"Disclose that the narration is synthetic" has been store policy for a while, and the way an author met it was a line in the description and a credit at the top of the book. This autumn the duty moves into the file itself. The EU AI Act's Article 50(2) requires providers of AI systems that generate audio to mark their output in a machine-readable form; it has applied to new systems since 2 August 2026, and systems already on the market have until 2 December. California's AI Transparency Act asks for the same thing in its own words: a latent disclosure embedded in every piece of generated audio, and a free tool anyone can use to read it. An amendment signed on 30 September removed that law's user threshold, so it reaches small providers as well as large ones.
For an author distributing an AI audiobook, the practical question is no longer whether to disclose. It is whether the file can prove you did.
Since 14 October, every audio file Midsummerr renders carries those marks, and anyone can check any file, free, at midsummerr.com/ai-marks.
What is in the file
Three marks, each doing a different job.
An inaudible watermark in the sound itself. This is the durable layer. In our tests it survives MP3 and AAC re-encoding, volume changes and EQ, which is roughly what a file goes through between your download and a listener's phone. Time-stretching removes it, and the page says so rather than pretending otherwise.
Content Credentials. Signed, time-stamped metadata in the open C2PA format, naming Midsummerr as the system that made the audio and recording when. Any Content Credentials tool can read them and verify them against the certificate we publish. Re-encoding a file strips its metadata, as it strips any metadata; that is normal, and it is the reason the watermark sits underneath.
A plain tag in MP3 files, for software that reads tags and nothing else.
The marks identify Midsummerr as the system. They carry nothing about you: no name, no account, no book title, no chapter. The only identifier inside any of them is a random one that means nothing outside our own records, so a marked file says "made with Midsummerr", not "made by this author".
Ready to try it on your own book?
Start your first chapter free →What changes in your production
In the studio, nothing. You cast, direct, edit and mix exactly as before, and previews played inside the studio are not marked. The marks go on when a file is rendered for delivery: the final mix of a chapter, the WAV master, the retail export, and the stems, which carry Content Credentials. Every file rendered from 14 October carries them. A chapter you downloaded before that date is marked the next time it is remixed or exported, so if you want an older download marked, export it again.
When a production mixes human-recorded audio with AI voices, music or sound effects, say your own narration with a produced score under it, the marks say the file contains AI-generated elements rather than that it is AI-generated outright. The distinction lives in the signed metadata and the tag, and the checker reports it.
The Download Centre now shows a short note beside every download, with a link to the page that explains the marks in full. Nothing else there has moved.
Checking a file, and keeping the receipt
Go to midsummerr.com/ai-marks, drop in an audio file or paste a link to one, and press Check. The result says whether the file is marked as AI-generated or as containing AI-generated elements, which of the three layers it found, and gives you a signed receipt: the file's hash, the detector version and the time of the check. The file is used only to look for marks and is deleted immediately.
Save the receipt alongside the master. "Can you show this edition is disclosed?" is a question that will come from retailers, distributors and, if you work with one, your publisher's compliance desk, and a dated receipt that verifies against a public key is a better answer than a screenshot of a description field. The same check works on a file that has already been through a store's encoder, because the watermark survives that.
One reading to get right: "no mark found" means no Midsummerr mark was found. It does not mean the audio is human-made, and the checker never says that it is.
What it is not
It is not a store label. Where a platform or the law requires you to tell listeners that the narration is synthetic, that duty stays with you as publisher; the file now stands behind the statement rather than replacing it. The listener-facing side, what to say in the description and the credits, is in what an AI narrator should tell listeners, and the store-by-store rules are in where to distribute an AI audiobook in 2026.
It is not a guarantee that any particular retailer's check will accept a file. Retailers run their own checks and change them. What we promise is that the marks are there, that we publish exactly what they are, and that anyone can verify them without an account.
And it is not optional. Our Terms now ask you not to remove, alter or defeat the marks. Ordinary encoding, mastering or format conversion that drops metadata as a side effect is not a breach; stripping the marks on purpose is.
For publishers and API partners
There is nothing to integrate. Every file the Partner API delivers carries the same marks, and each delivery response describes them, so a catalogue system can record the disclosure state of every chapter it ingests. The Partner API Terms set out who holds which duty under the two laws, and if you make the output available in California, the notice the Transparency Act expects a licensee to receive is in those terms too.
Why a public checker
A transparency claim nobody can test is only a claim. The checker is open to anyone, the certificate that signs the Content Credentials is published on the same page with its fingerprint, receipts verify against a published key, and the known limits of each layer are written next to it. That is the standard we would want from anyone else's audio, so it is the one we hold our own to.
None of this changes what you hear. Dracula is a full-cast production with an original score and sound design, and a marked file of it sounds exactly as it did last month. Download a chapter of your own, drop it into the checker, and keep the receipt.
