request, but you can't dump your discovery into a public consumer tool where others have access to what the model is learning from your data.

And notably, it put that same burden on every party equally when it comes to litigation. So I think our closing question on this specific topic is, Jeremy was worried that we've forgotten things about that we've learned from TAR, things like how to validate. The Kansas court is worried about where data lives. Are these the same problem, related problems?

Well, I think they're two halves of one problem. And the problem is discipline. Jeremy's talking about intellectual discipline. Hold out a control set. Don't grade your own homework, basically. And the Kansas court is talking about custodial discipline. Know where your data goes and whether you can pull it back. By the way, this is no differently than the rules of professional responsibility 1. 6(c), right? That's literally what we're talking about here. You have to know where the data goes.

But doesn't that go back to what we talked about all of our 10 guests so far as well, Brandon, and also last season, which is rule 1. 1, which is, in order to be competent, you have to understand the risks and benefits.

But I always say this. I contend that you can't be knowing— you can't uphold 1. 6(c) if you're not upholding 1. 1.

Agreed.

Right? There's no— you can't do that. So I think both are versions of the thing that Judge Kim asked for, which is the ability to stand in front of a court and explain in plain English what you did and why it was reliable. Now, a closed tool, right, so an enterprise version, something behind the firewall of an organization, with a validation protocol lets you give that answer. So an open tool with no control set leaves you really with nothing to say. But we have a lot more to say, Brandon. We have a lot more to say.

So let's transition into topic four, which is talking about authentication. And what I find deeply interesting is the deepfake frontier. So we've grouped together some other guests so far for this topic. So the fourth question really is this. This is the one where the guests seemed most unsettled, and for good reason. I think the courts are also. We have spent three segments on using AI to find the truth in the documents, right? Obviously very important. This talks about the other side of it. What happens when the document itself is a fake?

And you may recall, Brandon, that I raised this with several guests using a very small and silly example. I could sit down with a consumer AI tool and generate a receipt that says that you and I bought coffee together down the street, and it'll have the last four digits of my corporate card on it, and it literally looks authentic. It even looks wrinkled. And Robert Keeling did not sugarcoat the answer. He said, "It's an incredibly difficult problem, and unfortunately, the problem is only going to get worse and harder as the technology continues to prove," which doesn't really make us feel very good about this, Brandon. And if that one made you feel bad, Judge Kim piled on. His answer is really one that's worrisome for litigators and practitioners in that he said the deepfake problem doesn't really land on the judge.

In his opinion, it lands on the lawyer first. Quoting him, he said, "I don't know that it's going to make my job difficult. I think it's going to make the litigator's job very, very difficult, because during discovery, you're now going to have authentication issues to deal with." And he pointed out how much this changes how trials are run today. In his words again, "I can't remember the last time where I tried a case or presided over a trial where we had an authentication issue. It's just not an issue. But with Gen AI-produced evidence, it's going to become more challenging. " Yeah. And John Benson made the point that everyone, I think, should be thinking about as well. He's like, "We're looking in the wrong place," right?

Among the many interesting things that John told us, he said, "Everyone wants to talk about the dramatic corporate deepfake, the cloned CEO voice. " But Benson said that it's not where this surfaces first. He said, "It's going to surface in the family courts," which, by the way, I thought was very interesting, right? He said, "It's going to surface in those courts where judges are dealing with 10 family law cases in the afternoon when they've got a criminal docket to run in the morning." And as if these fears weren't enough, Steve Davis added another wrinkle to it, kind of piling on different eDiscovery challenges together, but one that makes authentication of this type of data even harder. And that's the fact that for some of these applications and ways in which you can create this information, the underlying data is often designed to vanish.

On ephemeral messaging, he said specifically, "Ephemeral, it's a cool word, and everyone loves it, and it really just means transient. Basically, the data that's there one day and then not there the next. " So put those together, Brandon. Evidence that is easy to fake and data that is designed to disappear landing on the highest volume, lowest resource dockets in the country. And Brandon, that is the one segment where I need to be direct with you and all of our listeners, and that is talking about the privileged questions, right?

