Writing a book with an AI book designer
A publisher-led AI book workflow combines the author’s subject knowledge, editorial direction and generative writing support. Planning, revision and human responsibility remain part of the work.
The publisher’s framework
In May 2024, Springer Nature’s AI book designer was a prototype being piloted across disciplines, subject areas and product types. It followed requests from authors and external editors about writing books with ChatGPT after its arrival in 2022.
The publisher’s five AI principles framed this work.
- Use AI for positive purposes with humans first.
- Mitigate bias and promote fairness.
- Be transparent.
- Keep humans involved and give them final approval of published content.
- Safeguard personal privacy.
Preparation, drafting and quality assurance
The publisher’s 2023 experiment brought together innovation management, AI, rights, public relations and editorial staff with authors, including an author who combined subject expertise with an innovative mindset. German television and newspaper journalists observed the work directly, while the team began without knowing what the result would be.
Months of preparation covered the book’s design, contents and subject before a hack day produced a raw manuscript of about 180 pages. During that day, an AI designer and an author worked on one side of the table, editors and authors on the other, exchanging prompts between them. Weeks of quality assurance followed; this was not a process in which one button produced a finished book. The book was produced faster than a conventional book, and about four pages in it explained the approach. The experiment attracted internal and external attention, was presented at the Frankfurt Book Fair and resulted in the German-language book being listed internationally for a book award.
The team treated academic book writing as a design process shared by authors, editors and AI. It reviewed each stage for possible AI support, identified parameters applying to book writing with or without AI and described the individual steps in detail. Combining the parameters and detailed steps was intended to support a fully automated workflow, although much of it remained manual in May 2024.
The author’s knowledge and editorial control
The prototype targeted people whose professions supplied knowledge worth sharing but who were not professional writers, including chief executives and biologists who lacked time, structure or enthusiasm for writing. Authors were invited to use AI if they could write the book without it. The workflow divided conventional book writing into seven major steps with many smaller steps and feedback loops within them.
The initial design stage collected background, target readership, title, a possible short description and keywords, as in the publisher’s usual author surveys. Information from the author and editor set the boundaries within which the AI could operate, with the aim of preventing hallucinations. Every output was editable in a field comparable to a word-processing file, and authors were obliged to change and add content rather than merely allowed to do so. The author supplied knowledge, while the AI supported structure and writing. Receiving the manuscript in its final format also saved production time.
Author and editor worked together with the AI, retaining human approval before anything reached an audience and human responsibility for the outcome.
Prompts and interface design
An AI expert developed the prompts and initial outputs with authors while the editor checked the results. The prompts continued to be revised frequently and iteratively as the systems changed. Support varied by author: someone who preferred to start writing on a blank page needed different help from someone who first needed bullet points. Each project identified strengths to build on and weaknesses or unenjoyable tasks for which AI might provide support.
An improvement identified after this experience was to involve user-experience and user-interface professionals earlier, rather than leaving the team to work those elements out itself.
In May 2024, the publisher reported satisfaction with the workflow, tool and results, while noting that more people could help with such projects and that further experience might change the assessment.