Podcasts

From PRIMO.ai
Jump to navigation Jump to search

Youtube search... ...Google search

A podcast is an episodic audio (and, increasingly, video) program distributed through an RSS feed or a platform such as Apple Podcasts, Spotify, or YouTube. For most of its history, making one meant human hosts, a microphone, an editor, and hours of post-production per episode. Generative AI has changed every step of that chain. Speech recognition turns raw recordings into editable transcripts. Text-based editors let producers cut audio by deleting words. Noise removal can make a kitchen-table recording sound like a treated studio. And since 2024, systems such as Google's NotebookLM (built on Gemini) can take a stack of documents and produce a natural-sounding, two-host conversation about them in a few minutes, with no humans speaking at all.

The key players fall into a few groups:

  • Model and research labs build the underlying speech and language models. Google DeepMind's SoundStream, AudioLM, and SoundStorm work laid much of the groundwork for long, multi-speaker speech generation. Microsoft (VibeVoice), OpenAI (Whisper speech recognition and voice generation), ElevenLabs, and academic teams (for example, the MoonCast and PodAgent projects) push long-form dialogue generation further.
  • Product teams turn those models into tools: Google Labs' NotebookLM team (led at launch by product manager Raiza Martin, with author Steven Johnson shaping its editorial approach), Descript, ElevenLabs Studio, Riverside, Podcastle, Adobe, and others.
  • Platforms decide what reaches listeners: Apple Podcasts, Spotify, YouTube, and hosting companies such as Spreaker, Acast, and RSS.com. They now set the disclosure rules for AI content.
  • Regulators are starting to require labeling of synthetic audio, most notably through Article 50 of the EU AI Act, which became applicable in August 2026.

The paradigm shift is about cost and control. A human interview show still needs people with something to say. An AI-generated show needs a source document and a few cents to a few dollars of compute. That makes it possible to give a personal audio briefing to one listener (your own research notes, read back to you as a conversation) and also to flood directories with thousands of low-cost shows. Both outcomes are now happening.


AI is reshaping podcasting at both ends at once: it makes human-made shows cheaper to produce, translate, and discover, and it makes machine-made shows almost free to publish. The open questions are whether listeners can still tell the difference, whether they care, and who is responsible for telling them.



AI/ML Podcasts

Shows come and go, rename, and pause. Entries below note changes we have verified; check each feed for current activity.

Curated lists

Technical and research-focused shows

News, business, and industry shows

AI, culture, and society

Shows made with AI (experiments and case studies)

  • Meet ‘Dudesy:’ The AI that hosts a comedy podcast with B.C.-born actor Will Sasso | Claire Wilson - BIV ... The show offers a look into how artificial intelligence will affect entertainment, merchandise and branding. See Recent Developments for the 2024 George Carlin lawsuit and settlement, which changed how this show is usually discussed.
  • Discover Daily | Perplexity ... an AI-generated daily news podcast voiced with ElevenLabs' long-form audio tools, an early example of a company publishing a fully synthetic show under its own brand.
  • NotebookLM Audio Overviews | Google ... not a show, but the tool behind many personal and published AI podcasts, including Spotify's 2024 Wrapped audio recaps.
  • Quiet Please network | Inception Point AI (now Liquid Studios) ... the largest experiment in mass-produced AI podcasts; see Recent Developments.

