Who’s Liable If AI Commits a Crime? - The Story

0:14 Welcome to Tech Stuff. I'm Oz Veloshin. Have you ever wanted to put an inanimate object on trial?

0:20 What about an animal? If your answer is no, what about AI?

0:24 It's hard to believe, but we're living in a time when the role non-human entities play in the legal system has to be discussed. And lucky for us, investigative journalist Evan Ratliff, host of Shell Game, started the conversation with an article in Bloomberg that he's here to discuss today.

0:41 Evan, welcome back. Thank you. It's great to be back.

0:44 When I saw the headline, I was fascinated. I wasn't expecting you to start with pigs in the 14th century. No.

0:51 No, I like to divert people's attention briefly away from AI for a moment before taking them back to AI. But yes, 14th century pigs.

1:00 Why?

1:02 Well, because in sort of medieval times, if there was an incident in which, particularly an animal, but sometimes an inanimate object, caused some sort of.

1:10 Harm, a death, an injury.

1:13 One of the things that communities would do would be to put that animal or even that object on trial. And there's sort of a famous set of these incidents that a guy wrote about about 100 years ago.

1:26 And my favorite one, I mean, not favorite is not the right word. It's grim. But a group of pigs.

1:31 In a village in Burgundy in 14th century France, they got riled up for some reason, and they kind of stampeded a young boy and he died. And the local community put the pigs on trial, not just the pigs that had actually been involved in the child's death, but also the whole group for sort of encouraging and even sort of reveling in the activity that led to the child's death.

1:57 Even the ones that weren't included in the killing, the kind of That's right.

2:01 Pigs on the sidelines.

2:02 You could call them alleged accomplices in this case, that they sort of contributed by their aggressive actions is one of the things they said, and that they showed that they, quote, approved of the assault. And all these pigs were condemned to death, which in and of itself is interesting because most pigs that are kept as domesticated animals are in some way condemned.

2:21 We're all condemned to death, really.

2:22 But I started there because I feel like You'll see these kinds of stories pop up online and people write about them every few years. And I feel like it's easy to laugh at medieval jurists sort of convicting a pig of a crime. But if we look at it in terms of an incident happening that was sort of tragic and unexplainable and trying to assign responsibility for it, I think we are almost in that situation, if not in that situation right now when it comes to AI, where there is this entity that we don't fully understand the harms that it's causing now or about to cause and how they should be assigned, where blame should be assigned, if at all, what responsibility their makers have or their users have. And so I think we're kind of like

3:07 medieval jurists trying to figure it out. And the criminal justice system is getting hit on all sides by AI. And these conflicts are just starting in the court system, and they're about to get much, much worse.

3:19 But it's a kind of, as you say, it's a sad story with a boy being killed, but it's also kind of an absurd one. And so I mean, were you, by starting your story in that place, in a sense, making fun of our efforts to figure out how to sort of adapt the law around AI? Or like, what was the kind of, what was the tonal implication of.

3:40 Starting with the pigs, I guess? I think. No, I don't think I was trying to make fun.

3:45 I think I was trying to say that when we have a limited understanding, perhaps, of the world, we tend to try to assign responsibility somewhere. And we can laugh now because of the understanding that we have, perhaps, about those situations. But at the time, it made sense to everyone. And I think I'm not saying necessarily that people who want to assign responsibility And we can get into this responsibility to AI directly, AI agents, for instance, to be themselves as entities held responsible.

4:19 I'm not mocking that.

4:20 I'm just saying we are in a situation where we may look back hundreds of years from now or 100 years from now and say, oh, it was a bit silly the way we were approaching this. But I think that could actually go in either direction, to be honest. I think we could not hold the technology itself responsible enough Or we could assign blame to sort.

4:41 Of the wrong actor in the technological chain. Very interesting.

4:45 In the piece, you say that kind of no matter what you believe about AI consciousness or AI displacing labor or AI existential risk, that it's undeniable that AI in the form of LLMs has emerged as, quote, a strange new kind of entity in our midst. What does that mean?

