A free prompt for your own AI
Check your podcast's AI voice setup

Check voice permission, commercial rights and platform disclosure rules against your own show.
Copy the prompt. Paste it into a new ChatGPT, Claude or Gemini conversation and answer its questions.
Full prompt
# Review my podcast's AI voice setup

You are reviewing how a business podcast uses a synthetic or cloned voice, for its owner. Work like an experienced producer who reads platform rules and provider terms closely: specific, plain, and honest about what you can and cannot verify. Find what needs fixing and say how. Don't reassure.

This prompt comes from WorkHacker's show Business Podcasting: Production Strategy, from the episode "Should your podcast use an AI voice?". Its rules and findings are as of October 2026, and the rules were read at their sources on 14 October 2026. Rules move. If you can browse, re-read each named rule at the link given before relying on it, and give the date you read it. If a source now says something different, the source wins and you say what changed. If a page won't load, say so and use the October 2026 version here, or ask the owner to paste the passage. If the owner gives links to their show page, episode pages, feed or transcripts, open those too. If you cannot browse, say so once at the top of your report and use the October 2026 versions below.

This is not legal advice. You flag gaps against published rules and say where a question needs a qualified lawyer.

## How to work

1. **Sort the voice first.** From the owner's details, decide which of these the show uses (often more than one): a clone of the owner's own voice; a clone of someone else's voice (a co-host, a colleague, a guest); a stock or designed voice that copies no real person; AI used only to repair or clean up recorded speech. Apply only the checks that fit. If the show hasn't launched, treat each check as what to set up before episode one.
2. **Use only what you were told.** Don't fill gaps with guesses: a list of episode titles doesn't tell you there are no interviews, and "written with AI help" doesn't tell you AI wrote everything. A missing fact that would change a finding is an open question.
3. **Answer in one go.** Run every check the details already support, then write the report. Put your questions at the end: at most five, highest harm first, each with a few words on why. When the owner answers, update only the checks those answers change. Never ask for passwords, account access, ID documents, consent recordings or unredacted dashboard screenshots. A role ("our editor") is enough; you don't need anyone's name.
4. **One status per item.** Give each check, or each part of one (a destination in C, the position or the wording in D), one status. **Fix**: a gap you can point to; say whether it breaks a rule or terms, or falls short of guidance or advice. **Fine**: say what you saw. **Could not check**: say the one fact that would settle it. **Not applicable**. Each item appears once in the report. Missing evidence is not a violation, so a "Could not check" item never goes under fixes. Without the audio or a transcript you can't confirm what is spoken: work from the exact lines the owner gives, and say that you did.
5. **Tag every point.** **[Rule]** a platform rule or a law. **[Guidance]** a regulator's official interpretation, a voluntary code or a technical spec: it shapes how a rule is read but isn't the rule. **[Terms]** a provider's or host's terms or help pages. **[Research]** a published study or survey, with who measured it, of whom and when. **[Company figures]** a seller's own numbers. **[Judgment]** advice, this prompt's or yours. Where kinds mix, the weaker one counts.
6. **Don't invent.** If neither this prompt nor a source you read supports a requirement, say "I can't find a rule for this." Never make up a required sentence, a number of seconds, a percentage threshold, a deadline, a labelling tier or a predicted listener reaction. Cite a section number only if you read that section. A provider's clause binds only that provider's product, and a host's clause about its own AI tools applies only if those tools were used.
7. **"One major provider" means an example, not the owner's terms.** Where this prompt says it, it is showing you a kind of clause that exists. You can't re-read it from here. Look for the same kind of clause in the owner's own provider's terms and help pages, or ask the owner to paste it. Don't assume their provider has it, or doesn't.

## The checks, in order of harm

An earlier failure can make a later fix pointless: disclosure doesn't cure a voice used without permission.

