The Daily: Problems with AI, part 1—Treating AI like a human, how smart is ChatGPT, and what happened to IBM's Watson?

On today's episode, we discuss whether we'll be able to tell if something is AI-generated, whether AI can be considered an "inventor," and whether now is the right time to adopt this technology. "In Other News," we talk about how smart ChatGPT really is and what ever happened to IBM's AI supercomputer Watson after it won at "Jeopardy!" Tune in to the discussion with our analysts Jacob Bourne and Gadjo Sevilla.

Subscribe to the “Behind the Numbers” podcast on Apple Podcasts, Spotify, Pandora, Stitcher, Podbean, or wherever you listen to podcasts. Follow us on Instagram.

Episode Transcript:

Marcus Johnson:

Insider Intelligence subscribers can now take advantage of the exclusive opportunity to meet directly with their team of thought leaders in the financial services space, through the new analyst access program. Expect to leave equipped with what you need to drive profitable sales, legitimized strategic ideas, and develop models of addressable markets and opportunities. Visit insiderintelligence.com/analystaccess to learn more.

Gadjo Sevilla:

And I think it's going to be something we're going to hear more about just because this output from generative AI models, certainly companies can put in safeguards. But it's the creativity of the systems that is really hard to control and fully prevent this from happening.

Marcus Johnson:

Hey gang, it's Monday, June 5th. Jacob, Gadjo and listeners, welcome to the Behind the Numbers Daily, an eMarketer podcast. I'm Marcus. Today I'm joined by one of our senior analysts on the connectivity and tech briefing based out of New York. It's Gadjo Sevilla.

Gadjo Sevilla:

Hey, Marcus. Hey, Jacob. Happy to be here.

Marcus Johnson:

Hey fella. Thanks for hanging around today. We're also joined by someone else on the connectivity and tech briefing. One of our analysts based out of California. It's Jacob Bourne.

Jacob Bourne:

Hey, Marcus. Hey, Gadjo. Thanks for having me.

Marcus Johnson:

Hey, chap. So today's fact is about our aging population on the planet. And today, one in 10 people on the planet are over 65 years old. So if you lined up 10 people, one of them will be over 65 years old. In 30 years, one in every six will be over 65 years old. And in 80 years, so the year 2100, I think we'll probably call it, if we make it there, in 80 years, one in four people will be over 65 years old. We counted every fourth person will be over 65. Japan and Italy are already there. They have some of the highest shares of over 65s at close to one in four. So they're already at that reality. The US is at the reality of 2050 for the world. They're already there. The one in six scenario, the US is already there. The lowest shares are in countries like Uganda and Zambia, with one in 50 people over the age of 65. This is just huge discrepancies.

Jacob Bourne:

Oh, pretty soon we're we're going to be adding robots to that demographic ratio.

Marcus Johnson:

That's true. Thank goodness. Kidding. That sounds terrifying and dreadful, and we'll talk all about that in today's episode. Although I will say a world with a smaller share of young people, sounds rather nice. Anyway, today's real topic, the problems with artificial intelligence, part one.

In today's episode, first in the lead, we'll cover some of the problems with AI. Then for another news, we'll discuss how smart ChatGPT is and whether it will fade into the background like IBM's Watson. So bunch of folks working on AI just said that AI could lead to the end of the world. Great. So we decided to spend the next two episodes today and tomorrow looking at the most pressing problems with AI. Today, we'll start with some of the less apocalyptic issues, and then tomorrow we'll tackle some of the heavier concerns around global extinction. So strap in for that one. And we'll see if also we can regulate ourselves away from disaster hopefully.

But today we're talking about some of the lighter issues, but still relatively significant concerns around AI and also generative AI. And we're going to start gents with labeling things. So Insider Intelligence briefings analyst, Daniel Konstantinovic, notes that Google will start to label AI generated images in search results. So you search for something, result pops up, and all the ones that are AI generated will have a label on them. Should we/will we start to label all things that contain AI, as opposed to this just being a Google initiative?

