AI-written books squeeze author revenue as Twitch gives creators a late opt-out
A new analysis finds substantial AI text in a fifth of self-published e-books, while Twitch adds a training opt-out. Together, the changes sharpen questions over creator consent, control and income.
Artificial Intelligence··Evening
AI share in self-published e-books and falling revenue per title
An analysis reported by The Decoder examined 14,419 self-published e-books released between January 2023 and March 2026 and classified 20 per cent of the catalogue as containing substantial AI-written text. Those titles took 12.1 per cent of sales and 11.3 per cent of revenue, while revenue per book fell for non-AI titles in 7 of 8 genres. The dataset came from one of five major US publishers and tracks about 500,000 Amazon titles, covering roughly 95 per cent of daily e-book sales; full texts were classified with the Pangram v3.3 detector, reported at a 0.04 per cent false-positive rate. Over the same period the catalogue grew 38.3 times while quarterly revenue grew 8.9 times. The share of new Top 25 entries with substantial AI text rose from near zero to 31 per cent, and the highest-grossing single AI title took 643,000 dollars on 80,431 copies. The figures place faster production and weaker per-title income for human authors on the same shelf. Classification rests on textual AI signals and sales data rather than naming every writing tool, yet the result shows substantial AI text reshaping how revenue is shared in the self-published e-book market.[1]
A training toggle that arrived on by default at Twitch
According to WIRED, Twitch has added a Generative AI Training toggle under Security and Privacy, letting streamers stop posts, streams and videos from training Amazon's AI models. The setting was on by default, and the platform has not said when that use began. More than 16,000 creators objected in a forum thread after the change, which came to light when head of community Mary Kish explained it on a livestream. Head of product Mike Minton said a default-on option was necessary because otherwise no one would participate, and said he assumes almost any public content is used for training with or without permission. Twitch's terms of service, in effect since March 2024, grant broad rights over user content but do not name generative AI training, and the toggle does not stop other uses described in the privacy notice, among them recommendations and AutoMod. The opt-out does not settle consent for the past; it only limits training use going forward. How long the setting stayed on and which models already ingested older streams remain open; Twitch has not published that timeline.[2]
Consent and income on the same table
The two developments move in different venues but sharpen the same tension: a creator's work can compete with AI text on the sales shelf while also sitting, by default, inside a platform's training use. The e-book analysis shows per-title income weakening in independent publishing and AI text climbing the top lists. The Twitch side shows consent arriving late through an incomplete switch, terms that do not name generative training, and a toggle that leaves recommendation and moderation uses untouched. Control over who processes content and who benefits is no longer only a copyright argument; it is a daily platform setting and a market-share question. On a shelf where human authors' revenue per book is falling, a streamer's default-on training switch puts consent and income on the same table. When control is granted, which past material it covers, and how per-title income recovers remain open; today's reporting makes those questions visible without closing them.[1], [2]