Generative AI can be used in several ways to make podcasts more successful

  1. Content creation: create new podcast content. For example, it can be used to generate new stories, script ideas, or interview questions based on the show's existing content and audience preferences. Research assistants grounded in your own sources (such as NotebookLM) are now good at producing episode outlines, guest briefing documents, and lists of follow-up questions. Treat these as drafts: check every factual claim against the source.
  2. Summarizing / transcriptions: speech recognition produces full transcripts that power search, captions, quotes, and accessibility. Apple Podcasts, Spotify, and YouTube now generate transcripts and chapters automatically, so your own transcript mainly serves your show notes, website, and corrections of names and jargon.
  3. Editing and production: edit and produce podcast episodes more efficiently. For example, it can automatically remove unwanted background noise or generate transcripts of the episode, making it easier for hosts to create show notes or captions. Text-based editing, filler-word removal, automatic leveling to loudness targets, and AI "co-editors" that carry out multi-step edits on request (such as Descript's Underlord) now cover most routine post-production.
  4. Advertising: create targeted advertisements that are more relevant to the listener's interests. By analyzing a listener's listening history, it can generate advertisements that are more likely to be clicked on or engage the listener. In practice, dynamic ad insertion and contextual targeting (matching ads to what an episode is about, using its transcript) are more common than per-listener generated ads, and listener-level targeting raises privacy and consent questions. Advertisers are also asking for ways to avoid buying placements on mass-produced AI shows.
  5. Translation and localization: voice-preserving translation and dubbing let a show reach listeners in other languages in the host's own voice. Human review of translations is still the norm for quality-sensitive shows.
  6. Discovery and repurposing: models pick highlight clips for short-form video, draft titles and descriptions, and generate chapters, all of which help new listeners find a show. With video now central to podcasting, clip selection has become one of the most widely used AI features.
  7. Synthetic voices and hosts: with consent, AI voices can patch a mis-spoken word, narrate an article version of an episode, or host fully synthetic segments. Platform rules now require disclosure when AI generates a material part of what listeners hear.
  8. Interactive and personalized audio: listeners can turn their own documents into custom audio briefings or interrupt AI hosts with questions, a format that sits somewhere between a podcast and a conversational assistant.

Producing Your Podcast

Youtube search... ...Google search

An earlier Nielsen survey found that half of all U.S. households considered themselves to be fans of at least one podcast, with 22% of these people identifying as “avid fans.” How to Start a Podcast: 7 Things These Experts Say You’ll Need | Tim Chan - RollingStone ...Three popular podcast hosts tell us how they started their podcasts, the equipment they use, and tips for keeping your audience entertained. For current audience numbers, see the Infinite Dial 2026 figures under Why This Matters above.

An AI-aware production workflow

The fundamentals have not changed: a clear concept, a good voice, a quiet room, and consistency matter more than any tool. AI mostly removes drudgery around those fundamentals.

Stage What AI can do Example tools Watch out for
Planning and research Outline episodes, brief you on a guest, suggest questions from your source material NotebookLM, general chat assistants such as ChatGPT, Claude, or Gemini Fabricated facts and quotes; verify against primary sources
Recording Record each remote guest locally, auto-level input, flag problems while you record Riverside, Descript Rooms, Podcastle No tool rescues a bad room; keep the mic close and consistent
Cleanup Remove noise, echo, hum, and mouth clicks; balance levels between speakers Adobe Podcast Enhance, Descript Studio Sound, Auphonic Over-processing makes voices sound robotic or "underwater"
Editing Edit by deleting words in the transcript; remove fillers and long pauses; carry out edits from plain-language instructions Descript (Underlord) Fully automatic filler removal can wreck comic timing and natural rhythm
Mastering Normalize to a loudness target and limit true peaks Auphonic, Descript, most digital audio workstations Aim for about -16 LUFS integrated and a true peak no higher than -1 dBTP
Packaging Draft titles, show notes, chapters, transcripts, and social clips Most editing suites; platform features on Apple, Spotify, and YouTube Generic AI-written descriptions hurt discoverability; edit them in your voice
Localization Translate and re-voice episodes Voice translation and dubbing tools Get consent for any voice cloning and have a fluent speaker review
Fully synthetic audio Generate two-host discussions from documents NotebookLM Audio Overviews, ElevenLabs Studio (GenFM) Disclose AI use in the audio and the feed metadata; do not imitate real people without permission

Disclosure checklist for AI-assisted shows

  • Decide what counts as material. AI noise reduction or transcription does not usually need a label. AI-generated voices, AI hosts, voice clones, translated voices, and AI-written segments presented as reporting do.
  • Disclose in the audio near the AI-generated portion, so the context survives when a clip is shared on its own.
  • Disclose in the metadata for both the show and each affected episode, as Apple Podcasts requires; use your host's AI-content setting if it has one.
  • Get written consent before cloning anyone's voice, including co-hosts and guests, and keep records of it.
  • Never fabricate quotes, interviews, or news events, and never imitate a real person in a way listeners could mistake for the real thing.
  • If you publish to EU audiences, check whether your content falls under the AI Act's deepfake or public-interest text disclosure duties.