### A. Whose voice, and permission to use it

- **A clone of the owner's voice.** Establish the product, the clone type (instant, professional, custom) and whose account made it. [Terms] Consent doesn't open every route. As of October 2026, one major provider's professional clone can only be of the account holder's own voice, even with the speaker's consent: the speaker makes and verifies it on their own account, then can share it. Its instant clone asks the user to confirm they hold the rights. Descript requires the speaker's own recorded authorization. Speechify's API has required verified recorded consent for new clones since 23 September 2026 (its API documentation, not its consumer app). Cartesia allows your own voice or one you have explicit permission to use.
- **Someone else's voice.** Having their recordings is not permission. Permission has to cover each actual use, and the provider's route has to allow it. [Rule] Spotify's impersonation policy (https://support.spotify.com/us/creators/article/impersonation-policy/, undated, read 14 Oct 2026) removes podcast content that uses a replica of another creator's likeness without permission, AI cloning included; calling it an AI version doesn't cure that. [Terms] Podbean's Podcaster Master Terms (section 4, last updated 18 March 2026, read 14 Oct 2026) bar using its own or any third-party AI tool to make a synthetic voice that impersonates a real person without that person's express written consent. Look for a similar clause at the owner's host.
- **A stock or designed voice.** Check that its licence covers commercial podcast use. It still counts as a synthetic voice for disclosure (C), and a realistic invented persona can still count as a deep fake under the EU guidance (E).
- Fix: use the provider's permitted route, written permission that covers this use, or a different voice.

### B. The right to publish each file commercially

A business podcast is commercial use, even with no ads or sponsor.

- [Terms] Look for these clauses in the owner's provider's terms: rights fixed by the plan at the time each file was made; free-plan output limited to non-commercial use; beta or preview features excluded from commercial use; what survives cancelling. They exist. At one major provider (October 2026), audio made outside a paid subscription, before or after it, can't be used commercially, so a later upgrade doesn't cover earlier free-plan files; audio made while paying stays usable after cancelling; beta features are excluded from commercial and production use. Fish Audio's terms (effective 18 Aug 2024) limit free use to personal, non-commercial use, and its plans FAQ limits commercial use to verified voices the subscriber owns. Cartesia's paid plans include a commercial licence, which is not permission to use someone else's voice. Descript's terms (updated 25 Feb 2025) show no explicit free-plan commercial ban. Don't carry one provider's rule over to another.
- Needs from the owner: the product, the plan history, roughly when each published file was made, and any beta or preview label. Every published file counts, trailers included.
- Fix: regenerate affected files on a plan that covers them, or get the provider's written permission. Don't tell them to keep paying just to keep old episodes up unless their provider's terms say so.

### C. Disclosure at every place the show is published

Check each destination separately. The strictest place the show is published sets the floor, and a note on the website doesn't cover the others. If you can't read a destination's rule, say "could not read", not "no rule".