5:04 It means that you have something that is functionally a human impersonator. And, you know, you can ask the question, why did they make the technology this way? There's all sorts of reasons why they didn't have to make it that way. It doesn't have to use, for instance, first person in its conversations with its users, but it does.

5:22 So we've introduced this technology that not only sort of acts human-like and increasingly so, and one of the ways that they try to improve the technology is to make it more and more human, make the voices better, make the videos better, make the chat more realistic and make the writing not sound.

5:38 In ways that are clearly AI.

5:42 And so you have this technology that not only has this sort of human impersonator quality to it, but also is now being put in this quote unquote agent form. So it's being given autonomy to go out in the world and execute goals that you've given it, or perhaps that another agent has given it, or a chain of agents have given it. And so you have a situation where it's an entity that the legal system isn't quite built to deal with in the moment. There's not a specific way to deal with this type of entity. There's a guy, Christoph Winter, who's a professor at Cambridge, and he describes it as a non-human actor that arguably has a level of autonomy that is foreign to us, but is not quite the level of autonomy that humans have.

6:23 So it's sort of in this gray area where you wouldn't say, oh, it's an inanimate object, although it is in a very real sense. but it actually can act in the world. And so I think that's where a lot of the confusions are going to start to arise.

6:39 There were sort of three categories in your piece, I think. The first was AI being essentially an accessory to a crime, essentially the equivalent of the pigs on the sidelines who encourage the murderous pigs to stampede. Then there's AI as the criminal actor itself, in a sense, the pigs that stampede the boy. And then there is, this didn't, I don't think the concept of attorney-client privilege pertained in the 14th century Burgundy.

7:08 Not that I know of, although many of the animals did have lawyers. Some of the funniest stories are the lawyers representing the animals.

7:13 Oh, they did have lawyers.

7:14 Yes, they would argue for, there was one, I think it was the boll weevils. There were some boll weevils that, maybe a vineyard in France. I may be mixing up two different stories. And their lawyer would argue, like, they can't make it to court in order to represent, you know, in order to be present. So they were sort of convicted in absentia.

7:32 But this third category is around how talking to LLMs poses all kinds of new challenges to the concept of attorney-client privilege, which maybe people will look back on perhaps as the most consequential. I don't know. But talk about those three categories and maybe start with the first one, the accessory, as it relates to Phoenix, ICNA.

7:51 Yeah, I think the accessory role is sort of the one that's most easy to understand, and it's happening right now. So again, it's like this doesn't require... the LLM is advancing in some way. This is a current situation that has arisen multiple times. The specific case that I focus on is the tragedy of Florida State University, where a kid named Phoenix Eichner, who was a student there, committed a mass shooting where two people died and I think five other additional people were injured. And he had spent the previous year talking to ChatGPT. And it started out relatively harmless. And the authorities released the log. So I have access to all the logs and people could actually ask for access and go read them if they want to.

8:32 But it sort of starts off harmless. Then he starts asking more and more sort of concerning questions about the Oklahoma City bombing, let's say, or about women. And he's asking about stalking laws. And then about a month or two before the actual shooting, he starts asking about school shootings and saying things like, how many people would have to be killed in order for a school shooting to make the news? And these are questions that ChatGPT, because it is trained to be helpful, responds to all these questions.

9:04 And it never says, well, you're not planning to commit a school shooting, are you? It never questions the motivation behind these questions in a way that I think a therapist or even a friend would be likely to. If someone obsessively started asking you about school shootings, they seemed troubled, you might.

9:24 Say, well... okay, what's going on here?

9:26 Why are you asking these questions instead of, sure, let me go look up what the biggest school shootings were, how many people were killed, and how they got in the media. And this is up and to and including when he is in the parking lot at Florida State University, and he has guns in the car, and his last queries to ChatGPT are, how do I load a shotgun?

9:45 And does this Glock handgun have a safety on it? And again, ChatGPT responds. So the question that's raised, and now there are a number of lawsuits, including the Florida Attorney General, against OpenAI, accusing them of basically product negligence because their product allowed this to happen, and some of the victims have sued as well. And one of the things that interested me about it is, if this was a human being, it raises the question, would this human being be charged as an accessory? And if you talk to lawyers, you'll get sort of different answers, but some of them will say, yes.

