Jenny Williams spent eight and a half years inside Google and Microsoft, including a stint on an AI ethics team. Her new novel, House of Liars, imagines what happens when six writers compete for a million-dollar publishing prize on a remote island - and one of them is an AI. The book arrives at a moment when writers are discovering their own work in AI training data, unpaid, with no legal remedy in sight.
The novel's premise grew out of Williams' time inside the industry. On the first day of the fictional residency, each writer receives an AI-generated version of the novel they came to write. Williams calls it "this kind of tempting object" - a draft that already exists, waiting to be read or thrown away.
Williams is direct about the anger directed at AI companies from the creative world. Her own first novel, The Atlas of Forgotten Places, appeared in the training data of large language models without her consent or compensation. "Whether or not it was legal, whether or not it's termed legal, it's clearly unethical," she said. "These companies were taking lifetimes of labor from writers and artists without compensation."
But she thinks the public conversation misses something important. "One thing that is not acknowledged enough," she said, "is really how tempting and seductive these technologies are."
Four problems, not one
Williams has four separate concerns about AI in publishing, and they don't reduce to a single legal question. First is the training data itself. Second is that the same models now compete with the writers whose work trained them, in the same market, at a different price point. Third is the absence of meaningful disclosure to readers.
That third problem has a concrete example. Amazon's self-publishing platform asks authors to declare whether they have used AI, but the answer never reaches the buyer. The form asks for a single yes or no. As Williams put it: "It's a binary that contains a huge spectrum."
The fourth problem is whether readers actually care. Williams is not sure they do - at least not enough for disclosure requirements to pressure the industry into changing.
The Limits of Disclosure Rules
Some of the disclosure questions are now matters of law. Article 50 of the European Union's AI Act took effect on August 2, requiring providers of generative AI systems to mark their output in a machine-readable format. Systems already on the market have until December 2 to comply.
Anthropic, which makes Claude, has signed the European code of practice and published its approach: an invisible watermark in generated text and a signed record on generated files. By the company's own admission, the mark is "not fully conclusive."
The limits of the system are obvious to Williams: someone who used Claude to proofread a paragraph carries the same mark as someone who fed it a single prompt and got a full manuscript back from it. "I don't think there's a there there," she said of AI writing. "There's no sentience behind it. There's no lived experience."
Williams offers an alternative reason to trust human writers, one that doesn't rest on technical disclosure at all: "I write not because I have answers, but because I have questions."
Why this matters for writers
The seduction extends far beyond a lazy afternoon or a stubborn deadline. Writers who choose not to use AI can already feel the pull of the tools as units of production. The platforms show you the possibilities are there. Williams calls the AI-generated draft "this kind of tempting object."
The distinction between using AI for limited tasks and protecting the position of the writer can dissolve when the line is drawn in front of you. The European rules will mark AI-authored output - but that distinction will deserve a scene if you consider both sides. For a writer, the devastating, practical question is less about which parts you used and more about whether you still own the curiosity behind the work.
For writing professionals, the US publishing industry still needs a position between complete non-disclosure and total capture of the AI question. Until then, the best defense is knowing today's AI and its current failures: AI training on unpaid work, the seduction of a "tempting object" already written, and replacements that don't compel readers with lived experience. That's the space AI for Writers maps out. The gap between the ethics problem of the training data, the narration, and the core tech itself won't close itself. You decide where you stand, and what is - and isn't - worth "compelling" about the way the work is produced.
Your membership also unlocks:
