When should podcast creators label their episodes as AI-generated? A clear, actionable set of rules built by studying what every major platform and regulatory body requires. Designed for podcasters who want a straight answer, not a legal essay.
Try the Interactive ToolThe Substance Test
Disclose when AI is doing the creative work your listeners came for.
The test is not "was any AI involved?" It's about substance. If AI generated or performed the core content of your episode (the voice, the narration, the conversation), that's substance. If AI helped with production (a jingle, cleanup, show notes), that's a tool.
Apple Podcasts uses the same concept: disclosure is required when AI generates "a material portion of the podcast's audio." The word material is doing the heavy lifting. A 10-second AI-voiced intro before 45 minutes of two human hosts is not material. A fully AI-narrated episode is.
Ask one question about each episode:
"Is AI doing the creative work my listeners came for?"
If AI is the performer (delivering the content listeners engage with as speech, narration, or conversation): disclose.
If AI is a production tool (supporting human performance with cleanup, music beds, jingles, effects, or behind-the-scenes work): don't disclose.
This mirrors how Apple, the EU AI Act, and the Partnership on AI all define materiality.
These are cases where AI is doing the substantive creative work:
| Use Case | Why |
|---|---|
| Fully AI-generated narration A synthetic voice delivers the episode | AI is the performer. The voice the listener came for is not human. Clearest case. |
| AI dialogue or conversation e.g., NotebookLM-style two-host discussions | Both "hosts" are synthetic. AI is doing all the performative work. |
| AI-translated episodes Your voice cloned into another language | The entire audio in that language is AI-generated. The substance of that version is synthetic. |
| AI voice delivering content segments AI reads listener questions, narrates sections | AI is delivering content the listener engages with as speech. That's substance, not production. |
| Self-voice cloning + AI-written content AI wrote the script AND generated the audio | AI is doing the creative work at every level. The ideas aren't yours, the words aren't yours, and the performance is synthetic. |
| Human voice + fully AI-written script You recorded yourself reading AI-generated content | AI created the substance your listeners came for. You performed it, but the creative work (ideas, words, narrative) is AI's. |
Self-voice cloning is the hardest case in this framework. It requires separating two things: the creative substance and the delivery method.
| What happened | Creative substance | Disclose? |
|---|---|---|
| You wrote the script. AI cloned your voice to narrate it. | Yours. Your ideas, your research, your words, your editorial judgment. | No AI is a production method, not the creator. |
| AI wrote the script. AI cloned your voice to narrate it. | AI's. The ideas, words, and performance are all synthetic. | Yes AI is doing the creative work at every level. |
| You wrote part, AI wrote part. AI cloned your voice. | Mixed. | Use judgment. If AI generated a material portion of the content, disclose. |
Why this distinction matters: When you write your own script and use your cloned voice to deliver it, the creative work your listeners came for (your perspective, your expertise, your personality, your editorial voice) is 100% yours. AI is just the delivery mechanism. This is closer to choosing a microphone than to delegating creative work.
The listener came for your ideas. Whether you spoke those ideas into a microphone, dictated them, or produced them via voice cloning, the substance is the same.
When you should still consider disclosure: Even when the substance is yours, if the cloned voice delivers a full episode and listeners would reasonably assume they're hearing a live recording, transparency is good practice. You don't need the AI tag for this, but a note in your show notes ("This episode was produced using my cloned voice") is a nice touch.
These are cases where AI is a production tool, not a performer:
| Use Case | Why |
|---|---|
| Self-voice cloning + your own content | Your ideas, your words, your voice. AI is the delivery method, not the creator. |
| Short AI-voiced intro or outro A produced jingle or bumper | Production element. Not the substance of the episode. The listener came for the human hosts. |
| AI-generated background music Ambient beds, theme music | Music supporting human content is a production choice. Like using a stock music library. |
| AI sound effects A ding, a whoosh, a transition | Functional audio elements. Same category as stock sound libraries. |
| Noise removal / audio cleanup | Production tools, not content generators. Every framework exempts this. |
| Filler-word removal (um, ah) | Editing, not generation. Every word the listener hears was spoken by the host. |
| Audio leveling, EQ, compression, mastering | Standard post-production, even when AI-powered. |
| AI-generated transcripts | Derivative of human-spoken audio. The audio itself is human. |
| Show notes, titles, descriptions, chapters | Text metadata. The listener's audio experience is unchanged. |
| AI-assisted scriptwriting Brainstorming, outlining, research, editing suggestions | You wrote the core content. AI helped with the process, not the substance. Apple, YouTube, and Meta all explicitly exempt this. |
| Research and fact-checking | Tool use, not content generation. |
Performer vs. tool.