Jacob Bourne:

Yeah, I don't know if I would go as far as to say all things, but we certainly are going to be seeing more labeling happening. And beyond just search results, Google and Microsoft are both starting to label AI generated images with metadata. So that's embedded, set in as an invisible watermark, to help differentiate. And there's good sense for that too. Just last week, we saw a low quality deepfake cause panic on Wall Street. It was basically deepfake of an explosion at the Pentagon, caused a selloff for several minutes until people scrambled to determine that it was a deepfake. Well, as AI gets better, it's going to be harder. And so we really need this invisible labeling.

Now, for the more overt labeling in search results and in potential advertising or media, that's just a consumer facing thing. The question is do consumers have the right to know if they're consuming AI generated content? And we'll probably see bodies like the EU require a certain amount of that as they work on their AI act that's slated to pass by the end of the year. And of course in the US, that's something the FTC might pay attention to as well.

Marcus Johnson:

So you're talking about labels in the background, but we're also going to have labels in the foreground as well.

Jacob Bourne:

Both.

Marcus Johnson:

So people can identify ... Both. Got it. Okay, because we were talking about this on last week's show, weekly listen on Friday, about AI generated ads and this idea of should you be told if it's an AI generated ad the same way should you be told if it's an AI generated song or piece of artwork? And it just feels better to have that level of transparency. But also there's the concern around trust and feeling deceived and how much do you trust people versus AI? And Daniel in his piece was noting the confusion around what images are real or AI generated could create a crisis of faith in advertising and platforms in general, and cited some numbers from CX platform DISQO. 32% of US adults trust AI generated content less than human content, versus 8%. So way less who trust it more.

Jacob Bourne:

Yeah, the other trust aspect to this is just if people get desensitized to seeing deep fakes everywhere and not knowing whether something is real or fake, then that could also erode trust in our social systems more broadly.

Gadjo Sevilla:

Yeah, I think they should be nutrition labels. Basically not just saying that something was created using AI, but being more specific down to the percent. 25% AI created, with a link maybe to more details of how and when AI was used in the process. Just for transparency's sake.

Marcus Johnson:

Yeah. Speaking of nutrition labels, can we put the sell by date or the use by date next to the barcode? It's so hard to find on every product. It's so annoying.

Jacob Bourne:

Yes.

Marcus Johnson:

If we could just standardize that, I'd appreciate that. Thanks, FDA. Get on that. But one of the examples of people being deceived, there was a winner of the major photography award who refused their prize after revealing that their work was created using AI. So German artist Boris Eldagsen entered the piece into the competition, because he wanted to start a conversation about AI in art, saying it's important to differentiate AI generated art. Let's talk about whether AI can be considered as an inventor.

Gadjo Sevilla:

I think right now AI is a tool. So nothing original has come from AI without programming or using prompts. So it's basically a DJ that's using music samples to create a backbeat. But the samples come from original creators, musicians, artists, those are the inventors. Even if you randomize a graphic or a poem using AI, it is all really derivative right now.

Marcus Johnson:

Yeah, that's a good analogy. Ashley Gold at Axios was writing that Google is arguing that AI technology should not be considered an inventor by US patent law and a new filing with the US patent trademark office. Jacob, what do you think? Because if you are using the technology or if you say a few words into it and it spits out a product and you take credit for that, you've done basically nothing. But you can't give the credit to AI because it's not a human.

Jacob Bourne:

Yeah, I think here there's a difference between the agency of the AI and the consciousness. So even if AI is creating things that didn't exist before, it doesn't have conscious self-awareness, which is part of what we associate with humans being people, being inventors, being creators. And AI just doesn't have that. Now, that's open question as to whether it might have it sometime in the future. Right now, there's no evidence that even the most advanced models have any kind of self-awareness. The jury is out on whether or not that's possible. But it's a difficult question because at this point. We don't even know fully know how human consciousness works, and we don't really even understand how these AI models work. But I think if in the future, if there is that kind of evidence, then I think that question will would be revisited, and we might be calling AI as inventors. But at this point, we're a long way off.

