The humble broadcast-radio host, whether a disc jockey or interviewer or reporter, has been going through it for decades now. The 1996 Telecommunications Act fueled the consolidation of local stations, decimating their staffs. The explosion of online radio, music and video streaming, and podcasting have upended ratings for shows on public airwaves. Phones and computers and smart speakers increasingly supplant radio sets. Funding for public radio is notoriously unreliable. It isn’t the best time for your modern-day Wolfman Jacks, or for any media profession.
On top of all that, your local DJ was already on the losing end of the artificial-intelligence revolution. Before the A.I. hype from last year, and even before the COVID recession demolished media ad markets, broadcast networks were gutting on-air talent at the both the national and collegiate level to trim budgets and automate programming: syndicating well-known shows and brands, prerecording and prearranging late-night broadcasts, training a roboticized voice to fill in the space when needed. Coupled with major streaming services’ dependence on algorithms and automation to curate playlists and make user recommendations—often with bizarre side effects—these developments make clear that the music industry anticipates the need for fewer humans down the line.
A.I. hasn’t yet finished killing the radio star, nor is it truly likely to anytime soon. But there’s a new digital buddy out there that might give hosts additional pause: RadioGPT, a new tool from the Ohio-based software company Futuri Media that fully digitizes the broadcast host as you know it. According to Futuri, which has worked with large corporations like iHeartMedia and Tribune Publishing, its “new and revolutionary product” combines a few tools: TopicPulse, a Futuri app that provides an automated way to scan media sources and pull out relevant topics for coverage; GPT-3, the large language model that powers the hit chatbot ChatGPT; and A.I.-voice “personalities” made by Futuri that can learn the info scraped by TopicPulse and aggregated by GPT-3 to read readymade copy live on air. Oh, and it’s trained to know all available facts about the music played by your station, so it can even intro upcoming tracks and provide trivia as needed. The RadioGPT beta is currently being tested by large radio owners in the United States and Canada, and to gauge from preliminary reviews, it seems pretty good. Accomplished enough, at the very least, to reawaken worst-case fears regarding the future of human radio jobs, no matter what you actually make of RadioGPT’s humanoid talent.
I recently spoke with two Futuri Media executives—CEO and co-founder Daniel Anstandig, and CFO Marty Shagrin—to discuss the reasons for creating RadioGPT, the potential use cases for the tech, and how to square radio-industry automation with the Futuri team’s “passion for entertainment and pop culture,” as Anstandig characterized it. Our conversation has been edited and condensed for clarity.
Nitish Pahwa: When did you come up with the idea for RadioGPT?
Daniel Anstandig: The idea for RadioGPT started before GPT was a mainstream concept. About three years ago, we started working on an A.I. voice project to help augment on-air personalities and bring voices to air shifts for radio and TV programming that are typically unmanned—when there aren’t people in the studio and there’s no live and local content available. Along the way, there’ve been limitations around the realness versus the synthetic quality of the voices, and there’ve been challenges we’ve faced around making the on-air hosts more humanistic.
Over the last 10 years, we’ve developed automation systems that directly interface with radio and TV stations and the automation systems that run what’s happening on the air. We had TopicPulse, which scans information from Facebook, Twitter, Instagram, and over 250,000 news sources so we can see what’s trending in a local market as big as New York or as small as Medicine Hat, Alberta. We’ve also been able to integrate GPT-3, which has made our scripts more conversational. That was a big breakthrough. At the same time, we had a breakthrough around our A.I. voices becoming more humanlike and more, I would say, interesting to listen to over long-form programming.
Marty Shagrin: We’ve built our company around helping broadcasters focus on local content that is engaging, that brings people of a community together. A.I. has been in development for a long time, so itself, it’s not magical. It’s about the way that it’s used to deliver against the promise of broadcast content, whose 90 percent national reach has been built on local and live content.
Something you mentioned, Daniel, is that there’s so much radio at this point that’s been automated, especially late-night programming. So from your view, RadioGPT is a way to bring back personality to those humanless shifts?
Anstandig: I do think there’s an advantage to being more current, and we have technology that helps personalities or talent to be timely. As an example, inside TopicPulse, we use A.I. to predict what’s going to go viral in a local market. So if a station is voice-tracking—that’s the industry term for what you’re describing—or prerecording their content with a personality, they could conceivably look ahead in that market to see what’s currently trending with their audience and what’s predicted to go viral, or determine what will be interesting at that time.
One of the advantages of RadioGPT is that it knows about an artist or a song or about a current event, so it can speak to a broad range of topics concisely and in an entertaining way. When we set up RadioGPT voices, this is not text-to-speech. This is setting up character and personality in A.I., with a perspective, with character traits, with a viewpoint that is presenting a certain view or idea to the audience. That’s a tremendous upgrade from listening to prerecorded content.