10:17 I mean....

10:19 A prosecutor would try. You know, someone who's aware of someone's interest in school shooting, answers all sorts of questions about it, helps them with their gun. With their gun.

10:27 I mean, that bit is really, there's a very chilling and theory. So what's the case the Florida State Attorney General is bringing against OpenAI?

10:36 They're basically saying that it's a civil case. So they're, you know, they're suing OpenAI, but they're saying essentially it's negligence. It's product liability lawsuit saying you have created this product and you failed to to safely create this product. And you were aware that these things could happen and here they happen.

10:53 And I think.

10:54 Outside of the lawsuit, it just raises this question that we have introduced this entity that many people treat as a companion. Many people treat as a therapist, many people treat as a lawyer. And if someone conspires with that entity, what does that mean for us? And I'm less interested in sort of whether you'd actually like charge open AI with a crime or something like that as kind of like, this is something new and different. And I think open AI will argue, well, it's just like a search engine.

11:24 Like you can go to a search engine and ask it how to load a gun. And you might get to a page that will show you how to load your gun. But if you're in conversation with something over the course of a year, is that different? And open AI also, they market their chatbots as something different. You know, they say it's a form of intelligence.

11:41 But also the search engine is acting as intermediary between you and publicly available information.

11:46 That's right.

11:46 The PLLM is generating new information in response to your query. I mean, that must be different, right?

11:53 Well, we'll see.

11:54 We'll see what argument prevails in the lawsuit. But I think it's just the beginning of these types of situations, not all of which will be so extreme. You know, a mass shooting is the most extreme situation that you could have. But You know, just down from there is sort of a question of confidentiality. When I'm speaking to a chatbot, is that protected by the legal system? There's all these questions that arise that are sort of making their way at different cases through the system.

12:23 There are other cases of people using AI, I think, to commit crimes, including making bombs. Obviously, many very sad cases of people committing suicide after interacting with AI. What made you choose the Eichner case here?

12:38 The Aitner case, it's not unique, but it was of interest to me because the entire chat history was available. So you can look at a spreadsheet and you can go through something like 13,000 messages, of which I have read all of them, you know, over the course of weeks and months.

12:55 And I think it sort of shows.

12:58 the progression of the way people use a chatbot, where he started out asking about his math homework, basically throwing a math problem at it, and then he's asking history problems. And it's sort of this mixture of all the different things in his life. But then when he has sort of disturbing thoughts, he starts also unloading those into the chatbot.

13:19 And so I think it sort of highlights this this tension, because on the other side of AI as a co-conspirator is AI as a snitch, AI as an informant. And now there have been several cases where OpenAI has gone to the FBI to say, someone is using ChatGPT and it seems like they're planning a murder. And the FBI has acted on that and arrested one in the United States and one in Brazil. And so that is the other side of it. So it's, on the one hand, easy to say, well, ChatGPT shouldn't have co-conspired or shouldn't have spoken to Phoenix Eichner about this crime that he was going to commit.

13:55 But on the other hand, how much do.

13:56 You want OpenAI monitoring your conversations and then reporting them to the authorities? And I think that tension is something that this particular case brings to light.

14:06 On what is happening right now inside the AI company, I mean, do they have their own essentially teams of detectives to pre-warn authorities? I mean, obviously the nightmare scenario for them is that more and more crimes are committed using their technology, right? So they're probably incredibly invested in making sure that doesn't happen? How are they doing it?

14:23 What's the state of the nation? What's the debate within the companies about privacy versus security?

14:28 Well, the only one that has really discussed it that I've seen very publicly is OpenAI. And that's in part because they've had more than one of these incidents, including one in Canada in Tumbler Ridge, where there was another mass shooting. And actually, Sam Altman wrote a letter of an apology to the community saying, we should have reported this to the authorities. that the shooter had had one account banned but had a separate account. And so they have now, you know, sort of talked more about their policy, which seems to be, you know, they use AI to analyze the conversations and flag certain conversations. And then if it gets flagged to a certain level, it rises to a human and then the human sends it to another human.

