LO
CapabilityCognitive automation2025-10-15
Expert readers who disfavoured prompted AI imitations of authors' styles favoured them once the model was fine-tuned on each author's complete works, a preregistered study found
Writer and authoroccupation page →
Event date / reported
2025-10-15
Evidence stage
CapabilityA demo, benchmark or paper shows the task can be done. Updates what the technology can do — not what employers will do.
Tasks this bears on
Writing fiction and books
Drafting novels, stories and non-fiction books.
Being augmented≈ Platform inference
Rights and licensing
Controlling and licensing the use of one's books and scripts, including for AI training.
Still human-led✓ Evidence-backed
Where this applies
A preregistered blind study comparing MFA-trained writers with frontier models writing excerpts of up to 450 words in the styles of 50 award-winning authors. With in-context prompting, AI text was strongly disfavoured by MFA-trained readers for stylistic fidelity (odds ratio 0.16) and quality (0.13); after fine-tuning ChatGPT on each author's complete works, those readers favoured AI for fidelity (8.16) and quality (1.87), and detectors flagged 3% of fine-tuned outputs against 97% of prompted ones. It measures short excerpts, not novels, and does not count the human effort needed to make publishable books.
What this means
The quality gap depends on training data: a model trained on an author's own books can out-write imitators of that author — which is why the rights over those books matter.
What it does not yet show
Short excerpts in blind comparisons; it does not show AI writing publishable novels or what readers buy.
What you can check
Open the arXiv paper 2510.13939 and find "fidelity (OR=8.16)".
Does it change the assessment?
No — and this stage does not move it either. A "Capability" record is real evidence, but it does not upgrade a task judgement on its own. The 2 linked judgements above stand where they were.
Source
Chakrabarty, Ginsburg and Dhillon — Readers Prefer Outputs of AI Trained on Copyrighted Books over Expert Human Writers, arXiv:2510.13939 (preregistered; submitted 15 Oct 2025) · verified 2026-09-30 · Claude (VOLO agent) · interpreted 2026-09-30 · Claude (VOLO agent)
Primary source — published by the party that did this, or the authority of record. No co-signature needed.