It’s interesting to hear that angle, because I think when it comes to DJs or radio workers who’ve seen the industry contract—especially in terms of headcount—some might be inclined to say: “It’s A.I. that’s pushing us out of our jobs to begin with. And now there’s this new A.I. gizmo to fill in for people who’ve been displaced.” I’m curious if you’ve heard from local DJs about what they make of an advanced tool like this.
Anstandig: We have heard from a lot of air personalities as well as station owners and managers. We’ve heard from virtually every major streaming service, as well. In general, everyone is leaning in to the opportunity to dream up innovative, creative ways to augment talent, or to help talent do what they do best and create more companionship on the air. As an example, if you have a show that is carried in multiple markets, imagine that show is able to use the same station host or that same host’s voice to narrate local weather, news, traffic, or to talk about an event that’s happening, even though the show is on a national platform. That’s tremendous. Imagine that you could take a show that’s presented in English but then, in near real time, translate it into Spanish and have that read by an A.I. voice. You could also look at situations where hosts want to add RadioGPT or A.I. as a co-host and interact with it.
We have heard from personalities that are concerned that somehow there’s a mission here to replace them. But that is the furthest thing from our mindset. Local is the key to attracting audiences, and you can bet every streaming service is going to innovate in this way. We’re talking to every major service about ways to use our system. Why shouldn’t radio do the same thing? Radio shouldn’t have unmanned air shifts without live and local content.
A.I. systems—whether for voice imitation or LLMs or aggregation—run on big datasets, and there are all sorts of energy and computing costs. What’s the financial burden of making a product like this and distributing it?
Anstandig: The cost of doing this well is tremendous because you are stacking on different technologies, some that are proprietary and patented by Futuri and created over many years, along with services like GPT-3. So there is an innovation cost upfront. We think, in the long run, that for companies to do this well, there will be continued investment, but obviously technology cost scales over time.
Shagrin: We think about the financial impact all the time. But that was probably seventh or eighth on the priority list. First and foremost for us is, will the audience like this? Is this going to be an engaging solution for our partners to deliver great content to their audience? If this can help them do that, then it’s something that we owe it to them to explore.
I’ve spoken with musicians over the years who are miffed at the algorithmification of playlists, which fuels this situation where a lot of the same big artists get picked by the same big playlists over and over again. And if you’re not on there, or unless your label’s doing payola, you’re not going to gain that distribution. With more systems like this in the radiosphere—which is still important to artists—what you would say to concerns about automated systems highlighting the same artists time and again?
Anstandig: There’s this sense that technology tends to amplify what it thinks is already popular. However, that all comes down to the complexity of the algorithm and the way the algo is built to serve a certain listener. I personally have been involved in it, not only at Futuri but in the music industry; I’ve published and written hundreds of songs. So I’m always concerned about how the algos are prioritizing or exposing different songs and artists to new audiences. What I can tell you is that I believe today there’s more opportunity for independent artists and non-label acts to find an audience than ever before. I think there used to be a paradigm that familiarity was the most important thing when it came to prioritizing what shows up in a playlist or in an algo. Now, there’s more nuance around what the audience really wants: Are there certain types of music and entertainment that are more appealing when they’re new or more novel?
Shagrin: Remember, the way RadioGPT works is that it’s hooked up to and synced with a radio station’s automation system. We can offer ideas, but they’re still dictating what is played when, and what that looks like. Local markets know more about their audience, and so they still dictate that, and they can still drive that.
Local news has been decimated in all markets, whether newspapers or radio shows or TV. There are all these attempts to try to revitalize it, including yours, it sounds like, where you scan social networks and news outlets for reports you can broadcast over the radio. But there can be a lot of misinformation or misleading news that spreads quickly, and that in turn can be picked up; I’m thinking about, say, the false reports that often surround a mass shooting. I’d love to hear what sorts of guardrails you have in mind to stanch that effect.
Anstandig: It’s always best to have a human in the studio when there’s a crisis. We’ve seen that in recent years with active shooter situations, local weather disasters. That being said, there are ways to surface information faster: looking at location-based data and combining that with niche publications, blogs, or traditional news outlets for verification. In the case of TopicPulse, even though we have over a decade of experience in social media monitoring and predictions and A.I.-powered discovery of stories, we still have a small human moderation team that observes what’s trending in the system, that looks into trends as they’re emerging in markets. Because we do recognize there’s a certain amount of assurance or accountability that comes with being on this side of media.