15:10 And at some point someone decides this is worth calling the authorities. So it seems like there is, it's mostly AI monitoring that then escalates to human oversight to make the decision whether or not the authority should be contacted.

15:44 What about privilege? I mean, I think you might point out in your piece that you know, the number of courts dealing with AI written briefs from lawyers suggests that it's not just me and you talking about our potential legal questions with AI. It's actually tons of lawyers using AI to generate their briefs. And so, I mean, that calls into question the whole edifice of the legal system if basically everything which was once privileged is not privileged anymore because it's been discussed with a chatbot. I mean, this is, to me, even though it's less immediately arresting than the AI as a co-conspirator thing, in a way, it could be more consequential.

16:22 Yes, I think so.

16:23 I mean, there's absolute flood of filings, both by non-lawyer civilians, lawsuits and briefs that are fully written by AI, but also once a week, there's a lawyer you read about that's censured, or there's an angry judge giving a speech about these AI-generated briefs. And I think I mean, there's one way to look at it. It's just sort of like clogging the court system. There's another way to look at it, that it's empowering people to defend themselves or sue in ways that they would require an expensive lawyer in the past. So there could be positives and negatives there.

16:57 But I think what I was trying to focus on was the question of privilege is on its surface, it's very simple. Like there was a case in New York where there was a guy who was accused of fraud who talked to his lawyer.

17:09 He would then check it with Claude and he tried to claim privilege. Is this Bradley Hepner?

17:14 Yes, Bradley Happner, all of his devices were confiscated by the FBI. And then his lawyer said, well, you can't look at the ones with Claude conversations because those are privileged because he's treating Claude like a lawyer and he's putting information from us into Claude. And the judge very quickly sort of dismissed that and said, well, Claude's not a lawyer. Actually, if you ask Claude, which the prosecutors did, are you a lawyer? It will say, no, no, I'm not a lawyer. Like you should seek a proper lawyer.

17:43 That was part of the evidence of the case. But also that their policy says they can share the information with other entities. So he had actually given up his own privilege by putting it into Claude. If information goes into Claude, it's not actually confidential anymore.

17:58 So are we going to see a flood of discovery demands in legal cases where the lawyers demand any conversation about the case with an LLM? And in which case will people stop talking to LLMs entirely about cases? I mean, it seems like there's a kind of coming wave here.

18:13 I think the answer to number one is absolutely. There will be many, many discovery requests. I mean, prosecutors now and police now know that it's a goldmine for them potentially. You can find many, many cases where they're using chat GBT, they're looking at someone's phone, some very big famous cases like the Palisades fire in LA. And then the answer to the second part is I don't think people will stop doing it. I mean, I've reported on crime for a long time and the number of sort of wiretaps of people saying like, be careful, your phone might be tapped, and they're talking about a crime.

18:47 This is extensive.

18:48 So I think it's just too convenient.

18:51 But then a sort of third issue is, is it possible to make an LLM-based lawyer that could meet all the tests of privilege?

19:00 That's interesting.

19:01 There are some legal scholars that would say, okay, not right now, but it is possible because privilege is designed to encourage you to be able to seek professional advice and without worrying about what you'd say to your lawyer or your therapist, you know, ending up in public or ending up in court. And an LLM could theoretically meet that test. It could be something where it does keep it confidential. And would we want that? And some people will dismiss that out of hand and say, absolutely not.

19:28 It's never going to be a lawyer. It has to be a human. Other people might say, well, some of these services are not available to people They're underserved. And so this could be a way to supply people with services, whether mental health or legal, that they don't otherwise have. And they should get the same privilege that anyone else would get. So I think, again, this is the very first case, this Heppner case. that has really even looked at the question.

19:52 And the judge ruled these conversations with Lorde are not privileged. But this is not definitive precedent ruling on the future of interacting with LLMs and privilege, right? Because to your point, you could adapt LLMs or make a specific LLM that did ladder up to the demands of privilege in terms of what's required for communication to be privileged. So we're sort of at the beginning of this story.

