Episode 023 Deepfakes and the Liar's Dividend

Welcome to AI Tools for Practicing Lawyers. Practical, no-nonsense guidance to help attorneys put AI power to work in their practice right now. We know AI has made it dramatically easier to create convincing fake images, audio, and video. No one disputes that.

But forgery isn't new, altered photographs aren't new, fabricated evidence isn't new. Litigans have been trying to fool courts for centuries. So I was really wondering, especially in light of the very prominent guest that we have today, what is it about AI that changes the legal analysis? Has AI fundamentally changed the law of evidence, or has it just made it easier for us to revisit principles that have always been there?

I'm Ron Drescher. I'm Heather Gardner. Our guest today is Judge Paul Grimm. Judge Grimm served for more than 25 years on the federal bench, including as the United States District Judge for the District of Maryland, where he became widely recognized for his work on evidence, civil procedure, and electronically stored information.

He is now the director of the Bulch Judicial Institute at Duke Law School, where he continues to write and teach on the intersection of law, technology, and the courts. Judge Grimm, it's very exciting for us to have you here today. So I still work with them on programs, but I'm no longer the day-to-day director. I'm the emeritus director, but uh I it was a wonderful organization to be part of.

And um the my successor is a fabulous retired district judge as well. So I'm still working in these vineyards, um, but uh with a little bit of a different name tag. All right. Well, thank you for that clarification.

I'm well familiar with the concept of emerettis as a retired bankruptcy lawyer. Uh but but Judge, you published last year an article with uh Maura Grossman. With Maura Grossman, uh a law school professor. Judicial approaches to acknowledging and unacknowledged AI-generated evidence.

What were you trying to explore in that? Sure. So just to go back a little bit for context, uh, I've been fascinated by the challenges that digital information presents from an evidentry point of view uh for quite some time. I happened to have the good fortune to join the federal court back in 1997, right when digital information explosion had occurred, and we were now starting to recognize that the word document, which had always been, you know, paper documents when I was a practicing lawyer, now included digital information.

And with the expansion of um of digital media and devices to access digital media, this just became a very big challenge on the discovery front. And uh after a while, I began getting concerned that the terabytes and ectobytes and and every other kind of bytes, huge amounts of electronic data were being exchanged in discovery far more than would ever be introduced into evidence. And so I started focusing in 2007 on how what are the evidentiary principles that are important in getting digital media embedded into evidence, audio, visual, text.

And and you you raised, start by a good point in that the rules of evidence in in the United States and probably elsewhere, they're really technology agnostic. Uh and and there's a there's a reason for that. First of all, all the rules of evidence in the federal system have to go through a very lengthy and public process through the uh statute known as the Rules Enabling Act. And so the Evidence Rules Advisory Committee can propose a new rule and then it goes out for public comment and then they will review it.

Sometimes they will have conferences to talk about the subject matter of it. They'll explore responses to people who say yes, no, modify. From there, if it's approved, then it goes to the Standing Committee for Committee for Rules of Practice and Procedure, sort of an umbrella procedural committee within the federal courts. If that gets the green light, then it goes to the Judicial Conference, which is sort of the uh Board of Governors of the Federal Judiciary.

From there, it goes to the United States Supreme Court. Uh, and then if they approve it, then by May, they must submit it to the uh Congress. And if Congress does not act to block it or modify it, it will take legal effect on December 1st after the of the same year that it was given to them by the Supreme Court. It takes three to five, sometimes longer, six years, to come up with a new rule of evidence.

And particularly now, but always it's been the case for at least our involvement in the legal community, technology changes quickly. And now AI technology changes instantly. All know that that the that the real focus now on artificial intelligence evidence, and particularly in the context of fakes or nonfakes, this is a product of generative AI. And I remember distinctly, because I've been writing articles with Professor Grossman since 2019, and it wasn't until 2023 that I had never heard of generative AI.

That was kind of like the turning point. Trevor Burrus, Jr. It was absolutely a turning point. Because now what you have is you have and and and and that was ChatGPT-3.

And so now we're we've probably gone beyond Chat GPT-5. We have we have Claude, you know, we have Gemini, we have so many of these different platforms which are there. And what it's done is it's made available for the public at large at zero cost or minimal cost, access to these very powerful generative AI tools that, with just a very small amount of training information, such as a 90-second clip, uh, if this was uh posted on the internet, uh 90 seconds of me talking and my image waving my hand is all you need to then go into one of these programs and type out what you want it to do, and it will create images of me and audio of me that is so realistic that it it was that that increasingly it's difficult for experts to detect the difference between real and fake.

Now, there's a a a uh uh internationally respected scientist by the name of Hani Farid. He was at the University of California University of California at Berkeley in their computer science or mathematics division. He's recently gone to Dartmouth, and and he was the go-to person, you know, when somebody says, hey, there's a fake, there's a there's a video being posted showing, you know, a missile attack on a hospital in Iran, uh, and you see, you know, uh very visceral images of injured people, you know, children, adults.

We think it's fake, is it real or is it fake? And he would be the one that could look at it and and using tools of analysis frame by frame by frame would be able to say, you know, this is fake. You know, you could tell the shadows don't line up, the you know, there's six fingers on a hand, the you know, the the film, the the the shadows are different, the the perspective of the of the rug pattern doesn't match, and say it's fake. Now the technology's gotten Sadagone good that he he agonizes over trying and you know to try and determine real from fake and and cannot do better sometimes than saying it has indicia of genuineness or indicia of fakery, but it can't say definitively.