Shagrin: The active shooter situation is near to me. I have a daughter who goes to Michigan State, where I know you went to school. So I was living that shooting situation in real time through my phone, as she was texting me and seeing how the news was covering it.
One of the things we focus on is that what’s relevant in a local market and to an audience is not always breaking. Breaking news is a bit of a different business than we are in. We are trying to identify those stories that are trending, relevant, most engaging to an audience. By doing that, we reduce the risk of reporting something that was first put out there and was incorrect. I know there were lots of incorrect stories being reported as that was going on at Michigan State.
I did have MSU on the mind when asking that—I’m really glad your daughter’s OK.
Changing the subject: Since the late ’90s, a lot of local radio stations have been consolidated or shuttered. How can A.I. revitalize local news and targeted music curation even when the corporate structures of radio markets may not be best suited for local regions?
Anstandig: First, I’ll just say that A.I. is not inherently biased or dishonest. It’s trained by humans. So when we think about RadioGPT and TopicPulse and our technology, we are training it with a media mindset. We’re ensuring that it’s trained on diverse, high-quality data sources, that the information generated can ultimately stand up to the trade winds of media, and that we don’t expect that we’re the last stop when we deliver content or information to a newsroom or to a broadcaster. That we know there are humans to verify and fact-check and think about regulation and ethical guidelines.
I think A.I. is part of the future recipe for media growth. It is certainly a way that costs can be reduced in certain parts of the industry, but it’s also an incredible growth mechanism. Information can be personalized and localized at a level that very few newsrooms have the capacity for right now. As an example, if you’re covering the Los Angeles DMA, you have such a wide range of cultures and pockets of community to cover—from Orange County to West Hollywood to Long Beach to Downtown L.A. There are all very different groups of people with differing lifestyles, and news that matters in one area doesn’t necessarily matter in another. There are ways, using stream-stitching technology and other tools, to personalize and localize news in a way that matches the interests and the needs of a hyperlocal audience. There’s data analysis that can go into the formation of that content, prioritizing which types of topics or stories should be covered by different types of people on the team. There’s a lot of efficiency built in that elevates humans working in news to do what they do best, knowing that discovery and being able to surface content and information and emerging trends is more efficient for them.
When it comes to investigative reporting, I believe that that’s boots on the ground. Humans are still going to accomplish investigative reporting better than A.I., but A.I. is a heck of an assistant.
To take it back to music, when you look back at the history of radio, there’s often some quirky DJ personality taking out obscure records for listeners. With all these A.I. systems where you can personalize things—text and images and even music—is human curation dead?
Anstandig: No. I know there’s still a tremendous passion in broadcast radio to elevate new and local artists. There’s more of that in some markets than in others because the risk tolerance of placing emerging artists that are unfamiliar or unknown is just different in different-size markets.
But I think—between social media influencers who specialize in music, bloggers, YouTubers, Instagrammers—there are more creative pathways for a new artist or track to find the right platform and audience. I think that’s only going to continue to grow.
In that vein, do you think radio stations should look into hiring A.I.-attuned people, like prompt engineers or coders or basic fact-checkers/editors?
Anstandig: I believe that radio and TV, and media in general, should do what they do best and bring the human element and local influence to what, partnering with technologists who can provide the platforms and systems that help them to elevate their creativity, intuition, expertise. We’re not attempting to replace the human element that makes the sales process or the creative marketing process special, unique, effective. We’re simply providing data-driven insights to help an account or marketing executive be more effective about creating powerful campaigns.
I think that’s a perfect example of partnering the unique spark of a human with a unique spark of A.I. And I think there will be more of a combination. I don’t think broadcast radio companies necessarily need to suddenly turn into computer or data science hives and start building the next megamachine
Shagrin: There’s nothing magical about A.I. in itself. It’s only magical if it delivers something that the people want. Media companies—most businesses, frankly—should find a way to put the technology to use as a utility to deliver the results that they want. It doesn’t mean you have to be an expert in the guts of A.I.
In your respective views, what do you think is the best use case for RadioGPT, and what is the worst?
Anstandig: I think the best case is creating more entertaining, informative, live and local content for every radio station in the world, and attracting and retaining and entertaining audiences. The worst is completely replacing humans who mean so much to their communities and to their advertisers. But I think there’s so much more upside here than there is downside for everyone, especially the audience.
Shagrin: You know, I’m the CFO. Cutting costs is not where this is the best use case. This is about delivering great content.
Anstandig: I look forward to the day that A.I. and human shows are in real, direct competition with each other. I think it’ll raise the bar for broadcasters and air talent. It’ll be entertaining, and the listeners will win.
Future Tense is a partnership of Slate, New America, and Arizona State University that examines emerging technologies, public policy, and society.