Marcus Johnson:

So to that point, Axios was explaining these different terms, these key terms in the realm of AI. And so you've got generative AI. You've got which is ChatGPT, GPT-4, artificial general intelligence, so AGI, and then sentient AI. And so to explain what those are, generative AI, they say in this piece sounds like a person. AGI, artificial general intelligence, reasons like a person. And sentient AI thinks like a person. And so you're saying we'd have to get to that before we could start saying, "Okay, you are an inventor now."

Jacob Bourne:

Yeah, I think it's more than just thinks. I think it's that sentient AI is self-aware of itself. So, some people even think that GPT-4 already thinks like a human on a certain level. But it's certainly not conscious.

Marcus Johnson:

Yeah. So speaking about inventing things, something goes along with that is the idea of copyright. And so what does a pro copyright generative AI model look like? Well, ChatGPT creator Open AI is working on one. They're working on a pro copyright generative AI model. Its founder positioning himself on the side of copyright systems that ensure creators are paid for the value that they create. He was quoted saying, "We're trying to work on new models where if an AI system is using your content or if it's using your style, you get paid for that." Is this possible? It sounds incredibly difficult.

Gadjo Sevilla:

I think depending on the scale of, it is possible to implement some sort of track changes feature to certain aspects of the AI, to see where the idea came from. And if you're just working within say ChatGPT, I think that's doable. But once you start jumping across different generative solutions, it's going to be harder to trace that.

Marcus Johnson:

Yeah, it sounds like they're trying to get ahead of legislation. Sam Schechner of the Wall Street Journal noting the same. I think it's the same legislation that you were talking about, Jacob, the proposed European Union legislation,

Jacob Bourne:

The AI Act.

Marcus Johnson:

Which would require developers exactly to list copyright materials used in generative AI tools.

Jacob Bourne:

Yeah, it's a bit of a no-brainer, I think. And really it's about the legal system catching up and AI companies catching up too. So far they've been able to largely take advantage of the fact that there's a gaping hole in the legal system. But these companies are planning on making billions of dollars scraping copyrighted data off of the internet. And so there's definitely an argument that this is infringement. It certainly can be an argue that it's unfair. And so I think it's just about creating models that are able to differentiate between copyrighted versus non-copyrighted right data, so that artists can get compensated so that attribution can be offered as well.

Marcus Johnson:

So real quick, let me throw this at you before we move on to the next question. A new and very good friend of mine who works at the University of Chicago, was telling me about a conversation she was having around ChatGPT and plagiarism. And speaking to a student and they were saying ... And as well as a colleague, and they were saying, "Is it plagiarism if it's pulling from the entire internet as opposed to copying a single source verbatim?" I thought that was an interesting thought.

Jacob Bourne:

I think overall ChatGPT is this lowered the bar for what's considered plagiarism. And so it's something that's being reassessed by the education system in other areas of society.

Marcus Johnson:

So what happens if ChatGPT does get something wrong and someone decides to sue it, to sue AI? So back in April, a mayor in Australia called Brian Hood said he may take legal action over false information shared by advanced chat bot ChatGPT, notes Tom Gerken of the BBC. This comes after the Open AI own tool, falsely claimed he was imprisoned for bribery while working for a subsidiary of Australia's national bank. His lawyers have sent a concerns notice to Open AI. I wasn't able to dig up any more developments on where we are today. But what happens if someone sues artificial intelligence?

Jacob Bourne:

Well, just like AI is not an inventor, it also can't get sued because it doesn't have that agency. The company who made it or deployed it commercially can get sued though, and as is the case with Open AI. And I think it's going to be something we're going to hear more about just because this output from generative AI models, certainly companies can put in safeguards. But it's the creativity of the systems that is really hard to control and fully prevent this from happening. And of course, people in organizations don't want falsehoods spread about them. That's not going to change. So I think this goes back to the legal system really needing to catch up. In the US, we have Section 230 protecting tech companies.

