Most of them don't say. Across the twelve agreements in our Publishing Database, two rule training out in writing: Kobo and PublishDrive. Voices by INaudio bars the uses that matter most for audio, a machine-generated voice or a new work made from your content, unless you give permission. Spotify for Authors has the same carve-out, but its licence also lists "training" and "modeling" among its purposes without saying what they cover. At the other end, Google's auto-narrated audiobook agreement grants a right broad enough to allow it. The other seven say nothing about AI. Silence is not an exclusion. It leaves the question to whatever licence you granted by uploading, and to policies that can change without amending the contract.
Kobo Writing Life. Its terms are direct: training on uploaded work is excluded by the terms — Kobo agrees that works will not be used to train generative AI. The same clause is open about what Kobo may still do with AI: read, analyse and process the work to categorise and tag it, judge whether it is suitable for sale, generate keywords and ad copy from it, and build recaps and accessibility features for readers who have bought it. Those are real permissions. They are also different from training a model on the text, and the contract separates the two on purpose.
PublishDrive. Its distribution agreement is just as plain: the terms state your content is not used for AI training or shared with third parties to train their models. Its research-and-development permission is carved out to exclude AI training on your content. PublishDrive's optional Publishing Assistant does send data to third-party AI providers for metadata and cover help, but only if you choose to use it.
Voices by INaudio does not use the word "training", but its terms close off the two uses that matter most for an audiobook. They do not authorise INaudio or its distribution partners to use your content to create a new book, ebook or audiobook, and the terms bar making a machine-generated voice from your audio without your permission. Either is possible only with your permission.
Spotify for Authors. Its Distribution Terms say nothing about AI, but the content licence in its separate Terms of Use does two things at once. As with INaudio, the terms bar making a machine-generated voice, or a new book or audiobook, from your content without your permission. It also lists "training" and "modeling" among the purposes Spotify may use your content for, next to product development, testing and anti-fraud work, without saying what is trained or modelled. So a cloned voice is ruled out without your consent, while other model training is neither named nor excluded. We record it as unclear rather than as a no. Spotify also offers optional AI narration of a manuscript using ElevenLabs voices, which is a creation tool rather than a use of a recording you supply.
Google's auto-narrated audiobooks agreement is the one record where silence is not the issue. It does not mention training either, but section 3.1(a) takes the book's text and affirmatively lets Google create derivative works of it, both to build the audiobook and for “testing and making improvements to the Google Services”. "Google Services" is defined as any product or technology Google or its affiliates provide or are developing, and the right survives termination. A right to make derivative works of your text to improve technology Google has not yet built is the clearest opening in the database. See Can I make an audiobook with AI narration?
Amazon KDP. The Terms and Conditions do not address training machine-learning models on uploaded content anywhere. It is not permitted, not excluded, not mentioned.
IngramSpark. The agreement is silent in the same way. The licence it grants lets Ingram reformat, edit and modify titles and metadata at its sole discretion, and the metadata rights outlive the agreement, but neither is an AI clause.
Apple Books. No mention either, but a narrow grant around the silence. Apple is appointed as your agent rather than taking a licence as a reseller, and what it may do is listed: sell the book, deliver previews, format and host the file, use your artwork for promotion with modification limited to resizing and cropping. Nothing in that list reads naturally as covering model training. Narrow is not the same as excluded, but it is a different starting position from a general-purpose grant.
Google Play Books. No mention of AI or training in the ebook agreement, but a wide licence. It authorises Google to use and make the content available "as part of and/or through the Google Services (or any of them)", under the same definition that reaches technology not yet built. Separately, Google's Book insights feature is a generative-AI companion that reads the book's text to answer readers' questions. It is on by default; Catalog › Advanced options › AI settings for the whole account, or Book info › Settings › Advanced settings per title turns it off.
Draft2Digital. No mention of AI or machine learning, but the terms grant the right to use manuscript and work files "for the purpose of improving and enhancing our systems for converting and distributing" books. That clause is scoped by purpose rather than by technique, so it neither clearly includes nor clearly excludes model training.
Draft2Digital digital narration. The audiobook agreement incorporates that same improvement clause, and adds its own right to make derivative works of your book "by all means now known or later developed". Neither clause names training, and neither excludes it.
ACX. The audiobook licence does not mention AI, machine learning or training. Under exclusive distribution it grants Audible an exclusive licence over the audiobook "in all formats now known or hereafter invented", which Audible may sublicense to affiliates or third parties. That is breadth without a statement either way.
A contract that says nothing about training has not promised you anything about training. What governs instead is the general licence you granted by uploading, which on every one of these platforms is broad enough to run a distribution business, plus whatever the company's current policy happens to be. Policies are not contract terms. They can be rewritten without your agreement and without notice, and the rewrite is lawful precisely because the agreement left the space empty.
This is why the answer here is different from the answer for books already published and scraped. That is a litigation question about work that has already left your hands. This one is about a document you sign before the file is ever uploaded, and it is the rare part of the problem where reading first actually changes the outcome.
Platforms increasingly run AI over catalogue content for categorisation, keyword generation, ad copy, recaps and accessibility. A company can do all of that and still not train a foundation model on your manuscript. It can also do all of that while quietly reserving the right to. The two questions have to be asked separately, because a press release about one is routinely read as an answer to the other.
No, and it would be a poor reason to. An explicit exclusion is a genuine point in a platform's favour and worth knowing about, but exclusivity, royalty basis, reach and payment terms all matter too, and no single clause settles where a book should be sold. Going wide means accepting several sets of terms at once and knowing which one is weakest, not finding the one perfect platform.
General information, not legal advice. Platform terms change; each assessment carries its own verification date. Records verified 4 and 25 September 2026; Spotify for Authors re-checked 12 October 2026.
Publishing Database
Author Answer