So we've gotten technology that's become amazingly sophisticated, hard to differentiate between real and unreal, and it's accessible to anyone in the world at little cost or no cost at all. And so while we've always had fraud, you know, to fake my signature required a little bit of talent, you know, to do a fake Mona Lisa, you know, he had to have some artistic skill. But you don't need to have any of that skill. You just need to have one of these programs and and a little bit of uh data to feed with it, and it can produce these amazingly impactful uh you know audio, visual, and it can produce music, it can produce code.

So what's happened is we've democratized fraud. We've made it available at an enormous scale. And and so as a result of that, what are the tools that we need to have to be able to deal with this? Do we need rules of evidence that are bespoke for AI, knowing that by the time we would get them through the process?

I mean, the the Evidence Rules Advisory Committee has for since 2023, Professor Grossman and I went down and we were we were um given the wonderful opportunity to speak with them, and we were urging the consideration of a a couple of rules, rules that might deal with acknowledged AI and unacknowledged AI, which I'll get to in just a second. And you know, they were they were interested in this, but they didn't think it required any action. Fast forward a couple more years, we did some more uh publications.

We actually drafted some proposed rules. One that would be an acceptable way that the rules could acknowledge that when you're offering evidence that you acknowledge is generated by AI, you could do this. And it would be a sort of a sub-rule of an existing rule, which is Rule 901B9, that says that you can authenticate non-testimonial evidence by showing it was the result of a system or process that produced reliable or accurate uh results. And it was a way that you could do that by you know giving notice and saying that you, here's the system, identifying the system, here's how it was tested, here's you know the data that it was tested on, and here's why we can show more likely than not that the result it produced here in this particular case is sufficiently reliable for the jury to see it.

So that was an example of a rule for acknowledged AI. But then you've got to deal with what happens if, Ron, you you wake up tomorrow and there seems to be this awful threatening voicemail on your phone from me. You know, you and I did an in-live person um program we were on earlier this year. We talked in telephone calls preparing for that a couple of times.

We spoke in breaks during that program. You and I have communicated by uh by phone and uh and and now uh digitally. So under the existing rules of evidence, if if that seemed to be my voice that you're familiar with by virtue of your prior interactions with me, rule 901B5 says that opinion as the voice is enough to authenticate it. That's grim.

So you say, I'm shocked. You know, we just did a program yesterday, everything seemed fine. I got this horrific voicemail from him. He's threatening to, you know, to, you know, he's gonna, he's gonna capture my pet turtle out of my backyard.

And if I'll I'll never see it again unless I send money to him, and and this is very upsetting to me. And and that would be enough to get it authenticated. Now I come in and say, this is this is not true. I didn't do that.

I was, I would never have done that. And now you see me and you hear me in the same voice that you saw in the in the voicemail, heard in the voicemail, and it seems to be me. How do you deal with that? And the issue is it's going to deal down with authentication, and you only need to authenticate evidence, meaning if it's non-testimonial evidence, not through a live witness, it has to be by a preponderance of evidence, which is only 51%, more likely so than not so.

And your testimony, that that's Grim's voice. I know it. That would be enough. And then what do I do?

How do I disprove that when it may be very, very difficult for me to prove that it's a fake? And so that is why we said that we urged the Evidence Rules Advisory Committee to consider a bespoke rule for what we would call unacknowledged AI. When you, as a proponent witness, say, I just I got a I got a voicemail from Grim. I know his voice.

It's right here on it's not AI, it's just his voice. It's just a recording of his voice on my device. And I come in and say, no, it's it's fabricated, it's a generative AI. It's not relevant if you go to court and you have sued me or I've been charged with communicating this threat to you.

It's not relevant to prove my guilt if it's not me. So it's key evidence. It's gonna have a great deal of impact because the jury will actually hear it and see it, potentially if it's allowed in. So how do you deal with this when one side says it's just a voicemail, it's not AI, it's just voicemail.

The other side says it's a fake. I didn't do that. And the the proposed rule that we came up with was called 901C, so it would be a new subsection of 901, which is the general authentication rule. And and the way we had originally proposed it, and the evidence rules committee sort of took that as a starting point and revised the structure a bit and did their own version of a 901c that would that would say that would work like this.

The proponent, which would be in my hypothetical U, you can come in and say, you know, this is a voicemail. I know Grim. I listened to it, that's his voice, that's my, I've I've I've heard that voice many times. There's no doubt in my mind.

You get to authenticate it however you want. Then for me to raise an issue for the judge to actually have to resolve, if my work is coming in saying it's fake, I can't just come in and say it's fake. It's easy for people to make a deep fake. How do we know?

How do we this? How do we that? Could it be this? Could it be that?

That's speculation. I would have to actually produce evidence that would be sufficient for the judge to conclude that a jury could perhaps agree with you that that's my voice, but also could reasonably conclude that it's fake. And if I did that, then you would have the opportunity to augment your proof if you wanted to, perhaps with corroborating evidence. And then the judge would decide whether or not, more likely than not, the evidence was authentic or not.

If the judge said it's equally likely that it's authentic and not authentic, the judge would not let the jury hear it. And this would be a rule that would at least give a framework for a lawyer who wants to challenge as a fake, this unacknowledged AI, a ability to do it, allow the freedom of the person offering the evidence to put whatever way they want it to authenticate. Because, as you know, there's you know, there are 10 ways of authenticating it under Rule 901, and there are 14 ways of authenticating under Rule 902, and they're examples.

They're not a universe. So it gives the proponent the agency to offer authenticating evidence however they want, requires the opposing party to come forward with evidence, not argument or conclusions that would allow a reasonable person to conclude that it was fake. The judge gets to hear it, and if it's real close on a critical issue, going to guilt or innocence, the judge would have the ability to say either that it's too close, I can't really determine, I'm not gonna let the jury hear it, or uh if you offered further evidence to uh corroborate it, then might allow it in.

