Law in the Digital Age: Exploring the Legal Intricacies of Artificial Intelligence [e323]

Full Podcast Transcript

NASIR: We are covering artificial intelligence as it applies to the legal world. Probably the most talked about issue in 2023.

MATT: A lot of question marks.

NASIR: Frankly, the dangers of it. I don’t know.

MATT: We call that a lose-lose.

NASIR: I would want to own that copyright.

MATT: It’s a level of human input.

This is Legally Sound Smart Business where your hosts – Nasir Pasha and Matt Staub – cover business in the news and add their awesome legal twist. Legally Sound Smart Business is a podcast brought to you by Pasha Law PC – a law firm representing your business in California, Illinois, New York, and Texas. Here are your hosts, Nasir Pasha and Matt Staub.

NASIR: Welcome to our podcast today! We are covering artificial intelligence as it applies to the legal world. One of the big news for this particular podcast is this is entirely generated by AI. I am talking, my image, everything here is completely through artificial intelligence generated. It’s a new technology. What do you guys think so far?

MATT: You know what’s disappointing? I actually came with the exact same joke or same line, and you stole it from me.

NASIR: Great minds think alike. Great comedians too as well? I don’t know.

MATT: I don’t know if we can consider ourselves comedians. We try, but… For attorneys, I think we can probably justify it, but in the general population, I don’t think so.

NASIR: Maybe in the realm of dad jokes.

MATT: That’s true.

NASIR: I’ll take that. From that standard, I thought that was pretty good.

MATT: Yes.

NASIR: This is really me as far as you can tell. In reality, maybe ten years from now, you wouldn’t be able to tell, but we are going to be covering probably the most talked about issue in 2023 which is AI or artificial intelligence.

I think the last time we had this kind of topic from a legal perspective to really parse out and hash out like this was probably for COVID-19. In fact, I think we did a whole episode on all the legal implications of COVID-19 – everything from working from home and the vaccinations and these kinds of things. But artificial intelligence in particular has brought in all these new legal issues.

One thing that we said back in COVID – if you recall, Matt – even though it’s bringing on new issues, it is still based upon old law. This is what happens with technology or a new pandemic. The law is slow to adjust, so there are going to be things that are based upon what’s happened in the past to build analogies to apply to the future. We’re going to talk about everything from copyright law to using AI in the workplace and some mishaps in using AI in business.

MATT: It’s similar to any emerging technology. At this point, it is very much in the forefront. Sometimes, we try to do podcasts that are evergreen. This one is obviously not going to be. If you listen to this a year from now, I imagine it is going to be much different the things we would say now versus 12 months from now. It’s definitely something that’s still emerging. We’re still learning. The law is very far behind, as you said, but we are going to at least try to place you in where it is right now today and go from there.

NASIR: Let’s get started. First, let’s talk about some of the things that AI has been doing, specifically in the workplace or as it applies to law.

The first thing I like to tell people in the very beginning is, when it comes to generative AI – most people know this already, but I am going to state it anyway just to level set – AI can produce what are called hallucinations. They are saying something that seems like a statement of fact that is absolutely true but is completely untrue.

What’s crazy about this is that, if you don’t do your research and actually double check these things, you can fall into a really bad trap. That’s what happened in a couple of cases in the actual court of law when attorneys actually relied upon artificial intelligence.

MATT: Yes, that’s the big thing. There’s this new tool that’s out there. Whether it’s in the workplace or not, people get a hold of it and are like, “This is great! It can cut the time I have to work on something in half! My efficiency is going to be much greater.” But like you said, you can’t rely on this 100 percent. You still have to double check your work. There are a couple of stories that we are definitely going to talk about on the legal side of things, but it’s not a cheat code, I guess you could say. I look at it as an assist; it can assist you with things, but it’s not an end all, be all, at the end of the day.

NASIR: At least not yet.

MATT: Not yet.

NASIR: Obviously, it’s getting so much better. For example, there’s a huge difference when it comes to hallucinations between ChatGPT3 and GPT4. Of course, that’s one generated model. There are plenty of others – Google Bard, et cetera.

This came out back in June or earlier of this year. I think GPT came out to the public some time in the late first quarter of this year. It was still relatively new, especially when it comes to lawyers actually using it in the real world or anyone in business using it in the real world.