How AI Podcasting Works

The Mechanism

There are two different pipelines worth separating: AI-assisted production, where humans still speak and AI handles the labor around them, and AI-generated production, where the script and the voices are both synthetic.

AI-generated shows (the NotebookLM-style pipeline)

  1. Ingest and ground the sources. The user supplies material: PDFs, web pages, Google Docs, YouTube links, or audio files (which get transcribed first). The NotebookLM team said in late 2024 that a notebook could hold up to 50 sources of up to 500,000 words each, far more than fits into a single prompt, so the system has to combine long-context reading with retrieval of the most relevant passages. Grounding matters because the hosts are supposed to discuss your material rather than general knowledge.
  2. Write a script. A large language model rewrites the sources as a dialogue. This is an editorial transformation, not a summary. The NotebookLM team described tuning for things that make human podcasts listenable: a clear narrative arc, a bit of disagreement or tension between the hosts, information revealed gradually instead of all at once, and analogies that expand on the source. The script carries speaker-turn markers and hints about where pauses, laughter, or interjections belong.
  3. Synthesize the voices. A speech model converts the marked-up script into audio. Modern systems treat this as a language-modeling problem over audio tokens: a neural codec compresses sound into a compact sequence of tokens, a model generates those tokens one step at a time, and a decoder turns them back into a waveform. Google DeepMind reported in October 2024 that its model could produce two minutes of two-speaker dialogue (more than 5,000 tokens) in under three seconds on a single TPU v5e chip. Much of the "human" texture, such as "mm-hmm," "right," short breaths, and overlapping reactions, comes from the audio model itself, which was fine-tuned on dialogue with realistic disfluencies.
  4. Post-process and package. The audio is normalized for loudness, sometimes mixed with music, and published with a title, description, transcript, chapters, and (increasingly) an AI disclosure.
  5. Optionally, let the listener talk back. NotebookLM's interactive mode (December 2024) lets a listener interrupt the hosts with a question. That requires live speech recognition, a fast language model response grounded in the same sources, and low-latency synthesis so the conversation does not stall.

Commercial systems do not publish their full architectures, so some details above are inferred from research papers and interviews rather than confirmed. The NotebookLM team, for example, did not confirm how closely its production model follows DeepMind's earlier SoundStorm paper.

AI-assisted human shows

  1. Record. Remote recording platforms capture each speaker locally at full quality and upload separate tracks, which avoids the artifacts of compressed video-call audio.
  2. Transcribe and diarize. Automatic speech recognition produces a transcript; diarization labels who spoke when.
  3. Edit as text. Tools such as Descript let editors cut audio by deleting words in the transcript, remove filler words and long pauses automatically, and regenerate a mis-spoken word with a voice clone (with the speaker's permission).
  4. Clean up the sound. Speech-enhancement models suppress background noise and room echo, a direct application of the Enhancement techniques used elsewhere in audio and imaging.
  5. Repurpose. Models draft show notes, titles, chapter markers, social clips, and translations. Spotify's 2023 Voice Translation pilot, for example, used OpenAI voice technology to re-voice episodes from hosts such as Lex Fridman and Dax Shepard in Spanish, French, and German while keeping each host's vocal identity, with human linguists reviewing the translations.
  6. Distribute. Platforms now add their own AI layer on top: automatic transcripts, automatically generated chapters, and links to things mentioned in an episode.

Chain of Physical and Information Processes

Every podcast, human or synthetic, is a chain of physical and information processes: air pressure becomes voltage, voltage becomes numbers, numbers get compressed, and the listener's ear turns pressure back into perception. A few principles explain most of what AI tools are doing.

Capturing sound

  • Distance and the inverse-square law. For a small source in open air, sound intensity falls with the square of distance, so sound pressure level drops by about 6 dB each time you double the distance from the microphone. Moving from 15 cm to 30 cm away costs roughly 6 dB of voice level while room noise stays the same, which is why mic technique matters more than mic price. Directional microphones used very close also show the proximity effect, a boost in low frequencies that makes voices sound warmer.
  • Room acoustics. Hard surfaces reflect sound and add reverberation. AI de-reverb tools estimate and subtract those reflections, but they work by inference, so they can smear consonants or produce a slightly processed sound when pushed hard.
  • Where speech lives. Most of the energy in speech sits below about 4 kHz, while much of its intelligibility (the crispness of "s," "t," and "f") comes from roughly 2 to 8 kHz. That is why aggressive noise reduction that cuts high frequencies makes speech sound dull even when it is technically "clean."