And that was a proposed rule. The Evidence Rules Committee considered that this past spring and decided they did not want to go out and publish it. They did come up with a different rule, which was proposed Rule 707. Again, they have not made the decision as to whether they're going to publish that either.

They'll decide that later this year or when they have another conference to talk about it and then decide whether they need to put it out for public comment. But this has been a this has been a three-year process where we still haven't even put a rule out for public comment. Well, Judge, do you need a rule? I mean, the the situation you talked about, which was really interesting, it's not something that I had considered.

I'd I had considered a situation where a litigant or an attorney who's present in court has forwarded some evidence that maybe they created that's maybe demonstrative evidence. But and so the litigant or the lawyer is before the court. But what you're talking about is a situation where some person who is not even before the court now has created that falsehood of your voice, and now that's coming into court, and no one in the courtroom can even say one way or the other if it's AI or if it's real.

But the we've had the opportunity to create phony recordings. We've had Photoshop, we've had computer-generated imagery that could pretend to be something it's not. Is this really an AI question, or does it go beyond that? Well, you're absolutely right, and you raise a good point.

In the absence of a rule, because we don't have a rule, we have to use the existing tools that we have. And the two tools that we would apply in a circumstance like this is as a proponent, you have to authenticate it. That's Rule 901B1, and it says you have to show more likely than not it is what you say it is. You could do that just with your opinion as to my voice.

Here's the problem: it's a very low barrier. It's only 51%. Now, what happens when you get this a video showing a graphic, horrific act, maybe a violent assault, or uh or a horribly injured person as a result of an attempt to rob them, or some very, very, very, very disturbing image, which would be the key issue in the case, right? It's always going to be the most important evidence in the case.

And now you have a tool to say, well, I was there, I looked at this, this seems like what I saw, that's enough to authenticate it. What does the person who wants to oppose it do when some of these fakes can be so well done? And that then says the second tool that you have to look at is prejudice. And there's an evidence rule 1403 that says that the judge can prevent the jury from hearing evidence if its probative value is substantially outweighed by the danger of unfair prejudice.

But that's a rule that favors admissibility, because the probative value, if it has proof value, and it has no probative value if it's not authentic. But if you've shown at least more likely than not, 51%, just a little bit more than a coin toss, likely that it's authentic, then you've established authenticity. And then to knock that out, because the evidence is very relevant, you have to have substantially outweighed by the danger of unfair prejudice. That puts a real true evidentiary burden on the opposing side to do something that may not be easily done.

And here's the problem. If you have a close case where one side says it's fake, the other side says, you know what, it's not fake, and it's a close issue, then it's not relevant unless it is determined by the person with the authority to determine it, that it was in the hypothetical I've been using me and not someone making my a fake of me through generative AI. And so under those circumstances, if you say it's me and you know my voice, and I say it's not me, let's say, for example, that at the time the voicemail was left, I was on a flight and there was no, there was no Wi-Fi, and I was flying with with a business associate, and they could confirm what I said.

And I come in and say I could not have made that video and or that audio and left it on his voicemail because I was unable to do that, and I could prove it with this witness. When you have that, and the possibility is it could be fake, it could be real, the judge doesn't get to decide if the jury hears it. The jury has to resolve the disputed fact. And the judge would have to say, all right, you've heard this.

Grim says he didn't do it, and here's the evidence that says he didn't do it. But, you know, Ron says he did, and here's the evidence that supports that. If you agree with Ron, then you can consider this evidence and give it the way you want. But if you agree with Grimm, you've got to disregard it.

But here's the problem they've seen it. And there's a lot of psychological tests that show that when you believe that you see something and heard something as it was created, that you can't just disregard it. And judges are famous for saying to a jury, ladies and gentlemen, the jury disregard that, give it no weight. And that's great as a legal fiction, but it's not so good as a matter of psychology.

And there's pretty good science out there saying that once you've seen something like that, it's now Altered the way you look at this evidence and this case, and you can't unring that bell. That's the worry that we have. You know, I've been following pretty closely the judicial efforts to wrestle with the problems that AI presents. Of course, we see it more often in the situation of hallucinated cases than the evidentiary deep fakes that you're talking about.

But really, really, can the judiciary, by rule making, address these very difficult issues? Aaron Powell What the rule, what the what the what the judiciary can do with appropriate rules is provide a clear test for what must be done when you have this situation. Allocate responsibilities between the person who is asking for it to come in, the person who's objecting to it, and under the 901c, the one that Professor Grossman and I originally drafted that the Rules Committee sort of followed that approach but but revised the language, what they did that was significant was they said that when you have that basic threshold shown by the objecting party, you got to have some evidence.

Maybe it's someone like you, an expert, or someone like Heather, an expert, maybe it's a corroborating witness, whatever it is, that is enough to raise a real issue as to whether that evidence is fake or whether it's real, then when that happens, the judge gets to be the one to decide before the jury has to hear that. And if the judge thinks that there's a likelihood that it could be, it's not authentic, then it's going to not be, it's going to be more prejudicial than relevant, and the judge is going to keep it out.

So it provides a system for doing that using existing concepts that the law is already familiar with. Who goes first, who goes second, who decides whether the jury hears it, it modifies that in a way that allows us to have a greater assurance that the evidence that's going to be given to the jury that may have enormous impact is legitimate enough for them to reasonably consider it. Aaron Powell Now in your article, you talk about a great concept called the liar's dividends. What's that all about?

Well, you know, you hear about deep fakes all the time. And and uh you hear about fake news, you you hear any time someone doesn't like something, they say, well, that's a fake. There have been some famous cases where celebrities have, you know, been on a on a podcast or on a program, and then later in litigation they uh deny that they said something and and the opposing side has it and wants to put it on there, and then they say, well, that's fake. Yeah, how do we know it's not real?

