Latest / Logixly Live Unsupervised Learning Podcast / AI In The Witness Box

Transcript

AI In The Witness Box · Logixly Live Unsupervised Learning Podcast

• 0:00
  SPEAKER_01: Everything was fantastic.
• 0:01
  SPEAKER_00: I snuck out early because we were ta I was talking to people all day long and there's just only so much time I can talk to people.
• 0:06
  SPEAKER_00: I run out of steam.
• 0:08
  SPEAKER_01: Yeah.
• 0:08
  SPEAKER_01: Well, hopefully we don't on this show.
• 0:11
  SPEAKER_01: All right.
• 0:11
  SPEAKER_01: With that, we're gonna welcome everybody to Logic Slee Live, where we discuss the biggest stories in AI, explain what happened what is happening, why it matters, and what to watch for next.
• 0:22
  SPEAKER_01: So today we got a newcomer joined by DJ.
• 0:25
  SPEAKER_01: Welcome.
• 0:26
  SPEAKER_02: Good morning.
• 0:27
  SPEAKER_01: Good morning.
• 0:27
  SPEAKER_01: We got Bakewell back as always.
• 0:30
  SPEAKER_01: Um I don't know.
• 0:31
  SPEAKER_01: I I'm personally just this just topic area today is super fascinating to me.
• 0:37
  SPEAKER_01: Um kind of touching on like what is an expert, who is an expert, and uh how AI should and does play a role in it.
• 0:45
  SPEAKER_01: So let me just recap the event.
• 0:48
  SPEAKER_01: Uh so this is kind of an older event.
• 0:50
  SPEAKER_01: This is something that happened back in January of 2020.
• 0:55
  SPEAKER_01: There was an explosion at a facility, it says a grinding facility in Houston that uh killed three people and destroyed uh 200 homes.
• 1:06
  SPEAKER_01: Okay, so obviously a huge, huge incident, huge accident.
• 1:11
  SPEAKER_01: Investigators um determined that there was kind of a um a malfunction in a in a piece of equipment that uh that you know was related to to causing of the incident.
• 1:24
  SPEAKER_01: So it goes to trial, you know, obviously time passes, and it goes to trial now in 2026, and 3M uh manufactured some of the equipment that was malfunctioning.
• 1:36
  SPEAKER_01: And as part of the trial, 3M hired a expert witness, okay?
• 1:43
  SPEAKER_01: And you know, I find everything about this just really interesting.
• 1:48
  SPEAKER_01: So this expert witness during the discovery process, the um, so the the expert witness would be for 3M.
• 1:55
  SPEAKER_01: So for the defendant, so for the plaintiff, the plaintiff's lawyer was going through the discovery process, came across a five-page document that I think this is interesting, that seemed to be created from Chat GPT, right?
• 2:10
  SPEAKER_01: Like you can all read it, read a document and see that's that's chat GPT creation.
• 2:14
  SPEAKER_01: Yes.
• 2:15
  SPEAKER_01: So because of that, the lawyer for the plaintiff said, you know what?
• 2:19
  SPEAKER_01: Let me see.
• 2:20
  SPEAKER_01: Uh I want access to the basically the prompts, the history, kind of like the the modern version of browser history, like the chat history between the expert witness and chat GPT.
• 2:33
  SPEAKER_01: So goes through and finds uh the chat history, and uh there were 350 pages of prompt history, and a couple of things really stood out.
• 2:45
  SPEAKER_01: One was again, this is an expert witness preparing uh basically a report as well as preparing for his testimony.
• 2:52
  SPEAKER_01: But one of the prompts that I think is at uh you know, kind of in question here is the expert witness said he wanted to um he asked ChatGPT to defend 3M and demonstrate that 3M was 0% at fault.
• 3:10
  SPEAKER_01: And and that that direction, that prompt was part of the prompts that we went into creating this expert report.
• 3:18
  SPEAKER_01: And uh furthermore, on as far as the expert report goes, it was a 30-page document that the expert witness agreed with under oath that 85 to 90 percent of this um expert, you know, the the document this expert witness uh produced, 85 to 90 percent of the report that was submitted was generated by Chat GPT.