It was a personal injury attorney – no offense, but I’m not too surprised about that – that had a personal injury lawsuit. Within the actual court documents, apparently, the attorney referred to more than one case, I believe – at least one of the cases – of a court case for some kind of legal precedent that did not even exist.

How did that happen? Well, very simply, he had ChatGPT write his pleading form and did not check the citation or actually check that that case existed.

MATT: I think you agree. I find this very funny in general. I have a couple of questions. This will be quick, but I can’t imagine it’s something he hadn’t done previously. I would imagine he could have pulled from a previous case or a previous file he would have done and just used that. But it cited cases that didn’t even exist. He brought it to court. I looked into it a little bit more.

Essentially, he admitted to what he did. His understanding was he could type in this information, and it would give him valid cases that he could cite. I don’t know what the end result was, but he had to go back and explain to the court why he shouldn’t be sanctioned for this. In my opinion, he should be because it’s ridiculous.

Like you said at the beginning, you have to double check it at this point. You cannot just rely on everything and assume that it’s going to be 100 percent accurate.

NASIR: For example, with a calculator, if I put in 2 plus 2, I am very certain that that calculator is going to put in the correct result. At the same time, I also understand that it’s still subject to human error. In other words, did I put it in correctly? Am I still checking to make sure, especially with a calculator that shows that?

In the same way, if a human being produces a pleading, I’m still checking to make sure that this is correct and makes sense and the case reference is actually correct. With that information together, if you have a bit of knowledge of how AI works – and we’re going to talk about copyright use in a little bit here, but – when the AI modeling goes out into the world and grabs these massive data sets, those data sets could have inaccuracies because it’s all produced by humans.

On top of that, if it’s a generative AI that is designed to create – and I think that’s something that people don’t realize – it has different purposes. In other words, it’s literally able to generate information, and it uses that information based upon its training with you. Because of how you ask the prompt in ChatGPT or others, it may think that you’re being creative.

I have done this with the kids. I don’t know if you’ve tried this. By the way, if anyone hasn’t done this, you should definitely try it. I’m getting on a tangent here, but it’s a good bedtime story. You say, “Let me tell you a story. What are some things that you want in it?” You can actually ask ChatGPT to create a bedtime story for kids of this age. You can put in certain character names that you want in there or other subject matters. My daughters always want a unicorn in there, for example. It creates these amazing kids’ stories. It’s really just incredible. But obviously, it’s just made up.

When you combine this made-up generative creative thought with fact-finding and things like this, sometimes the AI doesn’t know what you’re looking for, and you have to specify that. In a lot of ways, you can actually use the tool to generate these briefings, but you need to know how to use it – just like any tool.

MATT: It’s basically Mad Libs, but a good version of it, essentially.

NASIR: 100 percent, yes.

The other big story was with the Fugees. Apparently, one of the members – I don’t know who this was, and I don’t know if you know who this guy is – Pras Michel?

MATT: All I know is Fugees is Lauryn Hill, isn’t it? My knowledge is limited on this.

NASIR: Pop culture and sports – you lost me at hello there.

Anyway, Pras is a member of the Fugees. He was apparently found guilty in a criminal case for multiple political conspiracy charges. What’s interesting here is that he’s basically filing an appeal and making a claim for ineffective assistance of counsel.

Essentially, in criminal law, if you have an attorney that is really horrible and messes up your case, that is the basis for an appeal. It happens once in a while, but usually it’s things like “my attorney fell asleep during the trial” or completely missed a big issue or things like that. In this case, what he is alleging is that this attorney used this proprietary program called EyeLevel.AI which, by the way, most of these AI companies are just tapping into the API of ChatGPT. I don’t know if anyone knew that, but that’s 90 percent of the companies out there.

Basically, this attorney used this company or this software to generate a closing argument. Within the closing argument, the process attorney alleged that it made frivolous arguments, misunderstood the basic elements of the case, and also neglected to mention some of the flawed arguments or evidence of the government’s case. In other words, it was a bad closing argument. It’s definitely something interesting – something we haven’t seen before.

