AI and Creative Industries

LLMs have tripled the volume of new book releases since 2022 — but average quality has fallen while the upper tail remains intact, producing a “quantity shock with quality sorting” pattern that raises consumer surplus in aggregate but concentrates low-quality output among new AI-era entrants.

What It Is

Reimers & Waldfogel (NBER, January 2026) provide the first large-scale empirical study of LLMs’ impact on a creative market: book publishing. Using a dataset of 333,000+ ebook releases on Amazon from 2020–2025, they document: (1) new title releases nearly tripled between 2022 and late 2025, with some categories growing 10x; (2) average quality of LLM-era vintages is lower than pre-LLM vintages; (3) the top 1,000 monthly releases per category are of higher quality than before (though the top 100 are unaffected); and (4) new authors entering since the LLM influx predominantly produce low-quality work, while pre-LLM authors have accelerated their output at maintained quality.

The authors use a calibrated demand model to estimate that LLMs could raise consumer surplus in book markets by one quarter to one half in steady state — driven primarily by the large increase in the number of new products rather than by average quality improvement.

Why It Matters (for the Economy)

The book market is a useful early signal for how LLMs reshape creative labour markets more broadly. Several dynamics are likely to generalise: (1) entry costs fall sharply, creating a flood of new producers; (2) incumbents with established skills use LLMs as productivity tools and maintain or improve output quality; (3) new entrants using LLMs as a substitute for skill produce lower-quality work; (4) market structure bifurcates between high-quality incumbents and a large tail of mediocre AI-assisted work. This pattern has direct implications for any knowledge work where output quality is difficult to verify before consumption (journalism, consulting, legal documents, code).

Evidence & Examples

  • New book releases tripled from 2022–2025, with coincident timing matching LLM diffusion; some categories grew 10x (w34777.pdf)
  • Authors entering in the LLM era disproportionately produce low-quality work; authors who debuted pre-LLM and are using AI tools have accelerated pace of releases while maintaining higher quality (w34777.pdf)
  • Quality of top 100 monthly releases per category is unaffected by the LLM influx — the upper tail is robust; the top 1,000 releases are of higher quality, suggesting LLMs help more mid-tier work into the market (w34777.pdf)
  • A 2025 survey of 1,229 authors found 45% using AI to assist with their work; among non-users, >75% considered AI use “unethical” — indicating strong incumbent identity effects (w34777.pdf)
  • Nested logit calibration: consumer surplus in book markets could rise 25–50% in steady state, driven by volume increase even with average quality reduction (w34777.pdf)

Tensions & Open Questions

  • Is books generalisable? Books have low barriers to LLM-assisted production (pure text) and relatively easy quality measurement (ratings, reviews). Creative products where quality is harder to measure at point of sale (consulting decks, financial models, legal documents) may see more severe quality sorting problems — buyers cannot filter low-quality AI output as easily as readers can skip a poorly reviewed book.
  • The skill formation feedback loop: If new writers enter via LLM assistance without developing writing craft, the pipeline of future high-skill human authors may thin. The “incumbent quality advantage” documented by Reimers & Waldfogel depends on pre-LLM authors continuing to be productive — a one-generation assumption (see AI Skill Formation and Deskilling).
  • Platform and curation dynamics: Amazon’s rating system allows quality filtering, but other creative markets (music, visual art, journalism) have different curation mechanisms that may not handle the quantity shock as effectively.
  • Other creative sector evidence emerging: A 2025 scoping review (Springer; not yet in Raw/) examined GenAI adoption across 57 studies covering visual art/design, writing, performing arts, and spatial design (2022–2025). Key findings: shift from creation to curation and meta-creation; new literacies (prompt engineering, AI evaluation) emerging; expertise hierarchies being reconfigured — pattern consistent with book market findings. Music-specific: a 2025 survey of music practitioners found AI technology acceptance correlates with creative performance, but socioeconomic factors (age, gender, experience) influence uptake. No direct quantity-shock equivalent to the book market has been documented in music. [@cremieuxrecueil thread in Raw/ provides a signal in a related domain: AI cover letter generators destroyed the predictive value of cover letter quality for job hiring, suggesting AI-induced signal destruction in human-capital markets.] [source needed — web search findings, not yet fully ingested]
  • 🔴 TODO (NARROWED): Pattern of “quantity shock + quality sorting” from books appears likely to generalise. Evidence in music and design is emerging but no book-market equivalent empirical study in these sectors yet. Priority: seek controlled study in design or consulting output.

AI Labor Displacement and Augmentation · AI Skill Formation and Deskilling · LLM Commoditization · AI Native Software Development