Earmail uses AI throughout the process of turning written text into high-definition audio, but it never uses AI to replace an author's body prose. We think this distinction is critical, so here's exactly how we use AI.
What the AI does
- Adapts the piece for listening. We remove ads, tracking pixels, and footers; detect sections and headings; and check pronunciations. We do this under a strict fidelity rule: the AI model is instructed to never rewrite, paraphrase, summarize, condense, reorder, or shift the meaning, tone, or point of view of what the author wrote. It prepares the author's words for listening — it does not replace them.
- Narrates it. An advanced text-to-speech model reads the adapted text aloud. The narration is synthetic and we say so clearly — episodes are narrated by Earmail, not by the author, and we don't clone anyone's voice.
- Generates the extras you ask for. Earmail's weekly digests, reflection prompts, briefs, and Talk Back answers are AI-generated by design. Each of these is opt-in or starts with a question you ask. None is ever presented as the author's work, and none is published anywhere.
Where Earmail's own words enter the audio
The article itself is always narrated as written, and Earmail's own words appear in a few marked places:
- The introduction. Each episode opens by announcing the title, source, and original publication date.
- Chart and image descriptions. With “Describe Charts” toggled on, charts get a short spoken description we generate. It's description, not interpretation — but the sentences are ours.
- Footnote asides. With “Narrate Footnotes” toggled on, a substantive footnote is folded into the narration as a short spoken aside, which may mean rephrasing it to work alongside the rest of the body copy.
What the AI never does
- Write or publish anything publicly. Earmail produces no posts, no articles, no social content — nothing in a user's name or an author's name. It narrates episodes into a private feed that only the intended user can access. There is no directory, no sharing surface, no public artifact carrying an author's work.
- Summarize the original article when you send it. Narration is faithful to the original full text. (Note that Earmail digests and briefs condense — but these are separate, clearly labeled features each user must opt into.)
- Train on a user's mail. Content is processed by Anthropic (text) and OpenAI (speech) under API terms that exclude training. See our Security & Privacy page for full details.
If you write a newsletter
When your writing passes through Earmail, it's because one of your subscribers forwarded an issue from their own inbox to their own private feed. They hear the whole piece, in order, in your words — not a summary, not a remix. Nothing is republished, nothing carries your name anywhere public, and nothing is used to train a model. Earmail's job is to get your full piece heard by a reader who would otherwise have left it unread.
Questions about anything on this page? Write to hello@earmail.app.