MATT: Yes, I’m very interested in what this attorney even inputted into this EyeLevel.AI to even draft this closing statement. You can’t take everything from a full trial, encompass it into one input on some sort of AI algorithm, and get something. There are so many facts, obviously, that went into this.

He had to have gotten it, too. He got whatever the result was. Do you not read it before you go into it? You look at it and think, “That’s okay. We’ll just roll with this.” There are a lot of questions I have on this. I think it’s a relatively recent article at this point, but a lot of question marks.

NASIR: If you go to the website now – I hate to put it on blast, but frankly, I’m not the one that did it; the guy from the Fugees did it – this EyeLevel.AI is interesting because they focus on the truth of the data in a sense that their tagline is “build truthful AI” and the primary goal there is to basically prevent hallucinations.

By the way, I was right – it says trained with GPT4. I’m pretty sure that’s how they hooked everything up, but they seem to have a specific tool for legal where you are able to gather thousands or even millions of documents and be able to explore complex theories among the, which is interesting because I know a little bit about that.

When you are dealing with large amounts of data, unless they are actually using that data to train – which a lot of these guys don’t do, especially if you are using GPT4 – what they do is they have to take all this information in chunks.

For example, if you were to ask a question or want some kind of insight over millions of documents, in order to actually do it, it has to go chunk by chunk to look for that question and then put all that information together. This is just the limitation of how the technology current works, token limits, and so forth.

If you think about it, it’s like saying ten people go out. Everyone reads a chapter of the book and then all ten of you come back and let’s piece that all together. You can understand how that process is inherently flawed in the sense that one chapter doesn’t understand the context of another chapter – unless it’s learning off the other one. Honestly, most AI models, especially in practical applications, work that way. There are exceptions.

As the technology gets better and better, you can have an AI model that looks at the entire body of data, but I don’t think we’re quite there yet, which is why it makes it dangerous to rely upon this too much.

MATT: Yes, I’m looking at the site as well now. It’s basically every stage of the litigation process. Again, the closing argument, everything is going to be so fact specific. At this point in time, I don’t know how it’s going to be able to draft something that’s going to be beneficial to you. Like you said, there were factual inaccuracies. A major misstep by this attorney, but as anyone that has listened to this podcast, we love attorney missteps, so we had to cover these.

NASIR: It is strange because, in a closing argument, you are part of the case. By the end of the case, you know the case just as well as any computer. Maybe for an opening statement? I don’t know.

By the way, I know we’re giving them a hard time. I wouldn’t be surprised if the closing argument was just fine. A lot of these ineffective assistance of counsel cases are often not successful. I think it takes a lot for them to actually occur because the assumption is the jury that actually found this guy guilty for this, so I doubt it was won or lost because of the closing argument. Who knows?

MATT: You’re right. I’m also curious how this one singer even found out about this. I

NASIR: Yes, how did he know that that was written that way? Maybe the attorney told him. I don’t know.

Let’s switch gears. We talked about AI use in the courtroom, but I think one of the big things that came up this last summer are the strikes of the SAG-AFTRA strikes related to all of the writers and these kinds of things in Hollywood.

One of the big issues is the use of AI in writing – even giving credits and these kinds of things.

MATT: Yes, there are obviously multiple issues with the whole strike, but this was a big component with it.

If you look at it from the top level, you can say, “I can just take what I want to do, plug it into ChatGPT or whatever you want to use, and they can produce something that is maybe 80 percent or 90 percent as good as an actual person I have to pay. Why would I pay for somebody when I can just do this for way cheaper?” That was one of the big pushes with this whole strike – some limitations or restrictions on using AI in that context.

NASIR: Right.

They still haven’t really negotiated or settled on this AI issue. I think they resolved some of the other financial aspects of it. They have thrown a lot of interesting ideas.

Where they are probably going to be heading towards is there is not going to be a prohibition of the use of AI, but I think the main thing is the compensation associated with it in the sense that – again, I keep alluding to this, but we’re going to talk about copyright a little bit later in the show – when it comes to how much human involvement is in the production of this product.

Even if it’s something that’s generated by AI, it’s still proofread and signed off on by a human being. If that writer is spending five minutes as opposed to an hour on a project because of AI, essentially, I think the request is that that writer gets the same amount of money regardless, which is not – I was going to say “unfair” but that’s really hard to judge. I get the argument on both sides.