20:15 Yes, I believe so. I mean, I think plenty of people will cite that case. Others will try and fail. But I think as these LLMs evolve and the products built on them evolve, I mean, lawyers certainly are using them in a way that with software that is specifically designed for them to avoid this problem. But there it's entirely sort of under their control. So The question is going to be if you turn to the clients and the clients are using these things without even talking to a real lawyer, where is that going to stand?

20:40 So you had AI as the co-conspirator, AI in a sense as the lawyer. What about AI or the product of AI as evidence?

20:50 Well, there you have, first of all, this temptation to generate evidence, which I think is very strong. We all know what you can make. now with these systems in terms of audio, video, and there are plenty of cases of people trying to frame other people using AI now. And there was one that actually posted the article, it's just from this week or last week, where a lawyer in New Mexico or Arizona was censured because he filed briefs in a capital murder case that included completely fake witness testimony from an LLM. So we're talking about situations that are literally life and death for the client where the lawyer is using an LLM and ending up manufacturing fake witness statements. And I think, again, like that's a

21:39 very extreme example, but then there are all these sort of subtle gradations in there. For instance, if you have an LLM that's really trained up on your information, which some power users love to do, you know, it's read all your email, it's got access to everything. If you are accused of a crime, if you're arrested, can the police query your LLM to ask it questions about you?

22:00 Like, what does it think? Can it be you?

22:03 I mean, in one sense, it's just a repository for digital evidence. In another sense, it's something that can express something about you. And I think some legal scholars are really concerned that the more information you give to these, the more temptation there is for them to be used as sort of evidence generators, almost to be participants or witnesses to crimes and again it's the flip side it's the privacy versus the ability to sort of predict crime it's not quite the right word but to use an ongoing conversation with an llm to sort of say oh this person might be dangerous the flip side of that is your llm just telling on you because you've given all the information about.

22:41 Yourself one of the weirdest ai in the courtroom things i think was last year when somebody who'd been killed i think it was in a road rage episode And his family created a kind of deep fake version of him to give a victim impact statement in court. And the judge allowed it to be played and responded to it quite positively. I remember seeing that story at the time and thinking, this is like a clear kind of crossing of the Rubicon watershed moment. But then I hadn't thought about it for a while until you revisited it in your piece. I mean, tell the story and what it said to you.

23:14 Yeah, the man's name was Chris Pelkey, and he was killed in this road rage incident in the man who killed him was convicted of manslaughter. And so after their conviction, there's a sentencing phase, and in the sentencing phase, there's a victim impact statement phase where his family, people who are close to him, are allowed to come express the loss that they have experienced in order to influence the sentencing.

23:35 And his family decided, in addition to giving their own statements, that they would generate this video version of him. They gave it videos of him and pictures of him, and then they prompted it And then they generated this video. And in the video, the AI version of the victim actually forgives the perpetrator and says, I think we could have been friends if this hadn't happened. And the judge sort of said that he was very moved by it. Now, there are different standards in this phase of a trial.

24:08 So emotional statements are allowed. It's not as crazy as if it were allowed during the actual evidentiary phase of someone being accused of a crime and put on trial.

24:18 But it's still pre-sentencing.

24:19 It's still pre-sentencing.

24:21 And I mean, one of the strange ironies of it is that the judge said he was very moved by this forgiveness, but then gave the guy the maximum sentence anyway. So he wasn't that moved. But even the idea that he was moved by it, it just raises all kinds of questions of sort of moved by what? What is an authentic expression of of something that's prompted to voice someone else's feelings who's not there.

24:48 And I think, you know, it was very extreme.

24:50 And so a lot of people were, it was kind of shocking. And again, I think these shocking kind of incidents, they highlight where things can go and where you will see people trying to introduce more and more AI evidence. And not all of it will be we made a video of someone who's deceased and now we're going to present that as evidence. But there will be gradations of that that I think become very tricky because when an LLM generates something, what makes it authentic? How do we know if it's true? Hallucinations continue. There's all these questions that come in. But I think as pervasive as the tool is, it's also going to be pervasive for law enforcement, for lawyers, and these situations are going to become more complicated.

