How OpenAI Built a Billion-Dollar Ad Engine

[00:00:03] Welcome to WSJ Tech News Briefing. It's Friday, September 18th. I'm Julie Chang for The Wall Street Journal. AI voice cloning tools have gotten so good that they can now replicate accents, where unpacking why your accent might be a security vulnerability. Then, OpenAI CEO Sam Altman once called the combination of ads and AI, quote, uniquely unsettling.

[00:00:28] Now, the tech giant is pushing full speed ahead into the advertising world. We'll dive into the company's ad rush. Up first, say you get a call out of the blue from someone claiming to be from your bank. You might be skeptical, but what if the caller sounded just like you? New research shows that generative AI tools can now mimic distinct regional accents,

[00:00:55] using as little as 10 seconds of audio scraped from social media, podcasts, and other online sources. And apparently, that subtle sense of familiarity is often all it takes to lower our defenses. Journal contributor Heidi Mitchell sat down with us to explain why that is and how you can protect yourself from scams. Heidi, tell us more about this new research on how good people are at differentiating between AI and humans.

[00:01:24] So, when we hear something that sounds familiar, we just assume that it's real, especially if we have a super specific accent. I spoke to this guy in Scotland. He had 150 people participate in this study, and he told them, you're going to listen to these sentences and tell me if they're AI or if they're human. And they thought, overwhelmingly, they really thought that it was human when it sounded Scottish.

[00:01:48] And then when it sounded Dundonian Scottish, which is practically like another language, they really thought it was a human. Because they thought, how in the world could someone, I mean, 30% of the city of Dundee speaks this dialect. Even when he gave them warnings, they were believing that the Dundonian accent was trustworthy. Your story also notes that people can be more easily tricked into trusting a voice that sounds like their own accent. Can you tell us more about that?

[00:02:17] So, an interesting other piece of research that I stumbled upon in working on this piece is this researcher. She looked at 14 episodes of Shark Tank, and she analyzed the vocal expression of each of those pitchers that are pitching to those sharks. And she found that the closer that the entrepreneur sounded to one of the panelists, the more likely that panelist is going to make a deal.

[00:02:44] So, it's like, you sound like me. I trust you. I'm going to give you my money. So, then, marrying these two studies together, what does it mean that we're falling for these AI voices? I mean, unfortunately, I think it means we're at the risk of falling for scams more. They can just scale these things up and test them really quickly. Basically, remembering, though, that scammers don't want to spend a lot of money on any one person. So, they're just going to try and blanket as many people as they can.

[00:03:11] But basically, I think what it's going to come down to, according to the cyber experts, is that we're going to have to really think about the story that they're telling rather than the voice they're telling it in. Because they'll be able to sound just like your sister, just like your aunt, just like your daughter. And we are going to need to think, oh, does that make sense, what they're asking me to do? Is that something that she would say?

[00:03:33] And maybe just hang up and call that person back or text somebody else on a different channel so that you're not following the directions that they ask you to. Because also, when somebody calls you with a scam, and it sounds like your mom, which is a pretty common scam these days, your fear kicks in, and so you don't think so critically. You're more likely to react. So, we always want to tell people to slow down, think through this is a legitimate story that they're telling me. And it's challenging.

[00:04:03] That was WSJ contributor Heidi Mitchell. Have you ever come across an AI voice scam? If you're a listener on Spotify, let us know in the comments. Coming up, OpenAI has been making a massive ad push ahead of its widely anticipated public debut. But how is it going for the company? And for advertisers? More after the break.

[00:04:36] OpenAI CEO Sam Altman wasn't always sold on the idea of putting ads into ChatGPT. This is what he had to say about it a couple of years back, speaking at Harvard Business School. I kind of think of ads as like a last resort for us for a business model. But today, he's changed his tune. OpenAI has been rolling out ads on its flagship product, ChatGPT, since February. And it's been building out this new part of its business at a dizzying speed.

[00:05:05] WSJ Advertising Editor Suzanne Vernitsa joins us now to unpack OpenAI's ad rollout, its growing pains, and what it means for the tech giant's hotly anticipated IPO. Suzanne, can you start by telling us what OpenAI's ad business looks like? It depends on the day. The ad business is changing so quickly. We haven't really seen anything or any build on an ad platform this quick.