So you've got you know, you've got uh a lot of public awareness that this type of evidence can be fabricated. Articles on it, you've got people talking about it, you've got uh headlines about it, and what happens is that it starts to cause and you can have it in other contexts, by the way. So somebody comes out and and there's a a close political election and and someone puts a video of one of the candidates saying or doing something that might uh alienate the jury or the the voting public, and they come out and say, that's fake, that's fake.

I mean, these kinds of things happen all the time. So the public consciousness of the possibility that convincing audio and visual may be fake is has become accepted within the public's consciousness. Hold on one second. So uh what that means is is that you've got sort of a heads you lose, tails you lose, because you've got the possibility that fake evidence will be introduced and accepted as true when it's not.

That undermines the truth-finding function, right? And then you have the possibility that genuine evidence, a real video, a real intercepted communication will be dismissed by the objecting party as that's a fake, and the jury will say, yeah, we're not gonna, we're not gonna pay attention to that. And so that's the liar's dividend. The person who lies and says, oh, that's a fake, everybody knows these deep fakes are out there, you know, we you can't rely upon that, then you get a jury that doesn't know what they can rely on.

And that's a real problem because whether they accept as as authentic evidence it is not, or reject as inauthentic evidence it is, that undermines the fact-finding fairness that we rely upon juries being able to give us in our system. And that's what the liar's dividend is. You know, part of what I've always thought, or at least since I've been doing this podcast, is that I don't know that rulemaking will be particularly effective in dealing with these very difficult AI issues.

But I do think that lawyers can be proactive in how they draft their proposed pretrial orders to make sure that they have an opportunity to have a court consider the demonstrative evidence that they want to introduce and give the opposing party an opportunity to oppose that evidence long before it reaches a jury, so that there's a little bit of the opportunity to really kind of think everything through before you get the pressurized environment of a trial. Are you seeing lawyers try to address the problem in that way?

Well, in our in our paper, Professor Grossman and I not only talked about the evidentiary issues and what the what the things that you should be concerned about are and talked about proposed rules, but we also talked about what should what should be the approach. And the and one thing, and and this could be done either by a judge with their own procedures they use in their cases, or by a court adopting that for all cases in that court, or by a rule that require that if you're going to offer into evidence, evidence that is the product of artificial intelligence software applications, that you give notice of that by a certain date, that you produce a copy to the opposing side so that they have a fair chance to look at it and challenge it if they want to, uh, that you have a deadline for the party to say if they're going to challenge it.

Um, you have a deadline for them to disclose the evidence that they're going to rely upon to offer it and to oppose it and set it in for a hearing sufficiently far in advance of the trial, but the judge is not trying to figure this out in the middle of a trial, which is too late for you to do that. That kind of notice, uh discovery, motion, and ruling before trial is something that can be done right now by individual judges or by courts through local rules. And we are are strong proponents of that, because that is something that you need to do.

It doesn't always happen. So, for example, in in Maryland in 2023, Maryland statutory requirement that the General Assembly adopted said that in criminal cases, if the if the uh prosecutors were going to use a lineup, a photo array that a victim said, Yeah, that's the person that committed the crime, if the if the law enforcement had prepared the victim by giving the victim images of the person who is the charged defendant that were captured from facial recognition AI technology that because there have been some famous examples of of false identifications, that the um that what had to happen was they had to give notice to the defense in advance so the defense could then bring in an expert and challenge on the basis this is called witness washing.

So the police officer doesn't even know who the suspect is until they get a facial recognition hit. So let me give you a scenario. And they have a security camera in the store. Someone breaks in, you've got the video of them breaking in.

We've all seen video that's very clear, we've seen video that's not so clear, and and they look frame by frame by frame. They get a picture of the intruder who they think looks as clear as you can get it. They run that through a database, maybe someone who's been arrested before or a database for driver's licenses, they get a hit. But you don't know how many people they said was they couldn't, they couldn't say if it was hit or not, so you don't know what the error rate was.

They just get that hit. And then without telling the victim that, with that picture, they get five other pictures for a photo array. One of them is a picture of the person that they got from the database, not from the video. And then they go and they show them like a lineup, it's called a photo lineup.

And the victim looks at it and says, Yep, that's the one, and points to the person. Well, if they don't disclose that, then they've been using photographs to do these photo arrays for decades. And if they just think it was a photo array because the police officer first became suspicious of this person through ordinary, you know, police work, like a fingerprint analysis or a witness or a license plate or a partial license plate or something like that, and then gathered information to support that they were the one, and then got the pictures and brought in the witness, and it wasn't suggestive.

That's been going on for years. But if they first became focused on this one person who has now been charged because of facial recognition, then the legislature wanted that disclosed. Even though that statute required the disclosure, it wasn't done. Prosecutor wasn't purposely trying to hide the ball, just didn't realize it.

These things happen. And case goes to trial, it's a conviction for robbery set aside with prejudice, can't be brought again because of a failure to disclose. So disclosure and production and discovery and challenge and hearing is good. And that can be required, but it's not fail-safe.

Of course, no rules are fail-safe. So there are things that can be done to deal with this. Part of it's on educating people as to how uh they should do things, but I will offer you this suggestion. Hardly a month goes by for the last three or four years, where we haven't heard a story about some lawyer, and in one or two embarrassing cases, a judge has cited in a filing information, either factual information or legal authority, that they got through an AI search, but they didn't bother to check to verify that it was a real, you know, a real source.