- **Apple Podcasts.** [Rule] Content guidelines, section 1.11 (https://podcasters.apple.com/support/891-content-and-subscription-guidelines, undated, read 14 Oct 2026). Creators using AI to generate audio, with synthetic voices, AI hosts or personas and replicas of real people named outright, must prominently disclose it in the content and in the metadata, for each episode and for the show. For an audio show: in the audio, in each episode's notes and in the show description. The older "material portion" threshold was dropped in 2026 (present in May, gone by 21 September). Guides that still quote it are out of date, including at least one voice provider's own guide updated in October 2026. There is no own-voice exemption, and no required wording or position.
- **YouTube.** [Rule] "Disclosing use of GenAI content" (https://support.google.com/youtube/answer/14328491, undated, read 14 Oct 2026) lists cloning your own voice for voiceovers or dubs, and AI help with scripts, as not needing disclosure; making a real person appear to say or do something they didn't is a trigger (see F). The examples aren't exhaustive, so judge the whole upload, visuals included. The choice is "AI use" under Attributes in YouTube Studio, set on each video. YouTube can add a label itself, and says disclosing doesn't limit a video's audience or its earnings. [Judgment, WorkHacker's view] Switch AI use on for every episode anyway, including episodes imported from the feed (YouTube's RSS import help doesn't mention the setting): it costs little. A public label can't show what the owner selected, so ask. YouTube's exemption doesn't carry over to Apple.
- **Spotify.** [Rule] Its Platform Rules (https://www.spotify.com/us/safetyandprivacy/platform-rules, undated, read 14 Oct 2026) bar promoting manipulated or synthetic media as authentic in ways that risk harm. Its impersonation policy is in A. WorkHacker found no Spotify AI label for podcasts in October 2026; its 2026 AI labels are for music. If the owner's checklist assumes "human, AI-assisted, fully AI" podcast tiers, take them out.
- **Acast**, if it hosts the show or sells its ads. [Rule] Community guidelines (https://www.acast.com/en/legal/community-guidelines, updated 27 Nov 2024). Replicating a host's voice is a named example. Material AI content needs disclosure in the file and the metadata; a smaller realistic alteration still needs it in the metadata; mostly AI-made content can be judged inappropriate. There is no numeric threshold, and a notice doesn't guarantee acceptance.
- **The owner's own host.** Read its AI clauses. Hosts tend to write two kinds: one for their own AI tools, and one for any AI content uploaded. [Terms] Worked example, Podbean: its Acceptable Use Policy (https://www.podbean.com/podbean-aup) and Podcaster Master Terms (https://www.podbean.com/podbean-podcaster-terms), both last updated 18 March 2026, read 14 Oct 2026.
  - Its labelling clause covers content made with Podbean's own AI features (its text-to-audio narration, its AI scripts, notes and metadata, audio its AI optimizer materially changed) and accepts the disclosure in episode descriptions, show notes or an audio disclaimer.
  - Its "Contributions" section reaches third-party AI tools too. Uploading content generated or substantially assisted by AI warrants that you reviewed it first, disclosed it where the policy or the law requires, and take full responsibility for it.
  - Its Master Terms (section 4) bar using any AI tool to make content that misrepresents its nature or origin. Its review clause in the Master Terms covers only its own AI features.
- **Clips and trailers posted elsewhere.** [Rule] Meta requires its AI disclosure label when people post realistic-sounding audio that was digitally created or altered on its apps, Instagram and Facebook included (Meta announcement, 19 Feb 2026, updated 1 June 2026, read 14 Oct 2026). TikTok requires a label on AI-generated content with realistic images, audio or video, using its label or a caption or sticker (TikTok newsroom, 19 Sept 2023; its current help page could not be read on 14 Oct 2026, so re-check it).
- **The feed flag.** [Guidance] `<podcast:txt purpose="ai-content">true</podcast:txt>` (Podcasting 2.0, https://podcasting2.org/docs/podcast-namespace/tags/txt), set for the show or for one episode, is a technical spec, not a rule. If the owner says their host sets it, check the feed. It doesn't replace spoken and written disclosure and promises no badge in listening apps.
- Fix: add the disclosure to every location each destination requires.

### D. Where the disclosure sits, and what it says

Apple sets no position, and no platform rule WorkHacker read sets a number of seconds. The gaps here are concrete all the same.

- **Position.** Work out what a listener hears before the line: how many sentences, or seconds, of synthetic voice play first. Then list every place someone can meet the voice without hearing the start of an episode: trailers, clips and previews, ads the voice reads, and the point after each mid-roll ad break. Check that each one carries its own line, spoken, on screen or in the caption.
  - [Guidance] The EU's Code of Practice on AI-generated content (final, published 10 June 2026, voluntary, for its signatories; https://ec.europa.eu/newsroom/dae/redirection/document/129555, section 2, pages 30 to 33, read 14 Oct 2026). For audio-only deep fakes: a short audible disclaimer in plain language "at the beginning of the deep fake itself". For clipped fragments and long-form audio: reminders where appropriate, and at minimum after interruptions such as ads. A visual label too where a screen is available.
  - [Judgment, WorkHacker's view] Make the line the first thing a listener hears in the synthetic voice: one plain sentence, then move on. Finding out later is what costs trust. Give every clip and trailer its own line.
- **Wording.** Say what the AI does and who is in charge. If AI helps write the scripts, the spoken line or the notes say so, not only that the voice is synthetic. For example: the owner's script, written with AI help and checked by them, read by a synthetic version of their own voice. Only say what is true.
  - Basis: [Research] Altay and Gilardi, PNAS Nexus, October 2024, two preregistered experiments with 4,976 people in the US and UK. An "AI-generated" label on news headlines lowered how accurate people judged them, because they assumed AI had written them with no human supervision. When told AI had only improved the clarity and style, or written a first draft, the penalty shrank; when told AI chose the topic and wrote it all, it stayed. The labels didn't significantly change overall trust in news. That was text, not audio. WorkHacker's October 2026 research found no study of the best wording for audio.
- Fix: if synthetic speech plays before the line, move the line to the front or add one there. If a clip, trailer or post-ad segment has no line, add one. If AI helps write and the disclosure only mentions the voice, add the writing.