MATT: Yes, it’s tough.

It is very difficult. One concept is going to be enforceability. Going back to what we were saying earlier, if I was a writer and I took a script and tried to rewrite it using AI, I don’t know how they would ever know. That’s one thing. But let me see some of the points they have made.

What they were looking for was AI may not write or rewrite original material. It won’t be considered source material if they do use AI. The employer cannot consent to it, but there has to be the consent there. Then, a company must disclose if any materials have been using AI or incorporating AI-generated materials. It’s everything you would expect from the writer’s perspective.

But I do think you are right. It’s going to be somewhere in the middle – between usage and a strict non-usage. What extent they can use it is going to be the big thing, but we’re going to have to see. Again, it could be finalized by the time this episode comes out, but at this point, it is still a question mark.

NASIR: I don’t know. AI is doing such phenomenal things, but it is just not quite there yet.

MATT: Yes.

NASIR: For example, when it comes to writing a five-minute script or even a joke, I can see it doing that, but being able to create a body of work as far as what’s generally available? Here’s the thing. One thing that was freaky to me was that, when ChatGPT came out, I think everyone was amazed by it.

But then I started thinking about how this is what’s being made available to the public. From that perspective, I’m really curious as to what is not available to use because it requires extra computing power or frankly the dangers of it. I don’t know. Anyway, we can go on.

One of the things that I think everyone has come to realize, especially those that are really diving deep and embracing the AI technology is that they are using it in every sort of way – whether it is personally trying to figure out what kind of medical treatment they need, to getting advise on finances, to revising contract agreements, or what-have-you.

But a lot of that stuff that you’re putting into the prompt is pretty private information. In fact, very private information to such an extent that some companies have completely prohibited their workers or employees from actually utilizing AI because of this risk. There are some real-world implications for this.

For example, Samsung apparently leaked data through ChatGPT for not only some of the meetings that they were having, some code, some proprietary information, and Samsung had to shut it down completely because of the concerns of privacy.

MATT: Yes, that’s a story we were looking at, too. I think this is an instance of employees just not knowing the extent of how powerful this tool is, obviously. I’m sure there are provisions in their agreement that they have to hold information on the job confidential, so they wouldn’t intentionally type in source code. The meeting recordings was even weirder. I don’t know why you would ever do that.

NASIR: You know what it is? Has anyone used Fireflies? There’s something called Fireflies AI. If you’re not familiar, I’ll summarize it really quick. Basically, on all the Zoom calls, Google Meets, and Microsoft Teams, Fireflies.ai will actually join the meeting for you – not only for you or in addition to you, but it will audio record the meeting then it will produce a transcript. It’ll also submit this information and it uses AI. I’m pretty sure they tap into GPT4 to summarize, call out action items and these kinds of things. I believe they use GPT to transcribe it as well, but I can’t verify that.

I’m pretty sure, when they say meeting recordings, I’m pretty sure that’s what was going on. You have an internal Samsung meeting. You have a third party here. By the way, we have had these same conversations with our clients because, if I’m on a call as an attorney with the CEO and other management and we’re having a conversation, we use Fireflies, too. We have a third-party app that is recording the conversation that is essentially taking notes for us.

Now, because it’s an attorney-client communication, that’s completely privileged and protected. No one can subpoena Fireflies.ai to produce a recording of a conversation that I had or that Matt had with our clients – unless they are able to find some kind of exception. However, if you have a board meeting or if you just have an executive meeting and Fireflies is on, that call is recorded, and an attorney is not on the line or doesn’t otherwise qualify for an attorney-client communication, that is completely discoverable.

Everyone understands, when it comes to business, people are careful what they put in writing. It used to be that you don’t put it in email. Just text. But now, of course, in trial litigation, people know to grab those texts. Then, it was like, “Let’s use WhatsApp or Telegram.” We’ve seen in the Trump hearings and the court cases surrounding that, they were actually able to obtain – one way or another – these Telegram transcripts. It also became “don’t text me, don’t write me, don’t email me – call me,” but it’s funny because these Fireflies.ai and other bots join automatically. You almost forget about it.