It was a it was a um hallucinated source from AI, which can happen. And we've been hearing about lawyers being sanctioned for this for years, and yet it still happens. So we know about this. There's a quick fix.

It's like, look, I use this AI, I got it, I got the case of the case, I went and found the case, I read the case. I personally compared the case, you know, that I know was an authentic version of the case from the from the court reporters, from the reporters, and it says what I said it said. And I've, you know, I have kept my my uh duty of candor to the court. It's a no-brainer.

No one should be filing any authority in a court case that they haven't personally verified the accuracy of it. And yet, despite all the sanctions and all the news stories and some cases being tossed out and people being sanctioned, it's still happening. And so we are dealing in a situation where the the ease of using these techniques, the the I don't want to get too too too uh, you know, you know, I don't want to, I don't want to be overly dramatic, but it's the these tools are so seductive.

I mean, I don't think it's a day that goes by when I don't reach out to my good friend Genesis and say, hey, you know, uh, you know, I'm looking for someone who can, you know, help me with some repairs on this. Uh do you have any suggestions in my community that I can look? And I'm gonna, I'm gonna check them out. But I, you know, there's no yellow pages to look at anymore.

I don't know. And I can I can say, show your sources, and they'll tell me what they looked at and I can click on it. It's so amazingly useful. There's the cat's out of the bag.

And so people are still, despite fair warning, that at least in the context of a case, you don't cite it if you haven't verified it, and yet they still do. So no matter what you do, there's gonna be instances where people with this technology where they're they're not complying either with the procedure or the protocol or the rule, and and that can create problems. I want to ask you uh a different evidentiary question using AI. And this has to do a little bit, I guess, with machine learning.

Let's say you're you've got a situation where in Discovery, a series of photographs, uh microfiche of canceled checks are delivered and they're hard to read. One of the parties wants to use AI or machine learning to make it clearer, easier to discern. All right. Is that ever going to work?

Or are there protocols that that litigant could use to optimize their chances of getting that kind of evidence uh admitted? All right. Well, I I want I for your reader, for your uh for your audience, um, this is not a topic that we've discussed before. I did not put you up to doing this, but there's a specific example of that of an actual case.

In the state, in the case of state of Washington versus Pulka, uh, which was decided in 2024 as a state court case in uh King County Superior Court in Washington State, it was a criminal case. And the defendant was charged with a crime, and the and there was a factual issue about whether a cell phone video taken from a bystander that was grainy and dark, and and you we've all seen those videos where they're not Cecil B. DeMille quality. And uh, and and the it's clear that the defendant's got something in their hand that they're that they're holding, you know, and they're shaking it.

The question was, is it a weapon or was it a cell phone? The defense offered as an expert a film producer, not a Hollywood film producer, but sort of a local person that makes films for documentaries or will go to an event and make a film, like you go to the wedding or whatever. And and he came in and they and the defense offered it. There was notice of it, but so that's good, offered this in and said, you know, you can't really tell from this bystanders video that the police got when they investigated.

It doesn't really show what's in the hand. But if it's not a weapon, then the aggravated assault, assume for a moment that was a charge, couldn't be proved because it was a cell phone, not a weapon. And what we did was this expert used AI technology, and what it did was it cleaned it up, it sharpened the focus, it got rid of noise, it it got rid of, you know, it brought some light in there so that that you know what the lens couldn't see was clearer, didn't change what was there.

It just it just sharpened it up. The prosecution said, absolutely not. And the problem was is there was an evidentiary challenge, and the the witness, you know, testified, well, uh I don't know how this AI works. It's uh I use it in my films.

I make a film and I look at the film and I say, you know what, that's that's an important scene. It just didn't come out the way that I wanted. So I'm gonna run that section through this stuff, and all of a sudden, wow, it's real clear. It's what I meant it's be.

So I think that that's what was really happening, but it's really sharp. I don't know how it operates. I just use it. Well, the the prosecution came in with witnesses, and this happens.

I mean, I can remember when I was still on the court, you'd get these surveillance recordings, and there's, you know, there's static or an airplane goes by or a dog is barking in the next room, or or highway sound is there. And it's particularly for sound, you know, there are sound waves, and different kinds of sounds are at different waves. And the prosecutor brought a couple of forensic experts in who said there's a there's a legitimate area where you can clean up how accurate this is without changing the underlying event to make it look like something didn't happen.

That is not what it was, so the judge excluded it. So the point is that there are there are going to be examples where AI tools can be used to clarify something, which is an important point in the case where there's some doubt as to what the visual or the audio is. But it's gonna be important to make sure that it what you see is actually clarifying and not altering the underlying nature of what it is. And so that's gonna be the kind of thing that we're gonna want to be paying attention to.

And I think that if you have those either rules or procedures that require notice and disclosure and discovery, then the system has the tools available so that no one's surprised, no one's bushwhacked, and the judge can then rule, hopefully, before trial. So we're actually starting to see examples where some of this stuff comes in. There's another one, and I'm gonna say something too. There's an existing federal rule right now, and this is flying under the radar scope.

This is a fairly new rule. It's rule of evidence 107, just was adopted in the last couple of years, and it's called illustrative evidence. And it says, look, you know, illustrative evidence can be as simple as a drawing on butcher paper or you know, a chart that was done either by by hand or it could be generative AI computer simulations, right? And what this new rule says is that is that these illustrative aids are not themselves evidence.

It's not a visual of what actually happened. But if it's if it's helpful to the jury, for the jury to understand evidence that's already been brought into court or to understand legal argument by the lawyers, then the judge, as long as the judge determines that the probative value of this aid is not outweighed by the danger of unfair prejudice, can allow it to be seen by the jury. So imagine this. You say, I'm gonna prove X, Y, and Z, and here's an illustration of what the evidence will show, and you see this really sharp generative AI projection of what the person says, I'm gonna prove that with other evidence, witnesses and fingerprints and all these other things.