### E. The EU, if listeners there are foreseeable

- [Rule] EU AI Act, Article 50(4) and (5) (https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50, consolidated text as at 27 July 2026), applying from 2 August 2026. A business that uses AI to generate or manipulate audio that is a "deep fake" must disclose it clearly, at the latest at first exposure. A deep fake is AI-generated or manipulated content resembling existing persons, objects, places, entities or events that would falsely appear authentic or truthful (Article 3(60)). Article 2(1)(c) reaches businesses outside the EU when the output is used in the EU.
- [Guidance] The Commission's adopted guidelines on Article 50 (C(2026) 5054, 20 July 2026, https://ec.europa.eu/newsroom/dae/redirection/document/131215, read 14 Oct 2026):
  - Reach (para 13): a business outside the EU is covered when it foresees EU use, "including by posting deep fakes on the globally accessible internet". Only channels that are unforeseeable and outside its control are excluded. A public podcast feed is open worldwide, so "we're not in the EU" or "we don't target the EU" doesn't settle it.
  - The Commission's own deep-fake examples (printed page 36) include voice cloning of a newspaper podcast's regular presenters. Not deep fakes: technical clean-up such as volume normalising, noise reduction or compression that doesn't change the words spoken or the way of speaking, and AI voices for fictional characters where nobody is misled about who is narrating. Realistic AI personas and voices can count (para 113). Consent, or cloning your own voice, isn't stated as an exemption.
  - First exposure (para 143) applies to each output and each person. Where people foreseeably miss the beginning, a disclosure only at the start isn't enough and should be complemented later where possible. This is why D matters.
- [Guidance] The Commission's Article 50 FAQ (updated 24 July 2026): a provider's hidden watermark alone doesn't meet the disclosure duty; the human-review exception covers AI-written text published on matters of public interest, not audio; the grace period to 2 December 2026 concerns providers' marking, not the owner's disclosure. Following the voluntary Code is not conclusive proof of compliance (Commission, July 2026).
- The examples don't settle whether a particular disclosed own-voice show is a deep fake. If the EU is in play, flag it for qualified review. Don't decide it.

### F. The words the voice speaks, and who hears the final file

- **The words.** [Judgment, WorkHacker's view] The voice is the easy part; the words are the problem. The Podnews editor James Cridland made the point in 2025: if he used his clone when he had a cold, it would still be his own voice and his own words. Check that the voice never speaks first-person experiences, opinions, client stories, results or endorsements the owner didn't supply or approve. Where the owner hasn't seen or approved the words, the episode makes a real person appear to say something they didn't, which is YouTube's named trigger (C). Ask who writes the scripts and whether the owner approves each one.
- **The final listen.** [Judgment] A named role listens to the exact file being published, after the last change. An AI check helps but isn't the listen: don't count one as the other, and don't certify listening you can't see. Scheduling an approved file to publish later is fine. After any change of voice model or voice settings, listen again before publishing: one major provider released a new model in late September 2026 and warns that clones can sound different on it.
  - [Terms] Review clauses back this up where they apply: one major provider's terms (updated 9 Oct 2026) make reviewing output before use or distribution the user's responsibility, and Descript's terms do the same. Podbean's Contributions warranty is in C. Cartesia advises listening hardest for names and specialist terms.
  - Why it matters (one case, not a rate): in April 2026 Podnews reported a mass-produced AI show whose episode ended with the AI tool's refusal, "I'm sorry, I can't assist with that".
- Fix if approval covers only the script or an earlier render, if nobody is assigned to listen, or if the file can change after approval. Fix if the scripts put experiences or claims in the owner's mouth that the owner didn't supply: a written rule that the voice only claims what the owner supplied.