So the privileged questions from the last two segments got answered really fast. Authentication has not. There is no headline federal decision from the first half of 2026 that solves the deepfake authentication issue, because the tool for that is the rules of evidence, and the rules, as we know, move very slowly. And Keeling told us exactly what is pending. His words, "Right now, we have before the Rules Committee a proposed Rule 707." This is going to be so interesting, Brandon. All right. So Rule 707 relating to AI, he says, "Although it's not addressing deepfakes head-on, it's more about compilations like the like being used for trial purposes. " And I think that's the gap. Proposed Rule 707 addresses machine-generated evidence offered without an expert. But as Robert said, it does not take on synthetic evidence directly.

You made that timing point yourself during the interview.

I did. And changing the federal rules of evidence is a three-to-five-year process. And so multiple guests have said that to us. The technology moves in three-to-five-month cycles now, right? The way Gen AI is moving, we used to talk about this in three to five years. Like, what's the technology going to be like in three to five years? Now it's just like, what's it going to be like next month, really, right? So for the near term, the direct answer, I think, to all of our listeners is that authentication of AI-generated evidence is going to be handled case by case, judge by judge, on foundation of reliability using the rules that were written before any of this have existed. And there's one decision, actually, that's worth flagging because it shows courts already treat AI interaction data as ordinary evidence.

And that is the In re OpenAI copyright litigation infringement litigation matter with the discovery order requiring production of a 20 million record sample of de-identified ChatGPT logs affirmed in early 2026. So that matters here because it tells you that the prompts and outputs are discoverable at scale in the ordinary course. The evidentiary system is already ingesting this data, and what has not built yet is a reliable way to tell authentic evidence from synthetic.

So I think that takes us to a closing question, and I think it's a really hard one. If the rules are three to five years behind, what are litigators to do right now?

So, Brandon, I think it goes back to John Benson's three questions. One, relevance. Two, reliability. Three, prejudice. Reliability is the one that now carries the whole load, so you build the record for it early. You preserve the metadata as a core part of the case from day one. You bring in your forensic examiner, your Steve Davises of the world, before there is even a fight, and then you stop assuming authenticity. And Judge Kim said the era where a document is produced and nobody questions it is ending. So the teams that internalize all of that now, relevance, reliability, and prejudice, treating every key exhibit as something that has to be proven, are the ones I think will not get caught flat-footed. The rules will probably eventually catch up.

There's a lot of news right now about various different evidence rules and whether or not they're actually going to pass muster with the committees or not pass muster with the committees. And so the idea is that your case will not wait for them. So it's better to prep early and to have those three things in place: the relevance, reliability, and the prejudice. Okay. Final topic, Brandon. Topic five, the human verdict. So we're going to be talking about billing, teaching, access to justice. This is probably my favorite topic here, Brandon. So step back from the docket. If AI does the finding, the sorting, the first draft, what is left that is authentically ours and human? What is the human for? What is the value of our JDs having the esquire after our name?

And three guests came at that from three different directions. And Brandon, I mean, we have had some really great guests on NeLI Pod first half of 2026, but this was so interesting. We had a law professor. We had a law school dean, different law schools, by the way, and a chief justice that turned CEO. So Jonah Perlin from Georgetown started with the thing every firm fears discussing, which is the billable hour, and he refused to declare it dead.

He did.

Even though a lot of people are thinking that the AI revolution will kill the billable hour, he doesn't believe that's true. And the most provocative point that he had in our discussion about that specifically was one that should scare firms a bit more than the speed or cost savings do. In Professor Perlin's words, he said, "There's a world in which AI makes us better, but not quite as efficient. I think it's much scarier from the perspective of an outside counsel practicing lawyer that clients will just start saying, 'Well, actually, what you do is not that valuable anymore. '" Perlin's deeper worry is that clients decide the work itself was never worth what they were paying for it. And Robert Keeling framed the same question pressure, the same pressure from the individual lawyer side with a line that I think we've heard repeated in sort of different ways throughout the season so far.