Turning sound into data

  • Sampling. The Nyquist-Shannon sampling theorem says you must sample at more than twice the highest frequency you want to keep. Podcast audio is usually recorded at 44.1 or 48 kHz, which covers everything up to about 22 to 24 kHz, beyond the roughly 20 kHz limit of human hearing.
  • Bit depth. Each bit adds about 6 dB of dynamic range: 16-bit audio gives about 96 dB, and 24-bit gives about 144 dB in theory. Recording at 24-bit leaves headroom so a sudden laugh does not clip.
  • Compression for delivery. Uncompressed 48 kHz, 24-bit mono audio runs about 1.15 megabits per second. A 128 kbit/s MP3 is roughly nine times smaller; one hour at that rate is about 58 MB. Lossy codecs work by discarding detail the ear is unlikely to notice.

Loudness and true peak

Listeners hate reaching for the volume knob between shows. The industry therefore measures perceived loudness rather than raw peaks, using the ITU-R BS.1770 algorithm, which weights frequencies the way the ear does and reports results in LUFS (also written LKFS).

Context Typical target Notes
Apple Podcasts -16 LKFS (plus or minus 1 dB), true peak at or below -1 dBFS Apple's published recommendation, measured per ITU-R BS.1770-5
AES TD1008 (streaming and on-demand) About -18 LUFS for speech-centric content Leaves more dynamic range for talk programming
Music streaming (Spotify guidance for music masters) About -14 LUFS Music is usually louder than talk
US / European broadcast -24 LKFS / -23 LUFS Broadcast audio uploaded as a podcast often sounds too quiet

The true-peak limit exists because lossy encoders can reconstruct the waveform slightly higher than the original samples ("inter-sample peaks"), causing clipping after encoding. AI mastering tools measure these values and apply gain and limiting automatically, and synthetic voices need the same treatment as human ones.

Why long-form AI speech became possible: token budgets

Speech models built on language-model architectures have to generate a sequence of tokens whose length grows with the duration of the audio:

tokens needed ≈ duration in seconds × frame rate × tokens per frame

Attention-based models get more expensive roughly with the square of sequence length, so frame rate is the critical scaling knob.

  • DeepMind's 2024 dialogue model needed more than 5,000 tokens for two minutes, a bit over 40 tokens per second of audio, and ran more than 40 times faster than real time (under 3 seconds for 120 seconds of audio).
  • Microsoft's VibeVoice (2025) used continuous speech tokenizers running at only 7.5 frames per second. Ninety minutes of audio is then 5,400 seconds × 7.5 = 40,500 frames, which fits inside a 64,000-token context window. At the 50 to 100 frames per second common in earlier speech representations, the same episode would need 270,000 to 540,000 frames, and the attention cost would be roughly 40 to 180 times higher.

This is the core technical reason AI podcasts jumped from 30-second clips (SoundStorm, 2023) to two-minute segments (2024) to full 90-minute, four-speaker episodes (2025).

Recognition accuracy

Transcription quality is usually reported as word error rate:

WER = (S + D + I) / N

where S, D, and I are substituted, deleted, and inserted words and N is the number of words actually spoken. A 5% WER sounds small, but in a 9,000-word hour-long episode it means about 450 errors, often concentrated in names and jargon, exactly the words show notes and search depend on. Long-context recognizers now process up to an hour of audio in one pass (VibeVoice-ASR reports 60 minutes within a 64K-token context), which helps keep speaker labels and terminology consistent across an episode.

Voice identity and perception

A voice is recognizable because of a combination of pitch (fundamental frequency), formants (resonances shaped by the vocal tract), timbre, and prosody (rhythm, stress, and intonation). Voice-cloning models learn these features from samples; a 2025 PLOS One study built convincing clones from about four minutes of recordings each. Detecting synthetic audio means looking for statistical traces those models leave behind, and marking it means embedding signals (watermarks or signed metadata) that survive compression, resampling, and mixing. No single marking technique is robust for audio on its own, which is why the EU's 2026 Code of Practice calls for layered approaches.