Well, Rule 107B says that the jury can see this, but they can't have it when they deliberate, which is when they look at all the evidence together and actually rule on the case, unless all parties agree, or the judge finds that there's good cause for doing that, and limits the jury to say, ladies and gentlemen, I let you see these generative AI, you know, uh demonstrations, but these are not the evidence. They're just trying to help you understand what the witnesses testified to and what the other evidence was that you've seen.

Well, here's the problem. When you have these generative AI tools, you can have tools that can do amazing things. And I have another case for you. And this is a case this is a case, state of Florida versus Miguel ALBISU from Broward County 2025.

And in that case, what happened was you know, Florida has a stand your ground law so that if you have you're allowed to use deadly force, if you if on an objective, reasonable person standard, you had an objective, reasonable belief that you were in in danger of severe bodily harm. And the case involved the guy that owned uh an event venue uh in uh in a wedding reception, a post-wedding reception was being held there. He's there because it's his you know his bar, his venue. Things were getting a little bit uh boisterous.

You know, I'm I'm told that at weddings, alcohol sometimes flows, and and everybody was in a celebratory mood and partaking. And what the guy was charged with was discharging a firearm. Um no one was shot or hurt, but he was discharging a firearm. And the state said that he violated the law by doing that and and you know was threatening people.

His defense was look, it was dark. These four or five guys who were in the wedding party, big bruising guys, you know, they're they're liquored up and they come over there, and I was saying, hey, you know, you're you're you're getting a little bit boisterous here, you're gonna start to damage some of my stuff. I want you to stop it. They didn't like that.

They approached me and I feared for my life, so I discharged the handgun just to make him go back. And now the guy could testify to that, right? He he had the ability to testify and get on the stand and call witnesses. But here's what his lawyer did he said, Judge, you're not gonna really understand.

And the jury won't really understand how it appeared to my client just by seeing still shots of the place and talking to the witnesses and all that. But if you put these virtual reality goggles on, we have prepared an AI-generated illustration of what it looked like from my client. So the judge puts on the glasses, and so does the so do some of the other witnesses, and it purports to illustrate the underlying testimony about what the person saw. And it's like you're seeing it have to happen.

Now that was done. The judge agreed to do that. It's not an appeal. So you this is an interesting thing where you're not trying to actually even get the AI into evidence, but you're using the AI to illustrate other evidence.

But think of the impact of that. Think of the impact that has when the jury says, I, I, you know, I'm not walking a mile through that person's shoes. I'm actually seeing what they saw that could have tremendous impact, and it's not even evidence. So we, and that's a rule that opens the door for that.

And I don't know any cases where that issue has come up, but it's only a matter of time. So, you know, these are these are things that we're dealing with now in real time. This evidence is going to be used. So your answer is yes, lawyers can, there are tools that lawyers can use, generative AI in in creative ways to help cases that are going to be decided in court.

But each one of these comes with some special challenges that you got to keep in mind and and some protections that you got to put in place to make sure that you don't have unfair prejudice to the party against whom this evidence is going to be offered. One quick question. If existing AI, if existing evidence rules are inadequate for AI, basically, that's kind of what we're talking about. And they declined to put the new rules in place, for instance, that you proposed, or even the ones that they modified.

But we're discussing deepfakes. Deepfakes goes back over a decade. And so the the idea of the law outpacing the time seems, you know, silly to me at this point. Why has this not been addressed in the last dozen years?

I think that the difference is this. We we as Ron said at the very beginning, Heather, we know that there's fake stuff out there, but but it was, you know, it took some skill and some effort to make convincing fakes. And now and and you know, you weren't seeing it. I yeah, I I would have an occasional fakery issue come up in a case, but it wasn't present in every case.

Now what you've got is you have the ability for anybody to make this stuff and to offer it. And um and the stuff is so good now that the tells, you know, that you used to be able to have to say, I'm not sure about that, they're not sufficient to test to tell the difference between the two of them. And as a result of that, we've got very powerful technology that's only getting more powerful. We've got we've got tools that are available to everybody that that are going to be you know be so enticing that that people are going to do it.

There are already some cases where some lawyers have been sanctioned. Uh there was one case last year, I think it was, in Superior Court of California, in Southern California, where a witness, uh self-represented party in a civil case, completely fabricated images in a responding to a motion for summary judgment. And the judge ended up finding, in a very carefully worded opinion, that the judge said, this is why I find that this is fake, ended up sanctioning that woman and throwing her court out of case, case out of court.

So the problem is that you know, you don't need to have a computer science PhD in your in your uh in your back pocket to produce this stuff. You can do it all on your own. And uh and and the temptation is there. And, you know, when the lawyers who are the ones who should be most vigilant about this, continue to get sanctioned by courts for not doing basic things like checking cases that you can check.

I mean, there's a real case or not a real case. It's not, you know, it's not like there's some, you know, what was I supposed to do, Judge? Read the case, and yeah, and then not just take it because, you know, your buddy, your AI tool told you told you that's what it was, your AI companion. And so I think that the problem is that, yes, we've had deep fakes for I think that the the deep fake first got into the lexicon in 2017, which you're right, heather's like nine years ago, by the guy who was a Reddit user and his nickname was his his handle or whatever you call it when you go online was deepfake.

And what he was doing was very crude, but he was taking actors who were well known, you know, mostly women, and he would put their face on video of porn stars. So it looked like whoever was doing whatever this pornographic movie was doing was actually a famous actor or actress. And then, of course, they were putting it out on Reddit, and all the other people who spend all their time in their basement, you know, eating Cheetos all came out and said, Well, we can sharpen it up by doing this and that.