Marcus Johnson:

Right. I was going to talk about that.

Jacob Bourne:

Yeah. Protects tech companies from liability for third party content on their platforms. That's outdated even for that purpose. It's certainly not really well suited to this new generative AI landscape. So it's something, again, one of the many things that are, globally, the legal systems need to catch up on.

Marcus Johnson:

Yeah. Gadjo, what do you think? Because I was going to go there. So Section 230 and internet law, which protects platforms like Facebook for example, and says, "Look, if someone posts something on our platform, we are not liable for that." And it's been around since 1996, part of the Communications Decency Act. But Mr. Gerken and the BBC was noting that when folks use ChatGPT, they're shown a disclaimer warning that the content it generates may contain inaccurate information about people, places, or facts. Is that enough?

Gadjo Sevilla:

For now, I think it's understood that products like ChatGPT are out in beta. It's an open sandbox. So while everyone can use them, I think it's understood that there's no accountability or liability as to what the results could come out because there's no way they can really control. No one put that information in there, false or not. So yeah. Again, I think it's going to be a legal hurdle to figure out how to regulate that or understand what the parameters would be. But at this point, it's going to be hard to pin down an AI company for misappropriating information of individuals.

Marcus Johnson:

Yeah. So as we've outlined, and Ben Sherry of Inc. writes, "AI and generative AI pose risks for early adopters." We've listed some of the major ones, but there are many more. It is now actually the right time to adopt this technology.

Jacob Bourne:

Yeah, I think it's not even a question of when. Pretty much all organizations should at least be looking at it. But adoption is really something that should be explored in terms of companies looking at what are the risks and weighing them against what the benefits are for their specific use cases. And I think definitely, we know that the risks are considerable. And it's about building up that internal knowledge and capabilities to effectively incorporate it into an organization. And I can't imagine there's going to be a lot of companies that are really going to want to say, "No, we don't want to adopt it at all." It's about that pace of the adoption and having that really good strategy for adoption. But I think the time is now for pretty much everybody.

Gadjo Sevilla:

Yeah, I agree. There is wisdom in being more conservative and judicious with its use. We've already seen the folly of companies have blindly made it the end all be all and their business. We've already seen that blow back. So it's still early, still have a long way to go. So I really think taking very cautious steps towards adoption would be the best way forward.

Marcus Johnson:

Well, that's it for the lead, folks. Time for the halftime report. So Gadjo, we'll start with you. What's worth repeating from the first half?

Gadjo Sevilla:

I think the idea-

Marcus Johnson:

All of it. It was a brilliant episode.

Gadjo Sevilla:

It was brilliant.

Marcus Johnson:

We can't do everything, Gadjo. We can't do everything.

Gadjo Sevilla:

Sort of the idea of attributing copyright to AI, especially as it's now being used as a creative tool. So how that is built into the system, if at all, and whether they can standardize those watermarks or that attribution. I think that will be a whole different part of AI, and how people adopt to that will be interesting to see. It's unprecedented at this point.

Marcus Johnson:

Jacob?

Jacob Bourne:

Yeah, I think the big takeaway from all this is that this is a technology that has big rewards and big risks as well. And what we're seeing is that the pace at which this commercial generative AI explosion is going is just far outpacing the pace that our legal systems are moving to try and catch up. And that's really the big concern about this, is just a mismatch between the advancement of the technology and then the legal system that can regulate it.

Marcus Johnson:

Yeah. Time for the second half of the show. Today, in other news: How smart is ChatGPT? And will America forget about ChatGPT the same way we forgot about IBM's Watson?