Why This Matters

Scale

Podcasting is now a mass medium, and video has joined audio as a primary format:

  • Edison Research's Infinite Dial 2026 found that 58% of Americans age 12 and older (about 167 million people) consumed a podcast in the past month and 45% in the past week, both record highs. Eighty percent have ever listened to or watched one, and for the first time a majority (57%) said they had both listened to and watched podcasts. Growth was driven heavily by listeners age 55 and up.
  • YouTube reported more than 1 billion monthly active viewers of podcast content (as of January 2025) and repeated that figure in May 2026, adding that YouTube Premium members watched more than 800 million hours of podcasts in April 2026. These are platform-reported numbers that outside auditors have not verified.

Older figures, such as the Nielsen household survey quoted in the production section below, counted households rather than individuals, so they are not directly comparable with Edison's numbers, but the trend is the same: podcast reach has kept growing.

What AI changes in practice

  • Accessibility and learning. Automatic transcripts, chapters, and translations make shows usable for deaf and hard-of-hearing audiences, non-native speakers, and people who skim. Turning documents into audio helps people with visual impairments or reading difficulties and anyone who learns better by listening.
  • Personal audio. "Podcasts for an audience of one" are new: your own research, meeting notes, or reading list, discussed back to you. Spotify used NotebookLM to generate personalized audio recaps for its 2024 Wrapped campaign, and Andrej Karpathy published a show, Histories of Mysteries, built from Audio Overviews of Wikipedia articles.
  • Global reach. Voice-preserving translation lets a show reach listeners in other languages without re-recording. NotebookLM's Audio Overviews now work in more than 80 languages, and ElevenLabs Studio's GenFM generates two-host podcasts in 32 languages.
  • Production economics. Editing that once took several hours per episode now takes a fraction of that, which lowers the barrier for independent creators and small newsrooms.
  • Machine-made shows at industrial scale. At about $1 per episode, a synthetic episode can turn a profit with about 20 listeners if it carries programmatic ads. That changes the incentive structure of podcast directories (see Recent Developments).

Risks and limitations

  • Accuracy. AI hosts can misstate, overstate, or invent details, and a confident conversational tone makes errors harder to notice than they would be in a bulleted summary. Generated shows are interpretations of their sources, not substitutes for them.
  • Discovery and "slop." Critics, including podcast scholars such as the University of Sussex's Martin Spinelli, warn that floods of cheap AI shows will make independent human podcasts harder to find without large marketing budgets.
  • Impersonation and fraud. Voice clones that listeners cannot reliably distinguish from real people enable fake endorsements, fabricated clips, and scams, and they put pressure on any system that still treats a voice as proof of identity.
  • Rights. Using a real person's voice or style raises copyright and right-of-publicity questions, as the George Carlin estate's 2024 lawsuit against the Dudesy podcast showed.
  • Labor. Editing, transcription, translation, and voice-over work are the jobs most directly exposed to automation.

What remains uncertain

  • Whether listeners stay with AI-generated shows over months, not just sample them out of curiosity. Early evidence is mixed (see the Inception Point AI case below).
  • Whether disclosure rules change behavior: do labeled AI shows lose listeners or advertisers, and will platforms enforce the rules consistently?
  • Whether audio watermarking and detection can keep up with open-weight voice models that anyone can run locally.
  • Whether "engaging" can be measured well enough to train against. Researchers still rely heavily on subjective listening tests.

Recent Developments

NotebookLM grows from a demo into a family of formats (2024 to 2026)

Audio Overviews launched in September 2024 with a single "Deep Dive" format and no settings: you added sources and pressed a button. The feature went viral, and its development has been fast since:

  • December 2024: interactive mode lets listeners join the conversation and ask questions.
  • 2025: Video Overviews add narrated, slide-style visuals. In late August 2025, Google brought full-length Audio Overviews and Video Overviews to more than 80 languages, replacing earlier, shorter non-English versions.
  • September 2, 2025: three new formats joined Deep Dive. Brief is a one- to two-minute summary; Critique gives an expert-style review of your material; Debate has the two hosts argue different positions. Reporters have also spotted a "Lecture" format in testing.
  • Late 2025 to mid-2026: NotebookLM moved to Gemini 3 models (December 2025) and then Gemini 3.5 (June 2026), added infographics, slide decks, and data tables, and split Video Overviews into standard and more elaborate "cinematic" versions.

