AI Use
Podcasting is a medium for humans. Pinecast respects that.
You pay us to host your show as best as we can, not to replace you. AI at Pinecast is opt-in only.
When and why AI
Our philosophy is that AI should do the work that you can't or don't want to do. Transcribing audio is a great example: it's a slow, laborious process that's easy to get wrong. We give you the option to offload that work to the machines.
We also believe AI is a powerful accessibility feature. For instance, blind podcasters can have a difficult time preparing cover art for their shows. Having the option to generate simple, tasteful artwork can help these folks overcome an annoying but necessary obstacle to getting their show set up.
Responsible choices
We prefer to use only what we need and nothing more. Most tasks don't require massive frontier models from the top AI labs: using open models allows us to choose vendors that prioritize privacy and energy use over sheer performance.
No training, full stop.
We don't train AI models on your content. Never have, never will. Your content isn't for sale, and we're not trying to make more of it without you.
We stand by this: we do our diligence to ensure that all of the vendors we use to host your show comply with this policy as well: no training, full stop.
You have the choice
Some features on Pinecast use AI, like transcriptions. If you don't choose to run them, they don't get used. You're always presented with a choice.
Note: We use AI features behind the scenes to detect spam and protect our services from abuse. We choose vendors that handle your data responsibly, and it's not provided with any information that isn't otherwise made public.
What features use AI?
| Feature | Model and vendor | What it receives | When it runs |
|---|---|---|---|
| Episode transcription | AssemblyAI | The episode's audio | When you choose it for an episode. A podcast setting can choose it for you on new episodes |
| Listener feedback insights (Growth add-on) | Gemma from Google, a model running on Cloudflare | The text of your listeners' feedback and your show notes, with email addresses removed | Only when you ask |
| Cover art (paid plans) | FLUX.2 [dev] from Black Forest Labs and Gemma, both running on Cloudflare. | Your podcast's name and description, and the words you type about the look. The model that describes a cover gets only the image | Only when you ask, while you create or edit a podcast |
AI behind the scenes
Pinecast uses AI to check new episodes for spam and AI-generated audio. On a free account, a check can suspend the account automatically. Paid accounts are never suspended automatically. If your account was suspended and you think it's a mistake, email support@pinecast.com and a person will review it.
These checks are the only places where AI acts on its own. We see each automatic suspension as it happens, and we look into every mistake.
- The spam check sends the public text of a new podcast and its episodes, the domain of the owner's email address (never the address itself), the plan, and the sign-up date to gpt-oss-120b, a model from OpenAI running on Cloudflare.
- The AI-generated audio detector is an open-source model that runs on our own servers. Your audio isn't sent anywhere for it.
What our vendors keep
AssemblyAI deletes the audio and the transcript about a day after it finishes. Cloudflare doesn't keep what we send once it has answered. Neither one trains AI models on it.
We keep your transcripts as part of your podcast, and feedback insights for 48 hours. For cover art, we keep the images we make for you, like any image you upload. We don't keep the words you typed, or the description of the image.
AI agents that you connect
We allow you to connect your own AI agents to Pinecast with an MCP server. No AI runs on our side for this MCP server: only the service you use for your agent runs it. Ask your AI service provider if they store data: we have no visibility into this.
When you use our MCP server, it comes with built-in security. You can specify whether an access token has read, read+write, or read+write+delete permissions. Use only the access level that your agent needs to do the job you have for it. You're responsible for the security of your agent and the access tokens you might create for it.
Managing risks
We have three rules for managing the risks of AI:
- Pinecast doesn't rely on AI for any core functionality by default.
- If that ever changes, we'll tell the people who use the feature about its risks and remind them over time.
- Features on Pinecast that use AI never take destructive actions, like deleting something for good or moving money.
These are the main risks we watch for, and what we do about each one.
| Risk | What we do |
|---|---|
| A transcript gets words wrong | You can edit any transcript in your dashboard, or remove it. A podcast setting hides transcripts from listeners. |
| Feedback insights misread your listeners, or suggest a change you shouldn't make | They run only when you ask, they're labeled as AI, each point is tied to something a listener said. Insights are only advice: nothing on our end acts on them. |
| Generated art misspells your podcast's name or shows something you didn't ask for | You see each cover before you use it, and you can make another one or upload your own. The artwork is described with text when possible. |
| The description of generated artwork is wrong | The model tries to quote any text on the image exactly as it's written, so a misspelled name shows up as misspelled. It's labeled as written by AI. |
| Real people or their art can be mimicked by generated artwork | The window warns you not to ask for that, and our terms don't allow it. |
| A free account is suspended by mistake | Spam checks run multiple times to avoid false positives. We review every suspension and a human reviews appeals. |
| Someone hides instructions for the AI in podcast text or imagery | Our models can't use tools or reach any data. Model output isn't used in a way that can lead to unintended actions. |
| Personal information reaches a model | Each feature sends only the data it needs. Email addresses are partially redacted. |
| A model or its behavior changes without warning | We pin each model to an exact version, and change it the way we change any of our code. |
We can turn off any AI feature within minutes. We review this page and the practices behind it every year.