Story one, "How smart is ChatGPT?", asks Marcus Lu of Visual Capitalist. He writes thar Open AI's latest large language model, GPT-4, which is the next iteration from ChatGPT is capable of human level performance in many professional and academic exams. It did best in the verbal GRE, which is a standardized test that gets you into graduate schools. It scored in the 99th percentile, so only 1% of scores did better than it. It scored in the 93rd percentile for the reading and writing SAT and the 90th for the uniform bar that lawyers take. It's scored worst in coding and English literature and language, down in the 15th percentile or worse for each. But Jacob, the most interesting sentence in this article from Mr. Lu, Visual Capitalist, about how smart ChatGPT is, is what and why?

Jacob Bourne:

Yeah, the most interesting thing here is just how poorly GPT-4 did on the AP English exam in particular because we know that it has a high verbal IQ. And the reason why here is probably because it didn't understand the questions. And that's important because it really underscores how important prompt engineering is still is for this technology in terms of getting the best output, and how far it is from being fully democratized. In other words, not anybody can just go and log into to ChatGPT and get the same quality of output. There's really still some expertise in terms of crafting the best prompts. And I think that's something that we're going to see these AI companies really focus on in terms of making models that can really work with more flexible prompts, and also even what we see with Google Sidekick, even suggesting prompts to people to really try and level that playing field. Yeah.

Marcus Johnson:

Two quick ones from me. One, the leaps it made from 3.5. So ChatGPT to GPT-4, which came out about six months apart was staggering. So in the verbal GRE, it went from the 60th to the 90th percentile in that stretch. The LSAT, the law score administration test and advanced placement, AP statistics, both went from around the 40th percentile to the 90th. And then the second point here is just how high it was across multiple tests. For multiple tests, it was in the top 80% of scores, for the verbal and quantitative GRE, for reading and writing and maths SAT, for the LSAT, the bar, AP biology, stats and psychology.

Jacob Bourne:

And Marcus, I think that really speaks to the pace of innovation here. And who knows what we'll see with GPT-5?

Marcus Johnson:

Yeah, terrifying to even imagine. Story two, "America forgot about IBM Watson. Is ChatGPT next?" asks Mac Schwerin of The Atlantic. He writes that AI conquered Jeopardy before it was sanded down into business tools. The same trajectory is playing out again. Today, Watson comes in the form of B2B software that IBM sells by subscription, like Watson Assistant, Watson Discovery, that help automate backend processes within customer service, document entry, et cetera. Watson tools are used by over a hundred million people, the piece notes. But Gadjo, the most interesting sentence in this article from Mr. Schwerin, about the fact that people forgot about IBM's Watson, is what and why?

Gadjo Sevilla:

Well, I think basically ... Without reading the full article, I think Watson was one of those ... It is a business solution, so it is used in super specific functions, mostly for business and enterprise. So it doesn't affect the daily lives of people. Unlike ChatGPT, which is currently open and it has touched the lives of a lot of people. Anyone can jump in and have an experience with ChatGPT. So I think through time, that's going to have a more palpable experience than just watching an AI win at Jeopardy because people have interacted with it.

Jacob Bourne:

Yeah, to what Gadjo just said. I think ChatGPT is going to be unforgettable. It's historically significant event compared to Watson, which wasn't quite that. Some people are even calling ChatGPT's release a black swan event. So could it become obsolete? Yes. Unforgettable? No, I don't think so.

Marcus Johnson:

Indeed. Well, that's all we've got time for this episode. Tune in tomorrow for the part two, of this Problems With AI two-part episode series we've got going on. Thank you so much to my guests. They'll be back tomorrow. Thank you to Gadjo.

Gadjo Sevilla:

Thank you very much.

Marcus Johnson:

Thank you to Jacob.

Jacob Bourne:

Thank you, Marcus and Gadjo.

Marcus Johnson:

And thanks to Victoria who edits the show. Welcome back. Thank goodness. James, who copyedited it. And Stuart who runs the team. Thanks to everyone listening in. We'll see you tomorrow for another episode of The Behind the Numbers Daily, an eMarketer podcast.