Then it started being used for revenge porn, then it started being using for these other things. And now all of a sudden, if you you know you can go to the Oxford Dictionary for English language, and I'll bet if you look for deepfake, it's in there because it's now become pervasive. Probably. Yeah.

So all right. I I I think that that we've always had to deal with law being a lagging system that lags behind technology. But with this particular type of powerful persuasive technology, the problem is that it's developing so quickly that it seems to be outpacing. I mean, I I heard last year that there were predictions that this year, 2026, and of course we've got a few months left, that that they were going to get to general artificial intelligence that would exceed the ability of humans to be able to perform certain tasks.

And uh I don't know that anyone has claimed the ability to do that, but when Mora Grossman and I were first looking at this stuff, we were saying, ah, you know, that's not likely to be a case. Computer scientists were skeptical. They're not so skeptical anymore. Not so skeptical.

Judge, turning away from evidence and and cases, I have a kind of a hypothetical question to ask you. If someone gave you an unlimited engineering budget and said, Judge Grimm, build the perfect AI tool for judges, mediators, and neutral, what would that do? Well, I can tell you right now that that there is billions of dollars of investment money into these technologies. And there are various people who are starting to develop these.

Professor Grossman, my my friend and colleague of many years, I think that she at her university, which is Waterloo University in Canada, that's Canada's MIT, um, they got a grant to try to come up with a deep fake detector. So they they are, you know, their goal, if they can, is to come up with a with technology that would nothing's going to be perfect, right? That everything's got an error, right, right? But that would be very, very good at being able to, through automated artificial intelligence analysis that had been tested on data and had, you know, um results that you could say are sufficiently valid and reliable that you could say, well, I took this this uh video and I ran it through there, and it said, you know, conclusion was that it was fake or it wasn't fake.

So that's an example of something. If that comes out and that is demonstrated, and you know, then it comes out in the scientific paper, everybody looks at it, that could be a huge advance, something like that. And, you know, if if anyone could do it, you know, I I would never bet against Maura Grossman and her team. They're they're very, very, very good.

There's another company that's called Learned Hand, the famous judge, Learned Hand, and uh a young entrepreneur who I think went to, you know, went to either Yale or Harvard Law School, clerked for a appellate judge, knows what it was like to be a law clerk for a judge, and has developed a program for use in courts that would allow judges to have the benefit of the power of AI to do some things that can be absolutely transformative for what judges can do. So, for example, summary judgment.

You know, you you go through a case, you've got complicated issues, everybody spent a year and a half of discovery, you've got depositions, you've got you've got pages of pages of documents and all this, and and now the judge has got to sift through all these things and needs to get what do I have to focus on? I've got this ocean of records here. What do I focus on? Well, you can take AI and you can put a 300-page transcript in there and say, summarize in 10 pages the key issues of these.

I want the I want to know where each of these witnesses testified, and I want to know what page and what line they talked about these issues, and it goes. Now, you got that summary, and you've also got the transcript. So you or a law clerk can go in and verify that on page 34, line 12, they really did say that. That's hugely helpful.

You can outline, you can put certain things in there, like, you know, in bankruptcy court, you know, Heather, I don't know if you're a practicing attorney, but if but if you are, you've done this yourself a million times. You know, there's what I call the wind up. So somebody files a motion and says, we are gonna say that this is not sufficient cash collateral. I have no idea whether that makes any sense at all to a bankruptcy lawyer, but that's the issue.

Well, there's a standard. There's a bankruptcy rule, maybe something in the statute that says this is what it is. Somebody's got to decide what that is. And, you know, you could you could put all that stuff into this thing and say, come up with the with the strongest argument why this is cash collateral, or what's the strongest argument why it's not?

And it can go out there and if it's particularly if it's a bespoke legal AI tool, so it's not trained on the internet, but trained on real cases. You could train it on all the published bankruptcy cases ever decided. It comes out and says, well, these are the arguments, and this is the most important order and all this. So those could be powerful tools to help reduce the time and cost associated with what we do in court.

They could powerful tools that could perhaps help pro se litigants come up with pleadings that were genuine and could allow an issue to be decided on the merits, not the kind of stuff that AI that that self-represented people come up with because they don't know any better and they can't get a lawyer to help them. So there's some benefits that we can have. And right now, there's a ton of investment money going into these tools to be able to try and take that next step to make things faster, cheaper, better.

And the fact that we know this is happening is that in 2024, I think it was August of 2024, uh, remember that's just a year after generative AI broke into the public consciousness, the American Bar Association's Ethics Committee promulgated Formal Ethics Opinion 512, which is about 15 or 20 pages long, and it's ethical issues associated with the use of generative AI in the practice of law. So when the ethics people have already come out and given you 20 pages of advice as to how to ethically use this stuff, the genie's out of the bottle.

This stuff is there, it's being used, and lawyers are using it. There's an old saying that good news, bad news to lawyers, good news is you're not going to lose your job to AI. Bad news is you may lose your job to a lawyer that knows how to properly use AI. Um and it's a double-edged sword because the ethics rules require you to be proficient in technology so that if you don't check the output and verify it before you file it in court, then you have some potential ethical problems you've got to deal with.

But on the other hand, it's gonna get to the point where tools that you can do that AI can do well and better and cheaper than a human, like summarizing deposition transcripts, you know, coming up with a chronology, those kinds of things, you're not gonna be able to put a human being charging $300 an hour, spending 12 hours summarizing four depositions if you could use it for AI and it costs, you know, does it in a fraction of the time. There are tools that are going to be out there.