• 3:44
  SPEAKER_00: Katie, you're saying that.
• 3:45
  SPEAKER_00: Wild, wild west, man.
• 3:47
  SPEAKER_00: It's so interesting.
• 3:49
  SPEAKER_01: Yeah.
• 3:50
  SPEAKER_00: Because like I think there's there's not enough context on what is those 350 pages ahead of it, right?
• 3:56
  SPEAKER_00: Like in those prompts, yes, how much of it?
• 3:59
  SPEAKER_00: Because I think the prompt that that's in question, right, was you are an expert witness, you are defending 3M, created an exceptional expert witness report that shows 0% liable.
• 4:11
  SPEAKER_00: Right.
• 4:12
  SPEAKER_00: So much like, oh, that was kind of shady, right?
• 4:14
  SPEAKER_00: Like but is it is it that he used Chat GPT, or is it that you know it's that final report he didn't edit the report, but like how much of that fit 350 pages like to put 350 pages together of prompts, right?
• 4:32
  SPEAKER_00: He put some expertise into this.
• 4:33
  SPEAKER_01: So I I should correct 350 pages of chat GPT conversations.
• 4:38
  SPEAKER_01: So you're right.
• 4:39
  SPEAKER_01: It's actually prompt and response.
• 4:40
  SPEAKER_01: So maybe not all prompts, but still that's quite a bit of that's quite a bit of back and forth there.
• 4:47
  SPEAKER_00: 100%.
• 4:48
  SPEAKER_00: Yeah.
• 4:48
  SPEAKER_01: What about you DL?
• 4:49
  SPEAKER_02: But was he doing the legwork to prove zero percent fault?
• 4:54
  SPEAKER_02: Right.
• 4:54
  SPEAKER_02: So then we start thinking about well, what should a witness be in a in a trial, right?
• 4:59
  SPEAKER_02: Right.
• 4:59
  SPEAKER_02: They should have a little bit of neutrality, they should potentially, you know, make the case based on their opinion.
• 5:06
  SPEAKER_02: His opinion was I want to find zero percent fault, so hey, chat GPT, help me do that.
• 5:12
  SPEAKER_02: And that was potentially the premise of the whole reason why he had that prompting um conversation back and forth.
• 5:20
  SPEAKER_00: Right, but I feel like you this is DJ's dream world, and DJ's dream world is very nice where people do the right thing, but uh when you're at defense, you're hire well, defense or planes have you are hiring an expert witness uh who shares your opinion.
• 5:34
  SPEAKER_00: You're not just going maybe we should have that.
• 5:37
  SPEAKER_00: Maybe that could be our new thing.
• 5:38
  SPEAKER_00: We could get expert witnesses that are hired by the court and just a completely neutral person, but his goal wasn't to be a completely neutral person, let's be honest.
• 5:48
  SPEAKER_00: When you look at the legal system, his goal is to approve the point that the person who's paying him would like to hear.
• 5:56
  SPEAKER_02: So did he do a disservice to himself to his potential client, right?
• 6:03
  SPEAKER_02: Because obviously, you know, the way it plays out, the the prosecuting side, right?
• 6:08
  SPEAKER_02: Like they they ended up using him against the defense.
• 6:13
  SPEAKER_00: Good job, prosecutor, right?
• 6:14
  SPEAKER_01: Yeah, yes, and so that that detail is this was an expert witness hired by the defense.
• 6:20
  SPEAKER_01: Um, and it says here they paid him roughly$90,000 for his analysis at$475 an hour.
• 6:27
  SPEAKER_01: So hired by the defense, and then after discovery, once the prosecution saw the history and saw the the 0% um responsible, 0% at fault line, then the prosecution actually called them as a witness to kind of bring out that.
• 6:46
  SPEAKER_01: And I think it's kind of interesting because it's like I think there's a number of things why this is interesting, but you know, he I think he really is an expert.
• 6:54
  SPEAKER_01: I mean, based on his credentials, he is 20 years of an expert.
• 6:58
  SPEAKER_01: But at the same time, in this example, I I guess you'd have to go into the really the interactions to make a really good um opinion on this.
