Skip to content
April 16, 2026 | Empirical Study

Can good writing be generative? Expert-level AI writing emerges through fine-tuning on high-quality books

Chakrabarty Tuhin, Dhillon Paramveer S.

ai-authorship creative-writing quality-assessment fine-tuning expert-evaluation human-ai-collaboration

What This Paper Found

Researchers at Columbia University ran one of the most rigorous tests of AI creative writing to date. They recruited 28 MFA-trained writers — people with graduate degrees in creative writing — and asked them to emulate the style of 50 critically acclaimed authors. Then they had three large language models attempt the same task. The results were evaluated blind by 28 expert judges and 131 lay readers.

With standard prompting (giving the AI a few examples of an author’s style), expert judges preferred the human writing 82.7% of the time. No surprise there — a prompt isn’t a deep understanding of voice. But when the researchers fine-tuned the models on the authors’ complete published works, the preference flipped: experts chose the AI-generated writing 62% of the time. Lay judges preferred the AI writing across all conditions.

The most striking finding wasn’t quantitative — it was qualitative. In post-study interviews, the MFA writers described experiencing something the researchers called an “identity crisis.” Seeing AI output that expert judges rated higher than their own work eroded their aesthetic confidence and forced them to question what “good writing” even means. These weren’t casual hobbyists; they were professionally trained writers confronting a machine that could emulate literary voice more convincingly than they could.

Why This Matters for Your Family

This paper isn’t about ADHD, autism, or co-parenting. It’s about a question some readers have about this book: how much of it did the human write?

The answer is: the voice, the lived experience, the 3am meltdowns, the co-parenting negotiations, the late-diagnosed recognition of your own neurodivergence — that’s all human. What AI contributed was structural discipline (ensuring every chapter follows the same skeleton), citation verification (cross-checking 1,200+ references against clinical guidelines), metaphor consistency (keeping the nautical vocabulary coherent across 71,000 words), and localization (adapting the manuscript for four national editions).

This study demonstrates that AI can produce stylistically convincing prose — prose that even trained experts prefer. But stylistic emulation is not the same as autobiographical authenticity. An AI fine-tuned on Ross Greene could produce sentences that sound like Ross Greene. It could not produce sentences that come from being a late-diagnosed ADHD parent sitting in a car park after a school meeting, shaking.

The distinction matters. If you’ve wondered whether an AI-assisted book can be “real,” this paper suggests the answer is more nuanced than either side of the debate admits. AI writing can be good. What makes a book yours is something else — something this study’s MFA writers felt slipping away when the machine outperformed them, and something they couldn’t quite name.

What You Can Do Today

  • Read our Appendix A (“AI and Authorship”). It explains exactly how AI tools were used in this book — not as ghostwriters, but as research crew. The voice is human; the verification is machine-assisted. This paper provides the academic context for why that distinction matters.

  • Talk to your co-parent about AI tools. Many families use AI for IEP letter drafting, therapy note summaries, or medication-tracking apps. This research confirms that AI-generated text can be high quality — but it also confirms that human judgment about what matters remains essential. Use the tools; don’t outsource the decisions.

  • Consider what “authentic” means in your parenting. The MFA writers in this study lost confidence when a machine matched their craft. Parents of neurodivergent kids know a version of this feeling — when a therapist, teacher, or book seems to understand your child better than you do. The answer is the same in both domains: expertise and lived experience serve different functions, and you need both.

The Original Paper

Chakrabarty, T., & Dhillon, P. S. (2026). Can good writing be generative? Expert-level AI writing emerges through fine-tuning on high-quality books. Proceedings of CHI 2026 (to appear). https://doi.org/10.48550/arXiv.2601.18353


This research summary is for informational purposes only. It does not constitute medical, psychological, or professional advice. If you or your child are in crisis, contact your local emergency services or the 988 Suicide & Crisis Lifeline (US), Lifeline 13 11 14 (AU), Samaritans 116 123 (UK), or Need to Talk? 1737 (NZ).

Original Source

Coming Soon