I suspect, uh, Ron and Heather, that that five years from now, these issues and questions are going to seem silly to people because it's like, of course, we do it. And it's just like anything else. You know, it's it's AI until it's sufficiently broadly used, and then we just call it software, right? Because, you know, things like spell check, that was AI.

But we're not all saying, oh my God, they used AI for spell check. It's like it's just software. And and so um, you know, the the the pace of the development of this stuff is going to come out with things that are being used. And the question is going to be how do we integrate this stuff in a way that allows us to harvest the promise of this faster, cheaper, fairer, but avoiding the pitfalls, particularly the pitfalls of misleading juries, because our rules of evidence, of course, in bankruptcy, you you th there are circumstances where you can have a jury trial, but it's very rare.

But but in the in the district courts and in the state courts, the the the the presumption in the rules of evidence is that juries will decide the facts. And so we we don't want lay jurors capable of being misled, confused, or deceived by technology and some of this generative AI technology, you know, that it's not transparent. They can't they can't tell how the machine got to the answer that it did. They can only say, well, here's how we trained it, and here's the data we trained it on, here's the error rate, and we produced this stuff and we tried it in some in these five cases where we knew what the answer was and it got the right answer.

So in this one thing where we don't know the right answer, we're we're betting that it got the right answer as well. Well, is that going to be sufficient? And and those are the kind of questions that we're gonna have to wrestle with. And and Heather, to your point, even if the even if the rules committee comes out and says, you know, we're we're just not, we don't feel we have the need to go out and do this now, they've been kicking it around, they've been having conferences, they've been hearing proposals for several years now.

And and the fact that they're addressing it in this fashion is good because they're getting more and more input from people who either want rules or oppose rules, and and that is influencing, okay, how big of a problem it is. Biggest problem that we had starting out is that the courts were saying, well, we don't have any examples of any problems, so why should we have a rule? We don't even know how we have problems. The answer for that is we don't know the false negatives.

We don't know the number of cases where we didn't know, we didn't have these disclosure rules, we didn't have these ability to test, so we don't know whether it got right or not, because no one even thought to to think about it. So that's the kind of a thing that we have to worry about. And uh so as we and and and some of these software folks were trying to come up, I mentioned the the learned hand, it's my understanding that there's some uh state appellate courts that are doing some pilot projects using that, and even at least one one trial court system that's using it.

So there's gonna be, I think that by the time you know five years goes by, a lot of these things will have been worked out. There still will be issues, and I'm quite confident there will be even new newer technology that will keep you guys doing podcasts for the foreseeable future uh for quite some time. Judge, before we break from this really watershed episode, I always like to ask this question of certain guests. What was the last wow moment you've had with AI?

I I guess I guess the wow moment that I had was uh when I was still had the privilege of being on the faculty at Duke and and uh I would I would very often have to be writing articles on a on a quick turnaround time or an editorial to prepare for, you know, to to moderate a a panel or or something. And you know, I I give you a concrete example. I I've been asked to give a uh a talk to um a group of lawyers next month on a particular topic. And and this is uh there have been some very prominent people invited to speak in the past on things, and I want to do a good job.

And and if I were to sit there thinking, okay, this is a complicated topic, how am I going to organize it, where I'm gonna start, what examples am I gonna give? Um, you know, I might spend I might spend, I'd do my research on my my base material, I'd look at it, I'd I'd make an outline, I'd think about it, I'd revise it, maybe I start writing and all this. And maybe I would end up spending five, six hours and I'd say, okay, this is this is a good approach. I can tweak it later, but this is a good approach.

Well, and I and the way I try to do this is I don't want my first exposure to what I'm going to do to come from AI. But when I think I've got it, you know, and I want to say, these are the four things you really need to pay attention because they're the most important. And I've got that and I've got it in my head already. Then I had put this into an AI tool and said, tell me, you know, the the most important factors that someone who wants this outcome should consider.

And and what are the tools best available to do that? And boom, boom, boom, boom, boom, boom, boom. You know, you've seen how these things are. They're just instantaneous when it comes out.

And you look at that and and maybe they came up with 10 and you had seven. Maybe one of them is not really, really there, but the the likelihood is there's one or two that you didn't think about. And they did it instantaneously. And as long as you're putting the sweat equity into it at the beginning so that you're not just relying blindly on it, it can be a very good way of checking your work.

And I've been astonished when I would, I give you an example. I've I've frequently been asked to talk about how we can improve the public's viewpoint of the court system and judges. And, you know, I've been I've been working in this vineyard for a lot of years and I can rattle off, you know, half a dozen things. But when I get sit down and get a speech and say, hey, you know, here are six things that the public can do.

Are there anything else that I've missed and comes up there and there's a nugget right there that I just never would have thought about? That was the moment when I realized that was my, you know, you said watershed. That's that's the way I would phrase it is that's my, we're not in Kansas anymore, Toto moment. I realized that, you know, this stuff could be used so well, so quickly, uh, so helpfully.

And as long as you guard against the uh you know, blind reliance, then I think that there's a lot of promise with this technology. But come on some of the problems with it is that the stuff is getting so good now that even the people who have spent the time developing it can't really tell you how it how it got where it got. And the more and if we ever do get to that magic moment where we've reached uh general artificial intelligence or artificial general intelligence, however they put it, and the machines are able to do things that humans can't, then that's going to be something that we have not had to deal with before.

And you know, that will be a an interesting time to try to figure things out. Judge Grimm, thank you so much for being here. It's been a great episode, and I thank you very much. My pleasure.

I appreciate the opportunity to visit with you. All right. We will see you all next time on AI Tools for Practicing Lawyers. That's it for today's episode of AI Tools for Practicing Lawyers.

Thanks for listening. We'll see you next time.