• 7:09
  SPEAKER_01: But he may be an expert, but at the same time, like it kind of feels like the expert opinion that was submitted was more of Chat GPT's expert opinion.
• 7:20
  SPEAKER_01: So it's like he's an expert, probably appears to be taking a little bit of a shortcut, maybe.
• 7:26
  SPEAKER_02: He laundered Chat GPT's output on his own credentials.
• 7:30
  SPEAKER_02: Yeah, I mean that's that's that's that's what he did.
• 7:33
  SPEAKER_00: Let me go.
• 7:33
  SPEAKER_00: If I'm the defense counsel, where I'm really mad here is if you're charging me by the hour, right?
• 7:40
  SPEAKER_00: I don't think you put those hours in.
• 7:42
  SPEAKER_00: I want to see your work breakdown on that bad boy.
• 7:44
  SPEAKER_00: Because I if you're using Chat GPT to do it, like I'm not sure that you would get quite that robust of a number of hours.
• 7:53
  SPEAKER_01: Yeah.
• 7:54
  SPEAKER_00: But yeah, there it's such a gray area because the other side for me, like anytime you talk about this, we've got federal law that if you put something into Chat GPT that is privileged, it is no longer privileged.
• 8:06
  SPEAKER_01: Yeah.
• 8:06
  SPEAKER_00: So did he put privileged information?
• 8:08
  SPEAKER_00: I would have those questions as the defense counsel.
• 8:11
  SPEAKER_01: Yeah.
• 8:11
  SPEAKER_00: But, you know, I think we've we've talked about this.
• 8:14
  SPEAKER_00: Like, Matt, I I would consider you an expert witness if I was talking about like, let's say, a radiology um model gets put together to identify whether or not somebody has a tumor and the model fails, somebody loses their life because the model failed, and we're gonna go to trial.
• 8:33
  SPEAKER_00: Okay, great.
• 8:34
  SPEAKER_00: I'm calling Matt Burseth as an expert witness, right?
• 8:36
  SPEAKER_00: You are you've done enough in computer vision for radiology that like you would be the expert witness.
• 8:44
  SPEAKER_00: Would you use Chat GPT or would you use a tool to help you?
• 8:50
  SPEAKER_01: 100% I would.
• 8:51
  SPEAKER_01: I mean, I you know, and in I I mean, I I actually try to think of um activities that I do where I am not using an AI tool to I I haven't brought this up before.
• 9:06
  SPEAKER_01: Like, you know, I'm trying to learn how to play golf again and I'm terrible.
• 9:10
  SPEAKER_01: But the tools where I can just record my swing and get awesome feedback right in my patio is amazing.
• 9:17
  SPEAKER_01: You know, I am terrible at cooking, terrible at cooking.
• 9:21
  SPEAKER_01: And now like I have an operational sous V and a cast iron, right?
• 9:26
  SPEAKER_01: Because I I get I get like like directions that I can follow, right?
• 9:30
  SPEAKER_01: It's so it's like what I I I a hundred percent.
• 9:33
  SPEAKER_01: If I would be an expert witness, I myself, because I like to prepare, I would be prepared, I would be asking Chat GPT, what kind of questions are am I gonna get?
• 9:43
  SPEAKER_01: Yeah.
• 9:43
  SPEAKER_01: Um I am an expert, but I would also probably say, um, what do you think about some answers?
• 9:49
  SPEAKER_01: What do you think about this?
• 9:50
  SPEAKER_01: Or if I believe, so this is where I feel like I can argue both sides of this.
• 9:54
  SPEAKER_01: If I really felt I'm an expert, and I'm I'm this the gentleman here that was the expert witness that said zero percent at fault.
• 10:02
  SPEAKER_01: If I really felt 3M was 0% at fault in my expert opinion, I would tell Chat GPT, hey, I think 3M is 0% at fault.
• 10:11
  SPEAKER_01: Help, what do you think justifies this?
• 10:13
  SPEAKER_01: And do I think those same things justify it?
• 10:16
  SPEAKER_00: But do we know that conversation didn't no idea.