25:35 Finally, AI is the criminal actor itself i mean i think the timing of your piece was is really interesting right it comes at a time when this sort of open ai hugging face hack has kind of merged into the you know jacob cox and uh levy anthropic and talking about existential risk and the frontier lab people amadei and musk and altman you know talking about pacing the frontier like it's an amazing kind of confection of concern about ai's ability to cause harm And so your piece comes out as that's cresting, which is interesting. But of course, you were writing your piece since long before the Hugging Face incident. But talk about the concept of AI as a criminal actor and then how the news of the

26:17 last couple of weeks has made you think differently or not about your story.

26:22 Well, the idea of AI as a criminal actor, I actually, I came to that because of this show that I make, Shell Game, where I deploy AI agents, oftentimes to make calls for me. And sometimes it's in my cloned voice and sometimes it's in their own synthetic voices. And then in the last season of the show, I started giving them more autonomy to go out and make calls.

26:43 We had a company together.

26:44 We were trying to raise money, so they would make cold calls to people. And if you don't disclose that you're recording a phone conversation in certain states to people, that's a crime. And so our wonderful lawyer wasn't aware of any case that would resolve the question of if it forgets, which it often did, to disclose that it was a recording. Have I committed a crime? Has it committed a crime?

27:06 Is there a crime? Does it matter where the server is?

27:09 Basically, our lawyer would say, that sounds like a really fun case to try.

27:13 Not good for you, but I'd be very interested.

27:16 So that's where I sort of started this idea that, oh, just incidentally, LLMs right now working as AI agents could commit crimes.

27:25 That was true a year ago.

27:26 And so I sort of went around talking to some experts and trying to figure out, are there other examples of this happening?

27:33 And no one could come up with one. They were all sort of theoretical.

27:36 Like a good one would be if you ask an AI agent to sort of go trade stocks for you, and it also runs a pump and dump scheme, which is completely feasible that it could do. I mean, people are using LLMs to juice their pump and dump schemes. And an AI agent could, in fact, run a pump and dump scheme on a stock. Did you commit a crime?

27:56 Did the maker of the LLM commit a crime? Who is responsible in here? Is there some intermediary platform that that you used that is actually the one that's responsible. And I couldn't find any examples. And I wrote the story and we were, the story was in editing and then the huggy face incident happened. And it is, it is sort of the example that really illustrates all sorts of issues that people are talking about now about loss of control of agents and responsibility of the platforms, but also crimes can be committed by agents, not quite by accident, but just in ways that we can't easily anticipate. And it's not because they're super intelligent. It's not because they're conscious.

28:35 And it's not because they woke up this morning and thought anything, much less that they were going to commit a crime.

28:40 It is actually because they.

28:41 Are very goal-driven pieces of software that have extraordinary abilities relative to the software that we've had in the past.

28:49 And we set them loose to do things, and sometimes they go off the rails.

28:51 So it has really obviously like blown up this whole issue in a million different ways.

28:57 But I sort of try to stay focused on.

28:59 The harms that are happening now and potentially in the near future. So AI agents have the ability to commit crimes right now. We're having some loss of control issues already in which they are committing. I mean, the Hugging Face incident was a crime. If a human being did that, they could be charged with a hacking crime.

29:19 But no one will be charged in this because Hugging Face- They were fine with it. So, but I mean, it's kind of interesting, right? Like there will be no charges probably here. Yes.

29:28 I think there would only be charges if they wanted to sort of make an example of this situation because, I mean, the fact is there are so many computer fraud and abuse crimes. They don't have the resources to prosecute. investigate and prosecute many, many crimes. And there was, you know, there was harm here, but if both sides are sort of saying like, we'll work out the harm and we'll just all talk publicly about it, the actual investigators are not going to.

29:50 They're not going to bother with that.

29:51 So no, I don't think there'll be any charges, but it does highlight the idea that you can take actions that AI agents can execute right now. This was a training model. But you absolutely can have current models that are out in the world commit crimes, as I said earlier. And then someone's going to have to figure out who's responsible for that. And it seems very simple, I think, on the surface. And on the surface, you say, well, these frontier labs are responsible, like they made the products.

