AI Is Rewriting Publishing's Social Contract — and Nobody's Holding the Pen

One-fifth of recent self-published books may contain 25%+ AI text. The industry has no disclosure standard, no reliable detection, and no consensus on what a 'book' even is anymore.

Coral Hart has written 370 novels under 21 pen names since 2025, using AI to operate at a scale no single human writer can match. She earns high six figures on Kindle Unlimited, where the algorithm rewards volume — a new book every month is the floor, not the ceiling. Hart's framing is that AI removes friction, not craft. The writer still provides structure, voice, and reader instinct; the machine accelerates execution. That framing is self-serving but not entirely wrong, and it describes where a large chunk of commercial self-publishing has already arrived.

The structural engine underneath Hart's model is Kindle Unlimited's per-page-read payment system. It was designed to reward engagement, not just sales — a rational anti-spam mechanism. What it didn't anticipate was that AI would collapse the marginal cost of producing an engaging-enough page to near zero. The algorithm now functions as a subsidy for volume producers who can maintain minimum reader retention. Hart represents the leading edge of that dynamic, not an outlier. A preprint study led by Tuhin Chakrabarty of Stony Brook University estimates that roughly one in five recently self-published books contains at least 25% AI-generated text. That number is almost certainly an undercount — it reflects detectable AI, not all AI. The actual penetration is unknown, and the industry has no disclosure infrastructure to find out. Hart's advice to authors is explicit: don't disclose unless legally required, because disclosure triggers harassment. That is not a stable equilibrium. The enforcement mechanism that has emerged instead is adversarial and chaotic. Journalists and social media users run manuscripts through AI detection software — primarily Pangram, which claims 95%+ accuracy and near-zero false positives — and publishers respond by cancelling deals. Hachette withdrew Mia Ballard's Shy Girl. A $2m debut deal for Jerry Falade collapsed. Thélyson Orélien was struck from France's Goncourt longlist. Every writer at the center of these scandals has been an author of colour, a pattern Falade named directly: the investigative zeal, he wrote, appears reserved for writers who already face elevated scrutiny. That disparity is real, unresolved, and sitting in plain sight. The legal terrain is equally unsettled. A class action involving around 500,000 titles won a $1.5bn settlement from Anthropic for alleged training-data infringement. Other suits are ongoing. Seventy prominent authors signed an open letter demanding the Big Five publishers commit to never releasing machine-created books. The publishers did not commit. The asymmetry there is clean: the writers have moral standing and no leverage; the publishers have leverage and are waiting to see which way the market breaks. The detection infrastructure itself introduces new injustice vectors. Evidence already suggests that non-native English speakers and neurodivergent writers are more likely to trigger false suspicion — not because the tools are biased in design, but because those writers' prose patterns diverge from the training distribution that defines 'human.' Pangram's near-zero false-positive claim is a statistical aggregate; at the level of individual authors in specific cultural and linguistic registers, that confidence interval expands considerably. The industry is treating a probabilistic tool as an adjudication mechanism, with career-ending consequences. What this looks like in 20 years depends entirely on whether disclosure becomes a legal norm in the next five. If it doesn't, the self-publishing ecosystem completes its transformation into a volume-production pipeline with human brand management on top. Mainstream publishing either develops credible verification infrastructure or cedes the 'authentically human' positioning to a small prestige tier — with everything else operating in the same grey zone as Hart's 21 pen names. The writers who lose are the ones for whom craft is the entire value proposition and who lack the capital or appetite to operate at algorithmic scale.