What this shows: Google is treating the AI podcast as one output type among many from the same grounded source set, rather than as a standalone product. What it does not show: independent evidence on how accurate or educationally effective Audio Overviews are compared with reading the sources directly.

Open long-form dialogue models: MoonCast, PodAgent, VibeVoice

Academic and open-source work has focused on the two hardest problems: generating long audio and making it sound spontaneous.

  • MoonCast (Ju and colleagues; arXiv March 2025; NeurIPS 2025) generates multi-minute, two-speaker podcasts in the voices of speakers the model has never heard (zero-shot). The team trained a long-context audio language model in stages and paired it with a script generator that deliberately adds spontaneous details such as fillers and hesitations. Their key finding: those scripted spontaneity cues matter about as much as the text-to-speech model itself. They report gains in coherence and spontaneity over baselines, but English speaker similarity still lagged.
  • PodAgent (Xiao and colleagues, 2025) uses a multi-agent "Host-Guest-Writer" setup to draft more informative scripts and selects fitting voices from a voice pool.
  • PodEval (arXiv, October 2025) is an open evaluation framework for podcast-style audio. It separates content (what is said) from format (how it sounds) and combines objective metrics with listening tests, because open-ended, long, multi-speaker audio has no single correct answer to compare against. A January 2026 preprint adds a benchmark for podcast script generation from multiple documents with multiple speakers.
  • Microsoft VibeVoice was open-sourced on August 25, 2025, with the ability to generate up to 90 minutes of speech with up to four distinct speakers. Within weeks Microsoft removed the long-form text-to-speech code from its official repository, citing responsible-AI concerns after reports of misuse such as deepfake audio; community forks preserved copies. Microsoft later released a smaller real-time model (December 2025) that ships voices in an embedded format specifically to reduce deepfake risk, plus a speech recognition model that transcribes up to an hour of audio in one pass.

The VibeVoice episode is a useful case study in the central tension of the field: the same capability that makes a four-person AI podcast possible also makes realistic impersonation trivial.

Listeners can no longer tell voice clones from real voices

In September 2025, researchers at Queen Mary University of London and University College London, led by psychologist Nadine Lavan, published a study in PLOS One asking whether AI voices are now "hyperrealistic" (judged more human than real humans, as has been reported for AI-generated faces).

  • What they tested: 40 voice clones of real people (each made from about four minutes of recordings using a commercial tool) and 40 fully synthetic voices generated without a specific human counterpart, compared against real recordings.
  • What they found: clones were judged to be real 58% of the time, while genuine voices were correctly judged real 62% of the time, a difference too small to support reliable discrimination. Fully synthetic voices were easier to spot (41% mistaken for human). There was no hyperrealism effect, but AI voices were rated as more dominant, and some as more trustworthy, than human ones.
  • What it does not establish: the experiments used short clips and modest participant numbers under lab conditions, not hour-long listening in the real world. It does not tell us whether regular podcast listeners notice synthetic hosts over a full episode, where prosody errors and repetitive phrasing accumulate.

The practical takeaway for podcasters is that "you can hear the difference" is no longer a reliable safeguard, which strengthens the case for explicit disclosure.

The AI podcast factory: Inception Point AI to Liquid Studios

Inception Point AI, led by former Wondery COO Jeanine Wright, became the best-known test of the high-volume model:

  • September 2025: The Hollywood Reporter reported that its Quiet Please network had more than 5,000 shows and published more than 3,000 episodes a week, each costing about $1 and about an hour to produce, with dozens of AI personalities as hosts. Programmatic ads meant an episode could turn a profit with roughly 20 listeners. The company said its hosts identify themselves as AI at the start of episodes. Later interviews cited even larger catalog figures.
  • Criticism: Podcast industry coverage labeled the output "AI slop," and critics argued that keyword-titled niche shows (pollen counts, specific animals) mainly game directory search and ad marketplaces.
  • August 2026: the company rebranded as Liquid Studios and repositioned itself around transforming publishers' existing podcasts and newsletters into video, clips, and translated editions, rather than generating shows from scratch.
  • September 2026: in an interview on the Lead Balloon podcast reported by Podnews, Wright described some of the earlier publicity as "a PR strategy to get attention," and Podnews reported that the Quiet Please network had shrunk to about 300 active shows that were no longer being monetized. Wright told Podnews the catalog itself was not a stunt.