• 10:19
  SPEAKER_00: We only have the report, and if you're a reporter, yeah, the juicy story is prove to me it's 0% at fault.
• 10:25
  SPEAKER_02: They quoted some of the prompts, but you know, to your point, they're not probably showing everything.
• 10:29
  SPEAKER_02: But I think that's that's where you have the human part, yeah, and then you have the assist part.
• 10:34
  SPEAKER_02: Right.
• 10:35
  SPEAKER_02: And if it's assisting in crafting a message or building a report with your data and your thoughts, yeah, much different than I want it to prove you know 0% at fault, and that's just my whole goal.
• 10:50
  SPEAKER_02: But if I have the backup, like you said, and I would put it in there, and then maybe even throw in that, okay, am I looking at this subjectively or not?
• 10:59
  SPEAKER_02: Right.
• 10:59
  SPEAKER_02: Right.
• 10:59
  SPEAKER_02: What are your thoughts?
• 11:00
  SPEAKER_02: You know, stress tested and stress tested a bit would be different.
• 11:03
  SPEAKER_00: I would think you're no less of an expert because you used a tool to prep pre prepare it.
• 11:08
  SPEAKER_00: It wouldn't make me question your expertise.
• 11:10
  SPEAKER_01: Right.
• 11:10
  SPEAKER_01: Now, what what if though, what if though, I feel like it's come we talked about this last time.
• 11:14
  SPEAKER_01: It come to me, part of it comes back to disclosure.
• 11:17
  SPEAKER_01: So it's like, okay, so let's say I give this expert testimony.
• 11:20
  SPEAKER_01: I mean, let's let's take the example, Katie, the example that you had.
• 11:23
  SPEAKER_01: Like it's me and it's a computer vision model for radiology, and this same thing happens.
• 11:29
  SPEAKER_01: Oh my gosh, they pulled Matt's chat history, watch out.
• 11:32
  SPEAKER_01: But but he says he asks for a prompt that says show that there's zero percent at fault, right?
• 11:40
  SPEAKER_01: That looks pretty bad.
• 11:41
  SPEAKER_00: It looks bad.
• 11:43
  SPEAKER_00: But is there more?
• 11:44
  SPEAKER_00: Yeah.
• 11:45
  SPEAKER_00: Is this the whole story?
• 11:46
  SPEAKER_00: Because man, if I'm a reporter, that looks terrible.
• 11:49
  SPEAKER_00: I want to talk about it because it looks awful.
• 11:51
  SPEAKER_01: But how much too of it is expectation?
• 11:54
  SPEAKER_01: Like, like, I don't know.
• 11:56
  SPEAKER_01: For me, when I think about the legal system, I think about like courtrooms that smell like you know, shoe polish and wood, and I think about you know, an uh a judge that maybe a little bit question.
• 12:08
  SPEAKER_01: You spend a lot of time there, but you know what I mean?
• 12:12
  SPEAKER_01: It's like it's like I don't think about technology.
• 12:14
  SPEAKER_00: Sure.
• 12:14
  SPEAKER_01: And if I think about an expert witness, I'm thinking about Matlock or you know, and and I just I I feel like some of it to me is almost the expectation when we talk about an expert witness, in my head, I am not thinking that expert witness was on Chat GPT all weekend preparing.
• 12:31
  SPEAKER_01: And so when I hear or read something, I'm thinking this is an authentic opinion.
• 12:36
  SPEAKER_01: And the reality is in the world we're in now, I don't think like where are legit how even if my golf game, so my golf game is gonna get fantastic, right?
• 12:45
  SPEAKER_01: Now, if you talk to me in six months and I'm you know, I'm a 10 handicap, am I gonna say, hey, look, I mean, I know I'm a 10 handicap, but I mean, come on, I would not be there without Chat GPT correcting my swing every night, right?
• 12:57
  SPEAKER_00: Right, but if you had a coach, would you be out there going, you know, it's not actually me, it's the coach.
• 13:03
  SPEAKER_00: That's a great point.
• 13:03
  SPEAKER_00: If he did all of the research, right?