A human host with an AI-generated jingle: AI is a tool. No disclosure.
An AI voice narrating the whole show: AI is the performer. Disclose.
Your own cloned voice reading your own script: your substance, AI delivery. No disclosure.
You recording yourself reading a fully AI-written script: AI created the substance. Disclose.
AI reading listener questions between human segments: AI is delivering content. Disclose.
Because "substance" can feel subjective, here are concrete scenarios:
| Scenario | Substance? | Disclose? |
|---|---|---|
| 10-sec AI voice intro, then 45 min of 2 human hosts | No. Production element. | No |
| AI narrates the whole episode, human only in intro/outro | Yes. AI is the performer. | Yes |
| You write a script, your cloned voice delivers the full episode | No. Your ideas, your words. | No |
| AI writes a script, your cloned voice delivers it | Yes. AI did the creative work. | Yes |
| You prompt AI to write a full script, then record yourself reading it | Yes. AI created the content. | Yes |
| AI voice clone translates the full episode to Spanish | Yes. Entire version is synthetic. | Yes |
| AI reads all listener questions between human segments | Yes. AI delivering content. | Yes |
| AI-composed 30-sec theme music, human hosts throughout | No. Music bed. | No |
| NotebookLM generates a full AI conversation episode | Yes. Both speakers are AI. | Yes |
| Human host, AI-composed score running under entire episode | No. Music is production. | No |
An earlier version of this framework said "any AI-generated audio triggers disclosure." We revised it because that's too strict and misaligns with how every major platform defines the obligation.
Even though the test involves judgment (is this substance?), the disclosure itself remains binary: yes or no.
The checkbox is not a punishment. It is a signal. Checking it does not mean your content is lesser. It means you're transparent about how your content was made.
This framework was built by studying what every major platform and regulatory body requires:
| Platform / Framework | Their Rule | This Framework |
|---|---|---|
| Apple Podcasts | Disclose when AI generates "a material portion of the podcast's audio" | "Material portion" = substance. |
| YouTube | Disclose "realistic altered or synthetic content." Exempt: scriptwriting, audio cleanup | Exempt list matches. |
| Meta | Label photorealistic AI video/audio. Auto-detect via C2PA | Audio substance maps to voice/narration rules. |
| Spotify | Three tiers: human, AI-assisted, fully AI-generated | Simpler binary, but compatible. |
| EU AI Act Article 50, effective Aug 2026 | Deepfakes must be disclosed. Fines up to 15M EUR / 3% global turnover | Prepares creators for EU compliance. |
| FTC (US) | Deceptive AI endorsements violate Section 5. $51,744 per violation | Prevents deception by defaulting to transparency. |
| TikTok | Disclose realistic AI people/scenes. Auto-labels via C2PA | Aligned for audio equivalents. |
| BBC | All generative AI use must be disclosed | BBC's rule is stricter (we exempt production tools). |
| Partnership on AI | Material = voice/likeness clones. Immaterial = de-noising | Substance test mirrors their materiality framework. |
| C2PA | Cryptographic provenance chain in media files | Not yet in podcast audio, but conceptually aligned. |
Created by Alberto Betella, PhD, co-founder of RSS.com. This framework was built by researching what every major platform and regulatory body requires, then distilling it into a clear, actionable set of rules for podcasters. See Sources & References.
This framework is offered as a public resource for the podcasting community. It is not legal advice. Each platform has its own AI disclosure policies, and those policies may change. Always check the specific requirements of the platforms where you distribute your podcast. By using this framework, you acknowledge it is general guidance and does not replace compliance with individual platform rules or applicable laws.
Try the interactive tool to quickly determine whether your episode needs disclosure.