The arc matters more than any single number. Mass-produced AI shows proved technically and financially cheap to make, but the strongest company in the space has since moved toward augmenting human-made content, which suggests that volume alone did not produce a durable business.

Platforms add AI features and disclosure rules

  • Apple Podcasts content guidelines now require creators who use AI to generate audio or video (including synthetic voices, AI hosts or personas, and AI replicas of real people) to disclose it prominently in both the content and the metadata of each episode and show. AI may not be used to deceive, for example by fabricating news or manipulating clips.
  • Automatic chapters and timed links. With iOS 26.2 (December 2025), Apple Podcasts began generating chapters automatically for English-language episodes that lack creator-made chapters. These are labeled "Automatically created," and creators can opt out. "Timed links" surface links to music, books, shows, and other podcasts mentioned in an episode. Spotify introduced automatic chapters earlier in 2025. Advertisers have noted one side effect: clearly marked sections make it easier to skip ad breaks.
  • YouTube requires disclosure when realistic content is meaningfully altered or synthetic, but exempts productivity uses such as AI-drafted scripts, ideas, and captions.
  • Hosts and networks including Acast, Spreaker, and RSS.com have added AI-disclosure guidance or settings, and some write the disclosure into the open RSS feed so it travels to every app.

Regulation: EU AI Act transparency rules take effect

Article 50 of the EU AI Act became applicable on August 2, 2026. For podcasting, the relevant duties are:

  • Providers of generative AI systems must mark synthetic audio (and images, video, and text) in a machine-readable way so it can be detected as AI-generated.
  • Deployers who publish deepfakes, including audio that convincingly imitates real people, must disclose that the content is artificial. AI-generated text on matters of public interest must also be disclosed unless it has gone through human editorial review with someone taking responsibility for it.
  • The rules apply to providers and deployers outside the EU when their output is used inside the EU.

The Digital Omnibus on AI, adopted in mid-2026, delayed the Act's high-risk system requirements but left Article 50 on schedule. Its only relief here is a grace period until December 2, 2026 for generative systems already on the market to implement machine-readable marking. A Commission-backed Code of Practice on Transparency of AI-Generated Content, finalized in mid-2026, recommends layered marking (for example, metadata plus watermarking) because no single technique is reliable for audio on its own.

Rights and likeness: the Dudesy settlement

Dudesy, the comedy podcast with Will Sasso and Chad Kultgen whose premise is that an AI writes and steers the show, released an hour-long special in January 2024 presented as an AI-generated George Carlin routine. Carlin's estate sued for copyright infringement and violation of his right of publicity. After the suit was filed, a spokesperson for Sasso said Kultgen had actually written the material, with AI used for the voice and imagery. The case settled in April 2024: the special was permanently removed, and the hosts agreed not to use Carlin's image, voice, or likeness without the estate's approval. The estate's lawyers presented it as a template for similar disputes. The case shows that how a show is presented (as AI-made) can create legal exposure regardless of how much AI was actually used.

Learn More

Additional Viewing

Load video
YouTube

Behind the product: NotebookLM | Raiza Martin (Lenny's Podcast, October 2024, about 48 minutes). Google Labs product lead Raiza Martin explains how NotebookLM grew from a 20% project into the tool behind Audio Overviews, what models sit underneath it, and how the team handled red-teaming and potential misuse. It is the best first-hand account of why the two-host format works and how a source-grounded AI podcast is built. Show notes and transcript: Behind the product: NotebookLM | Lenny's Newsletter

Load video
YouTube

What does it take to make a Podcast? Also tips on video recording for small business too | Small Business Training by Martin Brossman. A practical look at what goes into producing a podcast, with tips on recording video, aimed at small-business owners. It predates most of the AI tooling described on this page, which makes it a useful baseline for the human fundamentals that AI tools build on.

Further long-form listening and viewing

Sources and Further Reading

Audience research

How AI podcasts are generated

Research papers and open models

Industry developments

Platform rules and technical standards

Regulation

Retrieved from "https://primo.ai/w/index.php?title=Podcasts&oldid=46788"