[00:05:31] The business right now, the way it started off was very basic, and that was just seven months ago in February. And the way it works is when you're using ChatGPT, after you have a conversation, at the very bottom, you'll see a box that has ad attached to it. And it's pretty much a plain text box with a picture. They've introduced a carousel where you might see multiple pictures and different products from the same company in the one ad. But that's basically what this thing is.

[00:06:01] And they've been rolling it out. They're now in 50 countries. And every day, they update their systems. They're adding new tools. So it's maturing pretty quickly. And they're working at sort of breakneck speed to get this done. So just to dive into the tech behind it all for a bit, is OpenAI just utilizing browser cookies to sort of determine which ads to serve to each individual user?

[00:06:28] So no, they're not really using browser cookies to determine what ads to show. The way it works is the ads are served based on the intent or basically the signals that the system is getting from the types of questions and conversation that person is having with ChatGPT. So for an example, if I'm on ChatGPT and I'm looking for the best places to visit to find

[00:06:52] the most pristine beaches per se, OpenAI's technology looks at all the signals within my conversation. And then it has a raft of advertisers that are in the system. And they insert keywords and phrases and the technology picks out which ad to put at the bottom of this conversation that is the most relevant to the subject matter that the person's discussing.

[00:07:18] So obviously, if I'm in there talking about beaches, you might have the system serve you an ad from an Expedia or a hotel that's on Aruba's beaches or something similar. Suzanne, you've covered the ad industry for decades now. Is there anything that you noticed is unique about the way OpenAI is using ads? What's very interesting about when they launched is they hired a ton of meta executives, both

[00:07:46] on the tech side of their ad business and their sales side. And they're following a playbook. They're building the meta system and the Google system. And you can see it. And what's interesting is all the back technology seems to be the same. But what's happened even just this week, they actually are coming out with new ad opportunities, which are definitely outside of what we've seen before.

[00:08:10] So they began testing a couple of weeks ago ads that felt very native to the platform. And these are called sponsored agents. So the ad appears, as I described at the bottom, and when the person clicks on it, they're not actually sent to a company's website. What happens is another chat GPT conversation pops up, and this is a branded one. So this allows me as the user to have a conversation with an agent that's been fed all the information

[00:08:39] about the advertiser so that I can query about a product. And how well would you say OpenAI's advertisement business is doing so far? Well, it just hit an annualized run rate of a billion dollars a couple of weeks ago in August, which doesn't mean there's a billion dollars in the bank by any stretch. It means if they continue at this pace, they're going to hit that billion dollar figure in a year. Now, that's probably faster than most ad businesses we've ever seen being created.

[00:09:06] And as much as they're celebrating over there, it is a little shy of what others had expected. Some analysts thought it would be higher. But it's still early days, and advertisers are not thrilled yet because the ads where they sit are very low in the page and feel very disjointed from the conversation at the top. But it's not unheard of that there's a slow build here and that they're having sort of mixed results.

[00:09:31] Facebook in the early days, right before their IPO, had many advertisers that actually thought Facebook advertising didn't work. Some big advertisers like GM pulled out weeks before their IPO. And look at that ad business today. It's second largest in the world. Speaking of IPOs, OpenAI has a much-anticipated public debut, now expected next year. How important is its ad business for that IPO?

[00:09:59] What's at stake if OpenAI doesn't get its ad business right? The stakes are huge. It needs money to offset the cost of its free version of the service. They now have over a billion weekly users. And so taking advantage of this consumer popularity they have across the board, advertising is the one area that they can do that with. That was WSJ Advertising Editor Suzanne Vernitsa. And that's it for Tech News Briefing. If you're a listener on Spotify, be sure to leave us a comment.

[00:10:28] Today's show was produced by me, Julie Chang. Jessica Fenton and Michael LaValle wrote our theme music. Our supervising producer is Katie Ferguson. Our development producer is Aisha Al-Muslim. Chris Zinsley is the Deputy Editor. Lital Mollad is our Senior Director of Shows. And Samantha Hennig is the Wall Street Journal's Head of Multimedia. We'll be back later this morning with TNB Tech Minute. Thanks for listening.