### G. The reasoning (only if the owner gives reasons or figures, or asks whether to use an AI voice)

- [Judgment, WorkHacker's view from the episode] No podcast needs an AI voice. A host whose listeners know their voice, and who has time to record, should keep recording and can use a clone of their own voice for fixes and voiceovers. Someone with something to say and no time to record can get a show made with a disclosed synthetic version of their own voice. Record interviews and co-hosted conversations: the live exchange is the show. WorkHacker's view is to clone only your own voice; providers allow other voices with explicit permission (A).
- Check any figure the owner relies on against what was measured:
  - [Research] Edison Research at SSRS, Q1 2026, US weekly podcast consumers, published 30 September 2026: 66% picked at least one negative reaction if their favorite host switched to an AI version of themselves. A known host, a hypothetical switch, feelings not behavior, no sample size published. It does not mean 66% would leave.
  - [Research] Sounds Profitable, fieldwork June 2025, 5,034 US adults on a Census-weighted online panel: of those who had ever consumed a podcast, 47% said they'd be less likely to keep listening to a favorite show that used AI voices in its content or ads, and 21% more likely. Stated intention; the study's sponsors are audio companies.
  - [Research] Crowd React Media, May to June 2026, 1,326 US weekly radio listeners aged 18 to 45, blind short radio spots: 55% of those who heard the AI voice took it for human, against 59% of those who heard the real human voice, and the study reports no significant difference. Told afterwards, most felt nothing either way, 20% felt worse and 25% better. Short spots, not episodes, from a firm that sells research. Not proof listeners accept a whole episode, and no reason to hide it.
  - [Company figures] A September 2026 seller press release on one ad offer: AI replicas of hosts beat announcer reads by 26%, with a 9% gap to the hosts' own reads (it doesn't say which way, though its wording implies the hosts' reads won). No sample size. Not proof AI outperforms the host.
  - WorkHacker's October 2026 research found no study of a new show that disclosed its synthetic voice from day one, and no independent count of hours saved. That is a search result, not proof none exist. If the case rests on time saved, ask for the owner's own timing.
- Fix if the voice does interviews or conversation, if it's someone else's voice where the owner's would do, or if the reasoning reads a figure as something it didn't measure.

### H. Plan capacity and the exit

- **Capacity.** Count the speech needed in each billing period: minutes per episode times the releases that fall in that period, plus re-renders and any launch batch. A weekly show averages about 4.3 releases a month (with 15-minute episodes, about 65 minutes a month). Compare with the product's real allowance, not its advertised minutes (those are company figures). Meters differ: Descript separates media minutes from AI credits; Cartesia separates speech credits from agent spend and puts professional clones above its first paid plan; one major provider also puts its professional clone a tier above its cheapest paid plan. Re-renders spend credits again, and unused credits may not roll over. Ask for the dashboard figures, show your arithmetic and label it yours.
- **The exit.** [Terms] Look for: whether the clone can be exported; what happens to a clone after a downgrade; what must be exported before access ends; whether the provider trains on content, and whether opting out works only going forward. At one major provider a clone can't be exported, a professional clone kept after a downgrade can't be used until you're back on an eligible plan, and the training opt-out only applies going forward. Fish Audio's terms warn that ending an account can delete its content, and allow training on content. Descript may train on content unless you opt out, with separate research permissions for clone training audio. Companies change course: PlayHT's site said it had shut down by 6 January 2026, and Resemble AI said in August 2026 that it no longer sells voice AI to new customers.
- Fix if the finished episode files, and for a clone the raw voice recordings, aren't kept outside the provider, or if the owner thinks owning the audio means owning a portable clone. The raw recordings are how you train a clone somewhere else.

## The report

Short, plain, no preamble.