I’ve had multiple instances where you’re on a call and someone else is like, “Do you mind dropping off so we can talk about this?” because it’s something unrelated to that person. I’ve had a board meeting where someone had a conflict and we were going to talk about an issue that is related to him, so he got off the board, but Fireflies was still on. He would receive a transcript afterwards of what we discussed. People forget; these things happen. That’s my long-winded answer as to how they ended up getting media recordings uploaded to ChatGPT.

MATT: It’s definitely easy to forget, particularly in that scenario. I can definitely see that happen. At the end of the day, I think the takeaway for this is you’re going to start seeing policies and employment agreements or employee handbooks that have very strict rules on what can and can’t be used – anything to that effect.

Again, it’s something that’s going to be very important moving forward because you don’t want that scenario that you just proposed of everybody jumps off the call and you think you’re just talking to one or multiple people, but there’s also a recording that’s going to go out to all the other people. That’s a disaster scenario at that point.

NASIR: Right. We’ve written those policies. We’re already seeing those policies and handbooks – whether it’s separate or some specific one. Disclosures, too – if you want to disclose to whatever service you’re providing that there is the use of AI and things like that. That’s something that we’ve done as well.

MATT: Yes.

NASIR: Let’s dive into the copyright aspect. Was there anything that we wanted to talk about before we switch to that?

MATT: I think we’ve covered a lot of the fun items at least. Let’s jump into the copyright component.

NASIR: All right. It’s probably one of the more common questions that we get as attorneys when we talk about AI – this copyright issue. I find it fascinating.

The first thing is this concept of who owns the copyright of AI-generated content. That AI-generated content could be in the form of an image, in the form of a story that you read to your kids. Who owns that? The proponents suggest that the prompter is the one that owns that copyright.

Let’s back up. There are three categories of who could own the copyright of the work. First, nobody – in other words, it is not copyrightable. No one owns the copyright of that work.

MATT: That’s the first?

NASIR: Well, what are the options?

MATT: I would have made that third. That’s fine.

NASIR: I probably made that backwards. That’s the first thing that I thought of. By the way, I’m definitely going backwards. Let me start over because I was already thinking back.

The first is that the person who put the prompt in and asked to create this content owns the work. The second is the party that actually created the tools to create the work owns the copyright. By the way, I lied. There are three – not four. Third, it is possible for two parties to own the copyright of one thing. They have joint ownership. Fourth – which was number one but now is number four – no one owns it.

MATT: We call that a lose-lose.

NASIR: Lose-lose situation.

Interesting enough, the most recent case under the Copyright Office that it took, they actually took the opinion of the lose-lose situation in the sense that they felt there wasn’t enough human touch or human direction given to it in order to establish this. One of the ways that I explain this to people is the analogy – and I believe proponents actually use the same analogy – the concept of someone taking a picture with a camera.

When I take my phone or a camera, click the button, and an image, it is well-established copyright law that I own the copyright rights – that’s redundant – to that image. I can reproduce it and so forth. If anyone has ever gotten photos for themselves and paid a photographer to do that, everyone knows that you actually have to get copyright assignment in order to be able to use those photos otherwise.

Just because they are pictures of you doesn’t mean that you own the copyright. The photographer owns the copyright. No one’s going to argue that Sony, the manufacturer of the camera; or Samsung, the manufacturer of your phone; or Apple owns the copyright images. No one would make that suggestion. That’s why I think that, when it comes to the creators of the AI model, it probably doesn’t make sense.

MATT: Yes.

NASIR: You would agree?

MATT: Yes. When I was reading about this, the two previous instances that came to mind – was it at the Emmys? I’m trying to remember. That’s not exactly the same, but remember somebody took a selfie, and that was a whole ordeal.

NASIR: That’s right, yes. It was at the Emmys, yes.

MATT: I’m not good with that award stuff. The one that’s more applicable, do you remember when the monkey took that photo?

NASIR: Right.

MATT: It’s not a human. A monkey is obviously not a machine either, but I think that’s the best analogy for this – at least at this point. Who knows what’s going to happen down the road?

If it’s not a human and it’s a machine, only humans can have copyrights. That’s what they’re going with here.

