Rakuten Kobo rejected nearly half its self-published submissions last year over suspected AI use

Rakuten Kobo turned away 45 per cent of the books submitted to its self-publishing service, Kobo Writing Life, last year, and the company estimates that roughly 80 per cent of those rejections came down to suspected AI generation, a scale that would have been almost negligible only a few years ago.

Speaking at the CONTEC Buenos Aires conference in April, Tamblyn described the volume as overwhelming, comparing the wave of new "publishers" appearing on the platform to undisclosed groups of language models working in concert rather than genuine small presses with catalogues to build.

The company has long run a content pipeline to catch duplicates and enforce its existing guidelines around hate speech and medical misinformation, but assessing how a book was actually written marks a newer and thornier task.

More from Lifestyle

Detection software remains unreliable in both directions, prone to mislabelling human prose as machine-made and the reverse, and Kobo has been wary of letting flawed tools make calls that could damage a working author's reputation or sales.

The platform is preparing another round of testing on detection systems, though Tamblyn has been candid that distinguishing human and machine writing is growing harder rather than easier as the technology improves.

A flood that changed the rules

Industry-wide figures help explain the scale of the shift. Bowker data showed self-published fiction ISBNs in the US climbing from roughly 307,000 in 2024 to more than 477,000 the following year, a jump that publishing analysts have linked, even if not definitively proven, to generative tools lowering the barrier to producing a manuscript.

Kobo's own examples of the worst offenders include a title that simply compiled ten thousand apple pie recipes scraped from across the internet and presented them as a book.

For Tamblyn, the concern extends beyond protecting Kobo's customers from a poor purchase. He has framed the platform's role as one link in a longer chain running from author to editor to publisher to bookseller to reader, and argued that allowing the store to fill with machine-generated volumes makes it harder for the kind of original, harder-to-produce writing that the industry depends on to find an audience at all.

Authors who view AI training on their work as a form of theft have also shaped Kobo's thinking, a sentiment Tamblyn has likened to someone dismantling another person's furniture and selling it back to them in pieces.

When even a Nobel laureate gets asked

The disclosure question Tamblyn raises has already played out at the very top of literary fiction. Olga Tokarczuk, the 2018 Nobel laureate, set off a storm in May after remarks at the Impact Congress in Poznań were read by many as an admission that her forthcoming novel had been written with AI assistance.

The backlash was swift across Polish and international literary circles, and Tokarczuk issued a clarification within days, insisting through her publisher that she had written the book entirely alone over decades of solitary practice and had used a language model only for preliminary research tasks such as fact-checking or generating period-appropriate details for her characters.

She also spoke of a genuine sorrow at watching a slower, more solitary mode of writing recede, even as she defended treating AI as a tool no different from a search engine or an archive visit.

The episode underlines exactly the grey zone Kobo is trying to police at a much larger scale: the gap between AI as a research aid and AI as the author of the prose itself. Tamblyn expects more publishers to start asking authors for clearer disclosure about any AI involvement in a manuscript, which would eventually let Kobo hand some of that judgement to readers rather than deciding unilaterally what belongs on the virtual shelf.