• 13:06
  SPEAKER_00: If he spent 350 conversations back and forth with his friend that's also in this industry and knows things, yeah, I wouldn't judge him for that.
• 13:16
  SPEAKER_00: Again, it doesn't change my opinion or your expertise.
• 13:18
  SPEAKER_00: And on the other side, when you look at like a trial, I don't look at I think about the average person and think about understanding whether or not a valve is at fault.
• 13:30
  SPEAKER_00: I know me, I have no idea about anything mechanical.
• 13:32
  SPEAKER_00: Yep, I would have no idea how to read that report.
• 13:35
  SPEAKER_00: What I would know how to do is be in a jury, listen to the person, right?
• 13:40
  SPEAKER_00: And chat GPT isn't doing the response.
• 13:43
  SPEAKER_00: If he's got an earpiece and he's doing the response that chat GPT is telling him, that guy sucks.
• 13:49
  SPEAKER_01: Yeah, yeah.
• 13:50
  SPEAKER_01: Okay.
• 13:50
  SPEAKER_02: So, all right, let me throw one extra thing.
• 13:53
  SPEAKER_02: So you have an expert.
• 13:54
  SPEAKER_02: We all know they're an expert.
• 13:56
  SPEAKER_02: Yep.
• 13:56
  SPEAKER_02: But they provide you outputs and you immediately know they did it in claw.
• 14:01
  SPEAKER_01: I discount it.
• 14:01
  SPEAKER_01: I mean, because it's a decl, it's a disclosure thing.
• 14:05
  SPEAKER_01: Like if you tell me, hey, if you say, hey, this is all Chat GPT wrote every word of this, but I put in three hours of iteration, right?
• 14:15
  SPEAKER_01: Like, uh I mean, okay, that's your expert influencing that output.
• 14:20
  SPEAKER_01: Right.
• 14:20
  SPEAKER_01: That I mean, I I I feel like it's disclosure.
• 14:23
  SPEAKER_01: So I want to ask one question here.
• 14:24
  SPEAKER_01: I want to ask a question, which is I feel like we're very close to a point in time where uh a reasoning we we have an automated reasoning engine that is that is very good.
• 14:36
  SPEAKER_01: We're probably already there, quite honestly.
• 14:38
  SPEAKER_01: That that is actually a better better at reasoning than 99% of humans out there.
• 14:42
  SPEAKER_01: Okay.
• 14:44
  SPEAKER_01: How far are we from saying, or is it science fiction that every expert in the courtroom, there is no defendant expert opinion.
• 14:53
  SPEAKER_01: There is no it is we have the reasoning engine and we have the context, which maybe we can debate, but we have the context of the incident and we can debate what that is, but we're not gonna debate the reasoning and we're gonna have an expert opinion that is actually a piece of software.
• 15:10
  SPEAKER_01: It seems science fiction, but it's like this is actually probably pretty close to what just happened.
• 15:14
  SPEAKER_00: Okay, if my my question, if he would have rewritten that report, if he would have taken those 30 pages, rewritten it in his own words, yeah, we would not be having this conversation.
• 15:23
  SPEAKER_00: I am sure that is happening right now.
• 15:25
  SPEAKER_01: Yeah.
• 15:26
  SPEAKER_00: Where people, if you just don't make it look like Chad GPT output, yeah, it's gonna happen.
• 15:31
  SPEAKER_00: People are using these tools, but I'm gonna go back to I mean, I do it.
• 15:36
  SPEAKER_01: I say write this, take that output, and make it sound like me with my spelling errors.
• 15:40
  SPEAKER_01: Yeah.
• 15:40
  SPEAKER_01: I mean, I mean, you know, I mean, it's ever I think everybody does that.
• 15:43
  SPEAKER_00: And I've got my platform flip-flops on right now, so we're going back to the 90s.
• 15:47
  SPEAKER_00: Um think about long-term capital management.
• 15:50
  SPEAKER_01: Yeah.
• 15:51
  SPEAKER_00: Okay, it was arbitrage, right?
• 15:52
  SPEAKER_00: Like this is they were printing money.