And his quote was that, "It's not that Gen AI is going to replace lawyers, but lawyers who use Gen AI are going to replace lawyers who don't. " I think that I printed that and I'm going to post it on my wall in my office. It is something I think that everybody should print out and put on their desktop, because that is, I think, the quote, maybe the quote of the season, maybe outside of Dr. Jeremy Pickens as well. But Dean Lou Mulligan at UMKC, University of Missouri, Kansas City Law School took the same technology and pointed it at a different target entirely: access. And he gave us the number that should reframe this entire conversation. What he said was is that 80 to 90 percent of legal need in the United States is unmet because folks can't afford a lawyer.

And Brandon, that really stuck with me, and I think that completely changed my perspective on how I look at access to all these various different tools. And then he ended up saying, and I have not been able to stop thinking about his words here. He said, "For me, it's really important to not only think about AI as an efficiency tool that makes things faster, but as a market expansion tool. " I mean, that was powerful.

It was. It was definitely a reframe. And a reframe that I think in the world of eDiscovery, we don't think about that very often. And when we went deeper with him on this specific topic, when we started discussing what should stay human about this legal process, he was really specific. When the machine can hand you cases, what the lawyer brings is the part that the machine can't bring. And in his words, that is, "What my counsel brings is discernment, judgment, professionalism, compassion, and empathy.

" And then Bridget McCormack. Bridget McCormack is the former chief justice of the Michigan Supreme Court, now running the American Arbitration Association. And if you haven't listened to that episode, please listen to it twice. She gave us, I think, the most bracing answer of the entire season. And she said that she's tired of the hallucination panic, and she has a point about who exactly we are comparing this machine to. She said, I love this. She said, "If one more person talks to me about hallucinations, and I have to remind them about human hallucinations, including human judicial hallucinations, I'm going to lose my mind. We have an entire appellate system built to respond to some of the hallucinations that judges make. " And further to that, her north star, the thing that organized her entire court modernization effort, was a sentence worth pausing on, on who the system is for.

What she said was, and quote, "The courts are for the people. They're not for the judges. The people, for the most part, don't even really want the judges. They want to solve their problems. " I mean, just let that sit there and permeate for a few minutes, right?

I mean, that's the mission statement right there. It is. And I think that's the range. So Perlin worried the market decides lawyers are worth less. Mulligan says AI could open the courthouse to the 80 percent who are priced out. And McCormack comes in and says, "Stop measuring the machine against perfection and start measuring it against us." Brandon, the case law we have covered today, I think, connects to this directly. It really does, but it's kind of easy to miss. When you look at who the parties were in the decisions we cited, Warner in Michigan, a pro se plaintiff. Morgan in Colorado, another pro se plaintiff, who the court described as simultaneously the party and the advocate. These are exactly Mulligan's 80 percent who didn't have access before.

The people in court without a lawyer, now using AI to do the work a lawyer used to do. And McCormack pointed us to the data behind it. She cited a study of federal court filings finding that after roughly 20 flat years, filings jumped in 2025 from represented and unrepresented parties both. And the striking finding was that the cases dismissed for failing to state a claim stayed flat. So that surge reflected substantive demand. Cases that needed to be brought, these were claims people could finally file. And that's the access story showing up in the docket. So with all of that, this is the last question of the episode. Five topics, ten guests, a stack of new opinions. What is the one thing that you want a judge or a partner to carry out at this discussion?

That's a good question, Brandon. And I think that every hard question this season had the same shape underneath it. Hepner turned on whether a target of an investigation kept the competence. Jeffries turned on whether you can get your data back. The deepfake problem comes down to foundation. And the human question, which is probably the most important, the human question comes down to judgment. Whether the person using the tool supervises the output, explains the process in plain English, and keeps sight of who the system serves. And not surprisingly, it almost seemed like there was a method to our madness with all of our guests that we had here on NeLI Pod so far. But our guests told us that from ten different seats.

I mean, that's so interesting. And the courts are now saying it in ten different opinions, right? The rules have held, and the question is whether we do.

Yeah, that feels like a good place to end the conversation. We want to thank, again, all of our ten season two guests so far. And thank you for spending your time with us and with NeLI Pod.