Kobo says it will not train models on authors' books

The writer's problem

An author choosing where to publish an ebook wide increasingly asks a question retailers rarely answer directly: does uploading a manuscript feed a company's own AI models? Rakuten Kobo addressed this directly in a 29 May 2025 Kobo Writing Life blog post, stating the company's position on AI use across its self-publishing platform.

What the documents show

The post states two non-negotiables: 'We don't feed books into any LLM' and 'We don't use books to train any LLMs.' This is Kobo's own claim about its own systems, not an independent audit. The same post says Kobo does use narrower machine-learning tools built in-house: models to identify and quarantine content such as hate speech, which the post says only assist a human review team and are 'not shared with outside companies'; algorithms that analyze samples of book text, not full manuscripts, to improve search and categorization; and pattern-detection tools aimed at catching pirated or AI-mass-produced content. Separately, Kobo's own content policy bars 'Automated Content Generation,' defined as content generated primarily by automated tools that 'lacks genuine human effort, quality control, or is designed to exploit trends rather than offer genuine reader value.' Read together, the two documents describe a boundary: Kobo says it does not use uploaded books to build generative systems, permits its own narrower internal AI tools, and separately polices books that are themselves mostly AI-produced with little human effort.

The editorial choice

What counts as 'primarily' automated is Kobo's own call, applied case by case rather than a bright-line test in the policy text. An author or editor relying on this distinction should treat having used AI at some stage as insufficient information either way; the policy concerns how much of the finished book's value comes from unedited AI output, an editorial judgment about the finished work rather than a tool-by-tool inventory.

What stays with the author

Kobo's statements do not extend to what other companies in its supply chain, distributors, or the audiobook side of its business do with content; the commitment is scoped to Kobo's own systems as described in its own blog. Nor does the policy define a numeric threshold for primarily automated content, leaving borderline cases, heavily AI-assisted books with substantial human revision, to Kobo's own review rather than to a rule an author can apply in advance.

• Does this claim describe Kobo's own systems only, or does it extend to partners the book is also distributed through?
• How much of the finished manuscript's content and structure came from unedited AI output, versus human revision?
• Has Kobo's policy page been checked at submission time, since a living document can be updated?

Kobo's language is more specific than a general AI ethics statement, but it remains a platform's description of its own practice rather than an independently audited claim.