• 15:54
  SPEAKER_00: I mean, these guys were 20% year over year on a hedge fund and like good for you, but there's this one piece of context that it didn't know.
• 16:04
  SPEAKER_00: And I think if I recall rightly, with long-term capital, it was the Russian market crashed.
• 16:09
  SPEAKER_00: It was well, but it there was a thing that happened completely outside of any of the information that model had that caused all of the issues.
• 16:19
  SPEAKER_01: Yep.
• 16:20
  SPEAKER_00: So that automated reasoning needs to happen.
• 16:24
  SPEAKER_00: Yeah, but we need it to how do we provide it every single piece of context?
• 16:29
  SPEAKER_01: Right.
• 16:30
  SPEAKER_00: It doesn't know the maintenance records of what happened in that building.
• 16:34
  SPEAKER_01: But the guy might, I mean, it's like what's the baseline there too?
• 16:37
  SPEAKER_01: Because it's like, I mean, the guy doesn't know it.
• 16:40
  SPEAKER_00: But for sure, but he might, you know, if you're a human and you hear that, you're gonna, oh, I'm gonna go and talk to the maintenance guy.
• 16:47
  SPEAKER_01: Yeah.
• 16:47
  SPEAKER_00: Right?
• 16:48
  SPEAKER_00: Like all of that information that a human would do, if we provide that to a reasoning agent, reasoning agent is going to do better than a human.
• 16:55
  SPEAKER_01: Right.
• 16:55
  SPEAKER_00: But it is the things that it doesn't know that would give me pause.
• 16:59
  SPEAKER_00: Yeah.
• 17:00
  SPEAKER_01: So it's just how it's context.
• 17:01
  SPEAKER_00: Yeah.
• 17:02
  SPEAKER_00: How do we get the context to go into the system?
• 17:04
  SPEAKER_00: It's just a weird world.
• 17:05
  SPEAKER_00: It all comes back to context.
• 17:08
  SPEAKER_02: So think on this.
• 17:09
  SPEAKER_01: Yeah.
• 17:09
  SPEAKER_02: If that expert knew that all the reports he built were watermarked with a Chat GPT logo at the bottom, would he have presented them in court?
• 17:18
  SPEAKER_00: I I agree.
• 17:19
  SPEAKER_00: Disclosure.
• 17:20
  SPEAKER_00: Right.
• 17:20
  SPEAKER_01: I so to me, I'd be way more comfortable with that if that's the reality.
• 17:24
  SPEAKER_01: Otherwise, I think we have to think about like a clean room for expert opinions to be like you go, you go formulate your opinion in this clean room where you don't have access to technology we don't have access to.
• 17:35
  SPEAKER_00: So you know DJ's world, we've got expert witnesses that are honest and they're gonna just do the right thing every time.
• 17:41
  SPEAKER_00: And in Matt's world, our expert witnesses are getting paid like 15 times as much because they got to go isolate themselves for a few weeks.
• 17:48
  SPEAKER_01: Yeah, and maybe that's it.
• 17:50
  SPEAKER_01: Yeah, that's it.
• 17:51
  SPEAKER_01: Well, awesome conversation, guys.
• 17:52
  SPEAKER_01: Honestly, I feel like once this goes off air, we're gonna stay here for another half hour and talk.
• 17:57
  SPEAKER_01: But uh, that's it for uh today's episode of Logics Lee Live.
• 18:01
  SPEAKER_01: Want to thank uh DJ.
• 18:03
  SPEAKER_01: Uh you weren't as bad as I thought you were gonna be.
• 18:08
  SPEAKER_01: Can we thank Wendy and Sarah too?
• 18:09
  SPEAKER_01: And Wendy and Sarah, thank you guys as always, and Bakewall for the awesome discussion.
• 18:15
  SPEAKER_01: Um, as a reminder, if there was something here that was interesting or if there's something else going on in your world that's interesting, reach out, set up some time.
• 18:22
  SPEAKER_01: Happy to jump on a call.
• 18:23
  SPEAKER_01: Love to talk through what's happening in your reality.
• 18:26
  SPEAKER_01: And uh until next time, keep learning, even when it's unsupervised.