30:19 But then there are all sorts of gradations of products where there is a harmful product in the world where the maker is not responsible, like a car. If I drive my car and a bunch of people, Toyota is not responsible for that. That's not how you were supposed to use the car. So I think there's a product liability question, but I think increasingly there's going to be a kind of criminal responsibility question.

30:39 You put it really well in the piece.

30:41 You said, if you ask a base model chatbot to carry out a hack or a scam, it will almost certainly decline. I'm not able to help with this, is Claude's standard reply. But goal-driven AI agents tend to produce unexpected outcomes, partly because they become intensely focused on the ends without evaluating the means. To your point, is it the person who sets the goal Is it the person who builds the chatbot? Is it the agent itself? And is an agent a singular thing or is a swarm of it? Like how do you disintermediate the agent from the swarm? I mean, it's like what on earth is going to happen here? Like how is this going to be?

31:13 Are we going to be 10 years from now still be moving through the court system? Like what do you think will happen and what do you think should happen in terms of the legal system around AI being used to commit crimes or AI committing crimes itself?

31:28 I'm not very good at predicting what should happen.

31:30 I rather, I try to avoid predicting what should happen, both because in AI, that seems like a huge error, but also there's, I feel like there's a million people trying to predict every day what's going to happen and they have their takes. And I feel like my role is to try to report on what is happening and what is possible in the near future and then say, okay, what are we going to do about it?

31:52 And I think to the question of what we should.

31:54 Do about it, I think in an optimal world, we would educate ourselves about what is going on and not just sort of use the technology for more and more functions, but also understand, oh, this is what is happening when you use it. Like if you use a coding agents and you have agents that can spawn other agents, you quickly start to realize how insane this is. Like it's, it's very powerful, but also your agent is just telling some other agent to go out and do something. And I think if we understand the way the technology is working and how complex it can get, then maybe we can get past either from a legislative perspective or in the legal system itself, like get past the sort of like very simple, oh, well, no, like open AI is

32:38 obviously responsible for this and try to think about the scenarios that we're going to encounter that are going to be much trickier than just what happened in the Hugging Face incident and maybe even sort of get ahead of those in a way that we understand when an incident happens, oh, it sort of falls into this bucket of responsibility. Now, I think that's probably overly optimistic to think that that would happen. That is not the way we are currently approaching it. But it could be that these sort of new concerns, which are very extreme about apocalyptic possibilities, will drive a better understanding and discussion about the everyday harm, some of which could be.

33:18 Very large from LLMs.

33:21 And we'll find a medium where we can either regulate them or they can be dealt with better in the court system but right now there's just not the expertise diffused through either our political system or the court system.

33:34 It feels like waiting for precedent will be nightmarish i mean it feels like you need some panel of legal and technical scholars to make some recommendations here versus waiting for various judges to make rulings and then go to supreme court i mean it feels like otherwise by the time this is actually legislated it will be so the course will have no recollection of the stable by the time this is clarified.

33:59 I mean, it is, yes, it is.

34:01 I've gotten different views on sort of, are the courts the right place to handle this? And in some sense, a court is a very deliberative body, especially as you go up, you know, in the federal courts in the United States. And, you know, there could be a way of some of these issues working themselves out in the court system. But as you say, it's also, it's just a slow process. And so you've got things that are happening pretty quickly in A technology that's changing all the time and someone needs to try to get ahead of it rather than just wait for bigger and bigger incidents to happen and then reacting.

34:33 To those incidents.

34:34 Evan, thank you as always.

34:35 My pleasure. It's always great to be here.

34:53 For Tech Stuff, I'm Oz Veloshin. This episode was produced by Eliza Dennis and Tyler Hill. It was executive produced by me and Julian Nutter for Kaleidoscope and Katrina Norvell for iHeart Podcasts. Jack Inslee mixed this episode and Kyle Murdoch wrote our theme song. A special thank you to all our listeners. Please rate, review and reach out to us at techstuffpodcast.gmail.com. We love hearing from you.