A couple of months ago when I was approached by Sounder with the opportunity to help build AI for the world of podcasting, I was quickly hooked.
As a data scientist and technologist, I’ve closely followed the progress of language models over the past decade, including the meteoric leap of large language models (LLMs) in just the last year. ChatGPT’s arrival in particular was a transformational moment – not just for what was possible with a chatbot, but more broadly for what was possible for any data product rooted in natural language. For me, building products at the intersection of language models and podcasting was a natural blending of two things that I already savored on their own.
I’ve spent much of the last decade building data products for the world of video, leveraging textual information – titles, descriptions, transcripts, metadata – and other signals to create engaging experiences for publishers and their viewing audiences. But when it comes to language, I’d argue that no medium is as rich or ripe for innovation as the world of podcasting. It’s a world that is enormously popular, with 90 million weekly listeners in the US alone, and a medium that collectively provides an endlessly refreshing lens into the vast landscape of human thought and conversation.
That said, even with its massive popularity and growth, the concept of podcasting is barely 20 years old, and much of the ecosystem for listeners, hosts, and advertisers is still in its infancy.
Of course, it’s essential to understand that chatbots from the likes of OpenAI, Anthropic, and Meta are not end-to-end solutions for all the challenges facing the podcasting ecosystem today. While LLMs are fantastic generalists across a wide array of tasks, trained on everything from sentence completion tasks to answering multiple choice questions and arithmetic – they are not particularly expert in anything. Moreover, the companies behind these models are well aware that their LLMs carry troves of inherent biases that reflect the data they were trained on, cautioning against using them in high-stakes scenarios.
Sounder’s approach has been to embrace these language breakthroughs and combine them with domain knowledge, in-house annotation, proprietary AI, and importantly, AI guardrails, to responsibly develop and deliver tailored solutions for the world of podcasting. And as terrific as the in-house tech and team are at Sounder, it’s perhaps the team’s deep commitment to ethical AI that gives me the most confidence that our products will succeed in building trust within the ecosystem.
One recent step in that direction is Sounder’s recent research study with Urban One which shined a light on bias within the world of brand suitability, particularly in the context of Black content creators. Brand safety is not only a high-stakes endeavor for advertisers safeguarding their brands, but also for content creators who might be disproportionately labeled as unsafe, unfairly depriving them of revenue in the process. The study found that over 90% of content from Black creators was blocked by traditional keyword targeting approaches; by contrast, only 10% was flagged when using Sounder’s AI, in effect creating almost ten times more monetization opportunities by comparison.
Those steps on their own are far from the end of the story, though. For all the undeniable progress we see today in LLM technology, the AI community must continue to self-evaluate to ensure that models don’t unintentionally perpetuate invisible biases. At Sounder we believe this takes nothing less than a sustained, proactive effort to monitor and counteract biases. It’s this commitment paired with the power of language models and our talented research team that make me excited about creating transformative data products for the world of podcasting and optimistic about doing so in an ethical, equitable fashion.
