Half the Books on Amazon Weren't Written by a Person

The most rigorous measurement yet: AI-containing books topped half of 2025 releases on Amazon. What the NBER study actually found, what it missed, and what it changes for authors.

The flood is no longer a feeling; it has a number. Half of new releases, tripled output, a measurable quality gap.

The first hard measurement of the AI book flood — what it found, what it missed, and what it changes for authors.

Quick answer: By the most rigorous measurement yet, more than half of the books released on Amazon in 2025 contained AI-generated text. A National Bureau of Economic Research working paper by economists Reimers and Waldfogel found AI-containing books topped 50% of 2025 releases, while monthly book output roughly tripled between 2022 and late 2025 as large language models spread through the author community. The headline matters less than what the study found underneath: AI books measurably underperform on usage, established authors used AI as leverage without abandoning human writing, and the flood is a discoverability problem — not a verdict on the quality of your work.

The study, briefly

The paper — "AI and the Quantity and Usage of Creative Products," NBER working paper 34777 — analyzed roughly 60,000 books sold on Amazon, classifying each release by whether it contained AI-generated text. Its core finding is a supply shock. Monthly book releases tripled from 2022 to late 2025, and the share of new releases containing AI text topped half in 2025 — up from roughly a quarter in 2023, according to reporting on the study. The industry numbers agree. Bowker's ISBN data, reported by Publishers Weekly in March 2026, counted more than four million U.S. titles in 2025 — up 32.5% in a single year — with self-published output (3.5 million titles) up 38.7% and representing roughly 84% of the catalog.

| Measure (U.S.) | Figure | Source | |---|---|---| | AI-containing share of 2025 releases | more than half | NBER w34777 | | Monthly releases, 2022 → late 2025 | roughly tripled | NBER w34777 | | Total titles, 2025 (ISBN) | 4.17M, +32.5% YoY | Bowker / Publishers Weekly | | Self-published titles, 2025 | 3.5M, +38.7% | Bowker / Publishers Weekly | | Self-published share of catalog | ~84% | Bowker / arXiv |

What the study found beyond the headline

AI books are measurably less used. The most useful result: AI-containing books earn worse ratings, sales ranks, and usage — and their rising share is what drives the overall decline in average book quality. That is not snobbery; it is a measured signal.

But the flood is not uniformly worthless. A companion working paper tracking 14,419 Amazon genre-fiction titles found books with substantial AI text made up about 20% of the catalog but earned just 12.1% of sales — while a small minority of high-volume AI producers did reach the top of the market, and a few generated more than a million dollars in gross revenue. The flood is mostly noise; it is not all noise.

Established authors used AI as leverage. This is the counterintuitive finding: previously active authors increased their releases in the LLM era while keeping their non-AI output stable. Authors who were already publishing used AI to produce more — they did not swap human work for machine work. The flood came predominantly from new entrants, and most of their books sold almost nothing.

The consumer-surplus math is cold comfort. The paper's calibration finds AI books raised consumer surplus only modestly — roughly seven percent in the reported estimate — because supply got cheaper. That is a gain for buyers, not for authors. It coexists with a market where most indies earn little: 44% of surveyed indie authors make $100 or less a month (Written Word Media, 2025).

Why the real number is even bigger

If anything, these counts understate the flood. ISBN data misses Kindle-exclusive and Kindle Unlimited titles — the cheapest to produce and the fastest-growing. Detection is imperfect in both directions: some AI text goes undetected, some human text is falsely flagged. Every reliable measure points the same way. Amazon's own response is telling: since September 2023, KDP caps new titles at three per day and requires AI disclosure. Tech journalist Kashmir Hill's experience is the flood's absurd edge: she found an AI-generated, error-riddled biography of herself on Amazon, published by an "author" who released nine other books the same week.

What changes for you

Treat the headline as context, not verdict.

• Discovery is now a volume problem. Your book competes for rank against a catalog that grew by a third in one year. That is why publishing more rarely fixes invisibility — it adds to the already-flooded side of the market. The defensible position is an audience you own (the mechanics are in the audience-building guide).
• The quality signal is on your side. The data says AI-heavy books underperform on usage. A genuinely edited, continuity-consistent book is a real differentiator — exactly where craft and story-memory systems earn their keep.
• Disclosure and trust are table stakes. Readers cannot reliably tell AI prose from human prose, yet suspicion is ambient — romance author Jess Michaels told the New York Times the "glut of the A.I. work makes discoverability very difficult." Honest disclosure (see the KDP disclosure guide) protects you from the scandals eroding everyone's trust.

The reframe

The flood is no longer a feeling; it has a number. More than half of new releases, tripled output, a measurable quality gap — that is the market you publish into. The same data that makes the problem visible also shows the way out: AI is leverage for authors who keep their quality bar and their audience. The flood is real. The defensible position is to be the book readers can find, trust, and remember.