AI voices are taking over the internet

Episode Summary

Executive Summary: The episode examines the rapid rise of AI voice technology, showing how easy it is now to create convincing voice clones for editing, narration, accessibility, and entertainment. David Pierce tests multiple tools—Podcastle, Descript, ElevenLabs, and Apple’s Personal Voice—and finds that quality varies widely, with ElevenLabs surprisingly strongest. The discussion emphasizes major ethical, legal, and identity concerns as voice AI becomes mainstream.

Main Topics: Why AI voices are booming (Priority: 5/5): The hosts frame voice AI growth as the result of rising audio consumption, better models, and easier creation tools. Podcasts, voice messages, captions, and spoken content have made audio central to internet use. Voice cloning for editing and production (Priority: 5/5): Tools like Descript and Podcastle can generate a user's voice to patch mistakes, add missing lines, or avoid rerecording. The segment shows how these tools can be useful in small doses but still sound imperfect or unnatural. AI narration and synthetic reading (Priority: 4/5): Companies like 11 Labs are being used to read articles, webpages, and books aloud. The episode highlights how dramatically better synthetic narration has become compared with earlier robotic-sounding systems. Ethical, legal, and labor concerns (Priority: 5/5): The conversation explores lawsuits, narrator fears, celebrity voice cloning, and Hollywood strike issues. It questions who owns a voice, who can train on it, and what consent should look like. Accessibility and personal communication (Priority: 5/5): Apple's Personal Voice and Live Speech are presented as transformative accessibility tools for people who can't speak or who need assistance communicating. Similar features from Samsung and screen reader improvements are also noted. Testing the technology in practice (Priority: 4/5): Pierce records his own voice across platforms and compares results. He finds Podcastle mediocre, Descript decent for filler fixes, ElevenLabs strikingly convincing, and Apple's on-device system privacy-friendly but less natural than ElevenLabs.

Key Arguments: Voice AI is no longer experimental: it is easy to use, fast to train, and good enough to become a common feature in audio products. The most compelling near-term uses are practical ones—editing, narration, and accessibility—not full replacement of human performance. Synthetic voices raise serious consent and labor questions because they can replicate a person’s identity and potentially replace narrators, actors, and speakers. The technology’s usefulness for disabled users is real and significant, even if it also creates deepfake and misuse risks. Quality matters, but usability matters too: tools that are simple and fast may spread faster than the technically best ones. As the technology improves, the line between helpful augmentation and unethical imitation becomes harder to define.

Data Points: Year mentioned for iPhone launch: 2007 - Used in an example sentence that AI voices were asked to recreate. Podcastle training prompts: 70 sentences - Pierce recorded specific prompts to create a voice clone in Podcastle. Descript training data: about four hours - He uploaded hours of his own recorded voice from Vergecast episodes to train Descript. ElevenLabs setup requirement: about 5 minutes - The service says only a few minutes of audio are needed to create a usable clone. ElevenLabs creation time: about 90 seconds - Pierce says the voice was built very quickly after uploading audio. Apple Personal Voice prompt count: 150 phrases - Apple’s on-device setup required reading many phrases aloud. Apple Personal Voice setup time: about 15 minutes - Time required to record the prompts before generation. Apple Personal Voice generation time: a couple of days - The final voice took significantly longer because processing happened on-device. LinkedIn professional network size: over 1 billion professionals - Mentioned in the ad read promoting LinkedIn Ads. LinkedIn decision makers: 130 million - Mentioned as part of LinkedIn Ads targeting claims. Zapier companies automating: 3.4 million companies - Cited in the ad read to show scale of Zapier usage. Audible Captions lawsuit year: 2019 - Publishers sued Audible over its audiobook-reading feature. Audible settlement year: 2020 - The lawsuit over Audible Captions was settled the following year. AI narration in Apple Books: a bunch of audiobooks - Describes multiple books available with Apple-generated narration.

Pivotal Quotes: "I’m kind of like David, but kind of not." — David Pierce: He demonstrates Apple’s Personal Voice and jokes about how close the clone is to his real voice. "It’s not perfect, but it works pretty well, and it kind of feels like magic to use it." — David Pierce: He describes Descript’s text-based editing and overdub workflow. "What does it mean that I can create a replica this good and that they’re only going to get better and easier over time?" — David Pierce: He summarizes the core identity and ethics problem of increasingly realistic voice cloning.

Implications: Voice cloning is moving into everyday software, making editing and accessibility easier while intensifying deepfake, consent, and labor concerns. Listeners should expect synthetic voices to become normal—and to force new rules around ownership and use.