NASIR: The main question for the Copyright Office and any future court that is going to actually challenge this, but one of the issues was “how much human input was there?” When it comes to clicking a camera, it’s pretty predictable what is going to be produced once you click that camera. It requires some kind of human process to do that. Not to say that it doesn’t require some intelligence to be able to put the correct prompt.

We’ve already talked about how phrasing your prompting correctly can produce hallucinations and put results that don’t make any sense. But I think the language that the US Copyright Office argued that not the human is the true author of these cases, stating that the AI systems operate in an unpredictable way rather than being guided by tools. I think that’s the distinction that they made.

MATT: Yes, the line I had pulled was it’s in line with previous decisions that require a human author. Like I said, they’re looking at it as a machine, and to what degree it was used. This is obviously not a fully resolved ruling on what’s going to happen at the end of the day. It can obviously evolve over time, too. But at least that’s where it stands today – October 2023.

NASIR: Yes, and I have mixed feelings about it.

By the way, the whole story of the monkey selfie thing, I think some monkey got a hold of someone’s camera, messed with it, and ended up taking a picture of itself. The big question is who owns this photo? I think the end result was the same result in this case which was basically pushed out to the public domain – meaning no one owned the actual photo.

MATT: I want to say we even talked about this on a podcast however many years ago.

NASIR: I know we at least wrote an article about it or something.

To me, for those who have really spent a lot of time working with AI and actually creating content, one of the ways that I have used it personally is I need to write a summary of these issues. I put in ten bullet points that I want to basically write in a narrative format. That’s quite a bit of human input that I’m putting in.

In other words, it’s the difference between going back to the child story analogy. If I say, “Write me a children’s story about a princess in a castle and make sure there’s a unicorn. Make it five minutes long.” It produces a story. I can see that the production of that result is in the public domain. But if I give ten points – I want the character’s name to be this and this, the age is this; I want it to start out in this setting – and basically outline an entire plotline to a story, I want to own that copyright. I think that that is something a little bit different than what we are dealing with here.

MATT: I would agree with that. Going back to what they said, it’s the level of human input. If you are basically telling them everything you want to have in good detail, and they are just compiling it together versus you saying, “Write me a story,” there is a big difference between those two at the end of the day, particularly from the copyright perspective.

NASIR: Of course, it produces a result. You start tailoring it like, “I don’t like that person’s name. Change this section,” and you start making changes. What’s interesting is, if you took that same selfie of that monkey and started modifying it, creating a derivative work from a public domain, then you can actually own the copyright of that derivative work. That just blew my mind. That’s crazy, right?

MATT: Yes. Along those lines, the copyright is an original work of authorship. Obviously, the more that you keep feeding into it or finetuning it or adding more, I think you are definitely right. It becomes more and more likely that it’s going to look like it’s your actual ownership of those rights.

NASIR: Right. It’s curious. I feel like that would be one main way to get around this whole copyright thing – basically, creating a derivative work from the original.

The main case that it was an issue was an image or painting of some sort. If you produce the image and didn’t share that image with anybody else – no one has a copy of it – then alter it and create an original piece of work from that content, I think you could probably get that copyright.

MATT: I agree. We’re going to see. There’s going to be a lot of examples in the next few years. It’s not going to be a black and white thing.

NASIR: No, it will not be.

Something that I wanted to talk about is more about data privacy and leakage. We talked about the Samsung case, but something disturbing has been happening. In particular, there have been some data leaks on multiple cases.

Let’s go back to ChatGPT. They actually updated the terms of service later on. They were actually using the data that you were submitting to train their model. They don’t do that anymore. They’ve also restricted access. But I remember when Google Bard came out. That’s Google’s version of AI. One of the things they did right away was tap it into their search engines which, of course, would be very helpful.

But one of the things that it said when you were signing up in the terms and conditions, Google tends to be pretty straightforward in disclosing and using plain language. I think ChatGPT is there now; it wasn’t at first. But it basically said that there may be human beings that are reading your prompts for purposes of making it better. That definitely gave me pause before opting in which I have totally done, of course, because everyone knows everything about me anyway.

