How an AI podcast actually works (the segment, not the fake two-host show)

The search box sold you two hosts

Type AI podcast, how to make a podcast, or podcast intro and the pages that come back are mostly the same product: upload a document, wait, download a two-person conversation that nobody wrote line by line. Those tools are not wrong about their own job. They are incomplete if the next thing you actually need is a take you can drop on a track.

A podcast episode is usually a recording of people talking. The frame around it (the cold open, the mid-roll, the one-line correction) is audio you splice. Collapsing the whole show into a chat about your PDF is convenient until the sponsor copy changes, the guest said a date wrong, or you need the same voice next Tuesday.

The product page for this job is Podcasts. This post is the craft version: the three jobs hiding under one search, what we return as a file, and what Apple Podcasts and Spotify actually say. How to write the spoken column already lives in how to write a voiceover script.

Three jobs, not one generate button

People type the same phrase for three products. Mixing them up is how a recorded interview gets replaced by a document recap, or a 30-second intro gets written like a chapter.

Connective tissue on a recorded show

If the interview is already cut, do not regenerate the conversation. Generate the ninety seconds that would send you back to the booth: a fresh cold open, a mid-roll for the new sponsor, a one-line correction for a date the guest got wrong.

House rule of thumb: about 1,000 characters of script is about a minute of audio. The Podcasts page budgets about 500 characters for a 30-second cold open, and about a minute of mid-roll from a 1,000-character read. Speed varies by voice, so treat it as an estimate. If the drop has to land on a sentence, generate that sentence as its own file.

The conversation stays the recording. Listeners came for that. Nothing about the frame around it needs a second session.

A scripted episode is still a script

Some shows are narration, not an interview. That is a legitimate AI-podcast job: you write the episode, a voice performs it, you export the files. It is still writing. The engine will not save a page that was built to be scanned.

Write for the ear. If a sentence is hard to say, cut it before you generate. We already wrote that fight in how to write a voiceover script. This post assumes you have a spoken column, not a blog post pasted into the box.

Split the episode the way you would split a session. Cold open. Body. Mid-roll. Close. One twenty-minute generate is how you get a take you cannot patch without redoing the hour.

What the two-host button actually is

The tool that made AI podcast a search phrase turns your notes into a conversation. Useful when you want to hear a document once. Weak when you need a sponsor line changed, a name pronounced right, or the same host voice in episode 40.

We do not generate two hosts from a PDF. We do not invent banter. If that is the product you want, use the product that sells it. If you want a file you can still rewrite, stay here.

Same file, two engines

The usual failure is directing a draft, or drafting a keeper.

Write the words on Kokoro until they are actually the words. It is the unlimited draft engine: a clean, plain read, fast enough that rewriting a sentence is cheaper than arguing about the take. When a line has to sound like a person leaning in, move that line onto the engine that matches the job.

For a cold open that needs to hold, start on Grok Voice: five English personas, a confident characterful take. For a sponsor read with a cue in brackets, Gemini Flash is the engine on this roster that acts [warmly] or [upbeat]. The Podcasts worked example is that split. The longer routing argument is in pick the engine for the job. The short definition of a cue is what is a TTS cue.

Beds under an intro are a second export. Generate the voice here, generate a quiet instrumental in Sound and Music, mix them in the editor you already use. Duck the music under the voice. Do not ask the music model for the host.

Apple and Spotify's pages, not the rumor

Two different store rules get mashed into one internet sentence. Disclosure is about whether a listener might think a person is speaking when a machine is. Impersonation is about pretending to be someone else. They are not the same form.

As of October 2026, Apple Podcasts content guidelines section 1.11 requires creators using AI to generate audio, including synthetic voices and AI-generated hosts, to disclose that prominently in the content and the metadata for each episode and show. Section 1.12 forbids using AI to mislead, including fabricated news or clips presented as real events. A roster voice on an intro you wrote is the disclosure case, not a ban. Put a spoken line in the episode and a line in the show notes. We are not Apple. Read their page. Verify before you rely.

As of October 2026, Spotify's platform rules still do not ban synthetic narration on a show you actually host. What they police is deception: impersonating another person, and synthetic media passed off as authentic in ways that pose a risk of harm. A generated intro in your own show's voice is not impersonation. Cloning someone else's voice without permission is. Acceptable use draws the same line. Spotify's October 2026 human artistry update is about music recommendation and spam tracks, not a new podcast TTS ban. Same caveat: store rules move. Check the live page before you file.

Where it still fails

A fluent take can still be the wrong file. These are the misses we see when people treat the search phrase as a spec.

• Replacing a recorded interview with a recap of the notes.
• One twenty-minute generate for a drop that needed four files.
• An intro written like module three of a course.
• A mid-roll with no cue, then blaming the voice for a flat read.
• Cues on a plain-read engine, then blaming the voice for ignoring them.
• Asking the studio to host RSS, mix the episode, or invent a co-host.
• A clone of a real host (we do not offer that, and acceptable use would not either).

Honest limits

We generate files. We do not sit inside Hindenburg, Descript, or your host. The output here is audio you own: commercial use, no watermark, same as the rest of the studio (you own what you make here).

No two-host conversation from a PDF. No in-app DAW. No RSS. No caption file. The console API tab is still labeled soon; we will not pretend a public render endpoint exists.

Cue-following is Gemini Flash only. The other engines on the roster read plainly. We strip brackets so they do not speak the direction. We do not pretend they performed it.

A generation has a character cap. Split a long episode. Do not paste a 40-minute script into one box and call the truncation a voice problem.

Premium generation sits inside a flat monthly allowance (published free-tier premium budget: 10,000 characters, about 10 minutes by house math). Kokoro drafting stays unlimited so you can rehearse an intro without burning keeper budget on every micro-edit. Paid plans raise the flat allowance; they do not put you on a per-character anxiety meter.

Voice cloning is not a live path today. When it ships it will be English-first with a consent attestation. Until then, roster voices and personas are the honest path.

After the episode ships, Speech to Text will take the finished file and give you a transcript for show notes, pull quotes, and a record of what was actually said. It will not write the episode for you.

Cut one this week

The shortest proof is a drop you already owe, not a roundup of vendors.

• Pick a 30-second cold open or a 60-second mid-roll. Count it.
• Write the spoken column only. If a sentence is hard to say, cut it.
• Draft the words on Kokoro until they are the words.
• Move the keeper onto Grok Voice for a hook, or Gemini Flash if the line needs a cue.
• Export. Drop the file on a track in the editor you already use.
• If Apple is a destination, add the spoken disclosure and the metadata line before you submit.
• Listen once without looking at the page. If a word is wrong, fix the page, not the voice picker.

Open the studio, not the fake hosts

If you came here from an AI podcast or how-to-make-a-podcast search, the honest product is a spoken file you can hear this afternoon. Open Podcasts, or start in Text to Speech if you already have the lines.

When you want a timed spot with a cue you can audition, the worked example lives on Advertising. When you want the full studio, create a free account. Written to be heard includes the frame around the conversation, as long as we do not lie about what that file is for.