AI-written books are taking sales in self-published catalogues as Google makes its visible marker optional, Anthropic prepares a hidden text signature, and researchers propose generation-time stand-ins for copyrighted characters.
Artificial Intelligence··Night
AI-written books gain a place in the catalogue
AI-written work is no longer marginal in self-published e-book catalogues. An analysis of 14,419 titles released between January 2023 and March 2026 classified 20 percent as containing substantial AI-written text. Those books took 12.1 percent of sales and 11.3 percent of revenue, while revenue per book declined for non-AI titles in 7 of 8 genres. Among new Top 25 entries, the AI share rose from near zero to 31 percent. The highest-grossing AI title earned 643,000 dollars from 80,431 copies. Over the same period, the catalogue expanded 38.3 times while quarterly revenue grew 8.9 times. The data came from one of the 5 major US publishers and covers roughly 500,000 Amazon titles, said to represent about 95 percent of daily e-book sales. Full texts were classified with Pangram v3.3, whose reported false-positive rate is 0.04 percent. The method does not directly document how each book was made; it relies on a detector. Even so, catalogue, sales and revenue shares give synthetic text measurable economic weight in publishing.[1]
Visible marks recede as hidden signatures spread
Provenance is shifting from visible logos towards machine-readable signals. Google’s Media watermark setting in Gemini and Flow lets users remove the sparkle from images, video and music made with Nano Banana and Omni. Outside countries requiring a visible watermark, SynthID signals and C2PA metadata remain. Users can ask Gemini or Search whether an invisible marker is present. OpenAI and Meta also avoid visible watermarks. Anthropic is carrying this approach into word choice: in future Claude models, a cryptographic key will replace randomness when choosing among equally suitable words, and a third-party detection interface is planned. Models released before 2 August 2026 are due to receive the method over the coming months. The mark performs poorly on small samples, thins in fact-heavy passages and disappears when text is rewritten throughout. It also cannot prove human authorship. Invisible signals keep the surface clean while moving verification into another tool. Readers see less warning directly, and provenance checks depend on services able to recognise the marker.[2], [4]
Protection moves into generation itself
Provenance signals ask where content came from; new work on copyrighted characters changes what can be generated. Without altering the base model, the method inserts an anchor into the text representation as a stand-in for a target cartoon character. It preserves scene structure while removing identifying details. Erasure strength is adjustable, several characters can be substituted together, and the method transfers between models. Tests used Stable Diffusion and the newer Z-Image architecture. The approach seeks to avoid damage from deleting model parameters and the ease of bypassing hidden negative prompts by describing a character without naming it. The results are the authors’ measurements, and journal peer review is not reported. Growing AI-written book catalogues explain why separating synthetic work after publication is becoming harder. Google’s and Anthropic’s signals add provenance after generation, while the stand-in tries to prevent a copyrighted feature from appearing. One labels origin; the other changes the output boundary. As volume rises, protection is becoming layered across generation-time constraints and later provenance checks.[1], [2], [3], [4]