They did the right thing – first, by disclosing it, and not only disclosing it in the terms of service, but actually putting it out there so that people are clear. This is a good lesson for anybody writing a privacy policy or terms and conditions. Especially if you are building an app that is tapping into ChatGPT’s API and things like that, understanding how to pass along those terms and conditions and privacy policies that match what you are agreeing to with these different models.

But one of the most scary things that I think is out there that I think everyone should be concerned about is these AI models are using data that is not readily accessible to the world. I keep switching back and forth to copyright, but one of the big lawsuits going on right now is Sarah Silverman believes that they have access to her entire book. I guess she wrote a book, and they are able to summarize it, and they can only summarize it if they had access to the entire copy of it. By doing that, is she infringing? Let me get back to that in a second.

There is a suggestion that a lot of these AI models are using data that can only be found in the dark web, for example, or these massive databases that are not easily accessible by most people. Of course, the dark web contains a lot of data that has been hacked and then released out there. If you think about it, if I went on GPT and asked the question, “What is Elon Musk’s social security number?” I guarantee you, somewhere on the internet, his social security number is out there. If the AI model has used or accessed that data, in theory, it could produce his social security number.

ChatGPT and others have already put in protections to disallow you to do that, but what if those protections weren’t there? What if it wasn’t just Elon Musk’s social security number? What if it was your social security number? What if it was your personal address and your phone number? What if it was the marketing data that maybe leaked on the internet, and I can access what you purchased for lunch yesterday? I had a sandwich. You don’t need to know that.

MATT: You usually don’t eat lunch, I thought.

NASIR: No, actually, I did. That was a lie. I had lunch today – a poke bowl. That’s why, when I was trying to think of what I ate yesterday, I just made something up – a sandwich.

MATT: Yes, that’s standard.

NASIR: Copy it a lie, but I did have a poke bowl for lunch. It was very good.

I forgot what we were talking about – something about AI. You get my point. To me, that is the scary aspect of this whole thing. There are a lot of scary aspects of it, but it’s really hard to control, especially bad actors.

MATT: We might have mentioned this earlier, but the gravity of it – I don’t think the majority of people understand it. There is one thing where people are inputting information that they shouldn’t be inputting like we talked about with employees. Like you were mentioning, if it’s information that is on the dark web or that to your knowledge is not readily available, well, it could be, and it’s going to be a major thing. I don’t know what precautions are going to be put in place for this.

To me, the tool is always going to outpace any sort of ways to prevent it at the end of the day. That’s something to keep an eye on.

NASIR: But it gets adjusted. When social media and the internet came out, one of the big things was “are we going to be able to hold accountable internet companies from the content that other users are producing?” If we did, not only would it stifle free speech, but it would be almost impossible for these internet companies to regulate it or actually impose it without shutting it down. That’s been well-settled law even though there have been discussions of getting rid of that immunization of liability.

Similarly, if you have an open AI that generates content, can you hold that company responsible for copyright infringement? This is really fresh information, so forgive us for the details because it’s moving pretty fast. For example, Microsoft agreed to foot the bills of any of its Copilot customers who are sued for artificially generating content that might open them up to copyright infringement. I’m not even familiar with Copilot. First of all, that’s an incredible position for Microsoft.

MATT: Significant. You wouldn’t expect that.

NASIR: Yes, especially from a big company like that, but I understand the position because, from a legal perspective, it makes sense.

Going back to Sarah Silverman’s position, she is suggesting that the only way this AI model could have done what they did is if they had detailed or full access to the entire book which may have been done illegally. Then, it is able to produce this detailed synopsis of the book for others.

Now, there are two things – one is the access to the book itself. That may or may not be illegal. At the same time, you can go read a book, go copy a book, memorize it and so forth, but you are not publishing the book. It’s really that publishing that is the big issue where the real financial penalties come into play. It’s not to say that, if you copy a video game or copy a movie and watch it for yourself, you are not violating copyright law. That’s not what I’m saying when it comes to the actual real financial penalties.

Second is that it’s producing a detailed synopsis. It’s not reproducing the entire book. It’s reproducing a summary. The big proponents on the side of these AI models are suggesting that this falls well within the rules of fair use. Some of the factors of fair use is the purpose of it, the nature, the copyrighted work, the portion used. In theory, if most generated AI models are doing their job and creating a summary or creating a work that is derived from something else, it’s going to be incredibly unique.