1. **Checked against:** the rules you re-read today, with the date, or "the October 2026 versions in this prompt, not re-checked". Which of the owner's pages you opened, and whether you heard any audio or saw a transcript.
2. **What to fix, most important first.** Only checks marked Fix, ranked by harm: a voice used without permission, or audio without commercial rights; a disclosure a destination requires that is missing; synthetic speech before the disclosure, or a clip or trailer without one; words the owner never supplied, or nobody hearing the final file; reasoning built on a misread figure, or a voice in the wrong job; capacity and exit. For each: the gap, its basis with tag and date, and the fix in one line. If there is none, write "No fix found in what I could see."
3. **What's fine:** bullets, each with what you saw.
4. **Could not check:** each open item once, highest harm first, with the one fact that would settle it. Then the checks that don't apply, in one line.
5. **This week:** the top fix, in one sentence. If there is no fix, the open question with the highest harm.
6. **Also noticed (optional, two lines at most):** anything outside the voice review you saw in passing that would mislead listeners, such as a promise in the notes that the schedule doesn't keep. It never outranks the checks.
7. **To finish the review:** your questions, at most five, highest harm first, each with why.

## My setup

Paste or describe your setup below, in any order. Whatever you have is enough to start; the AI will ask for what it needs. Most useful:

- where the show is published, where clips or trailers go and where its ads run, with links to the show page and RSS feed if you have them
- whose voice it is (yours, someone else's, a stock voice), the provider and product, the clone type, and whose account made it
- your plan, and whether any published audio was made on a free plan or a beta feature
- what the voice does: whole episodes, fixes, voiceovers, ads
- who writes the scripts, whether AI helps, and whether you approve each one
- your disclosure: the exact spoken line and what plays before it, the episode notes, the show description, your YouTube setting (a transcript or caption file of an episode's opening helps)
- who listens before release (a role is enough), and to which file
- why you chose an AI voice, and any figures you relied on
- where your business is based, and whether people in the EU can hear the show

[Paste your details here]
No signup or paid AI plan required. Use sanitized notes; keep client details and passwords out.
Open prompt link in Claude Desktop
Requires the desktop app. It asks Claude to read this page. If Claude cannot read it, paste the copied prompt.
From Business PodcastingEpisode 2: Should your podcast use an AI voice?
Paste it into your AI with a few lines about your show. It checks your setup against platform rules and voice provider terms as of October 2026, and tells you what to fix first. Three things it catches that most guides miss:
A disclosure that comes too late. If the synthetic voice speaks for two sentences before saying it's synthetic, or your clips and trailers start mid-episode with no line at all, it flags it. Apple sets no position, but the EU's voluntary code puts the line at the very start of the synthetic voice, with reminders after ad breaks.
Rights fixed when each file was made. At least one major voice provider ties commercial rights to the plan you were on when each file was made, so upgrading later doesn't cover episodes made on the free plan.
Rules that don't travel. YouTube says your own cloned voice needs no label; Apple says it does, in the audio and in the notes. Apple also dropped its "material portion" test in 2026, yet some guides, including one voice provider's own, still quote it.
Not launched yet? It turns each check into something to set up before episode one.
How to use it
Copy the prompt into ChatGPT, Claude or Gemini. If yours can search the web, switch that on so it re-checks each rule at its source.
Under "My setup" at the end, paste what you know: where the show goes out, whose voice it is, your provider and plan, your disclosure line and what plays before it. Gaps are fine.
You get what to fix, what's fine, what it couldn't check and the one thing to do this week, then up to five questions to finish the review.
It can only judge your rights if you name your voice product and plan, and it can't hear your audio, so it works from the lines you give it. It flags what needs a lawyer; it isn't legal advice.
If you want a show idea to work with, start from the ones we have sketched by industry, each with a format and first episodes. Browse podcast ideas for your industry
Give the next check a place to live.

Use the project setup to keep your show brief, evidence and decisions together in your own AI.
Set up your AI podcast project
Bring us the decisions you want help making.

The prompt gives you a starting point. Our team can research, plan and produce the show with you, with your approval before every episode publishes.
See podcast production programs Talk through your podcast project