Of course, the Sarah Silvermans of the world or others – and there are multiple lawsuits; she was one of the first main ones – the question is “are these actually derivative works?” Going back to our monkey photo where you’re going from a public domain, you modify that and create a derivative work, that’s one thing. But if you take from someone else’s work and then build a derivative work, then that’s not the same. I think that that’s going to be the argument from people like Sarah Silverman and things like that.

MATT: I agree. There definitely is a clear difference between those two ideas, but we’ll have to see how this all plays out. I think hers was even a class action lawsuit. I don’t know if it was only comedians or writers or however she wants to be classified, but I can imagine she got others to go along, too. That’ll be the interesting thing.

We talked about it at the forefront, but if you want to be a standup comedian, you could say, “I want to make a joke with this stuff,” and it’ll probably produce something that at least would get a laugh. This was her book, so it’s different, but just thinking along those lines and what’s going to change over time.

NASIR: There’s one more category that I think is just frankly fascinating. I have tried this myself. I have personal experience – again, going back to reading to my children. Going back to the stories, I produce a story, but then I ask it to be written in the style of Dr. Seuss. It is amazing. Obviously, Dr. Seuss has a very specific writing style. Of course, any human being can replicate that, but ChatGPGT does a phenomenal job in doing so.

They have books written by Dr. Seuss but has a Dr. Seuss publishing company, and the style is Dr. Seuss, but it wasn’t written by Dr. Seuss. They already have that. Similarly, there was a song – and this became a hugely popular song – it was an AI-generated song called “Heart on My Sleeve” which basically imitated Drake and The Weekend together. It was actually submitted for consideration for the Grammy Awards. Of course, the question is (1) “who owns the copyright of that?” and (2) “is that allowed?” I’m sure Drake and The Weekend are not too thrilled that basically a computer can replace them.

I haven’t listened to it myself, but I have heard people actually thought it sounded pretty good.

MATT: Yes, I remember a couple of things from that. As you can imagine, a lot of artists were not happy with that. I forget which one it was, but there was one artist in particular that encouraged it and said, “Make these AI generations for me with my voice. If it’s good, I’ll put it out.” I’m trying to remember who it was.

NASIR: I have got to listen to this. Have you heard this before yet – Heart on My Sleeve?

MATT: I might have when it came out.

NASIR: I’m listening to it now. It’s catchy. It has a very catchy beat to it. It sounds just like Drake. It’s pretty crazy. Here’s another one. Hey There Delilah by Kanye West.

MATT: I did listen to that one because I was just curious. It sounds exactly like him. It’s eerie.

NASIR: Let me listen to this one. He’s a bad singer.

MATT: His voice is good in it, I thought.

NASIR: It sounds like him.

MATT: I think his voice is better than most.

NASIR: I don’t think his voice is nice, but it definitely sounds like him.

MATT: Yes, I think his voice is better in this song than his songs in general.

NASIR: Than normal?

MATT: Yes.

NASIR: This one is Kill Bill by Ariana Grande. Now, she has a good voice. Let’s see. I think her voice would be a good test.

MATT: That’s probably true.

NASIR: Now, is Kill Bill an actual real song? Or is this a new song? I don’t know.

MATT: I think it is.

NASIR: You think it’s what? There’s another song called Kill Bill? Okay.

MATT: Yes, it says SZA. That’s the original artist. Vanessa is going to have to help us out on this one.

NASIR: Yes, I wish I knew the original song. I could judge it better.

MATT: Going back to what I was saying, to bring it back full circle, the artist I was thinking of is Grimes. I don’t know if she’s married to Elon Musk, but she’s the mother of their child. She was promoting it and said she would split royalties 50 percent on any successful AI-generated song that uses her voice. Some musicians are embracing it. Why not? If it sounds okay. Illegally Sound Smart Business.

NASIR: We joked at the beginning of the podcast, but if we can get to a point where this whole podcast is AI-generated, then instead of someone listening to it, there could be a summary of the podcast itself that’s generated by AI or some AI reads the summary for them, then it’s like a computer is talking to a computer. That’s going to be awesome.