What is the right way to use AI for nonfiction writing?

The short answer

There is no single right way. There is a set of boundaries, and they move with the phase of the work. Keep for yourself the work where understanding forms: building the argument, synthesising sources, forming your claims, writing the sentences. Hand the AI the work that is genuinely mechanical: finding and retrieving sources, managing citations, spell-checking, formatting.

A 2026 paper on AI in research writing puts the reason in one line: for the work that builds understanding, “when a tool performs the work, the processing does not transfer to the writer; it does not occur” (Chauncey, 2026). Offloading the mechanical work, the same paper says, “is benign and often beneficial.”

What should I keep, and what can I hand to AI?

Why does it matter most for nonfiction?

Because in nonfiction the argument is the product. A review of 109 studies on AI writing tools, with interviews of 15 writers, found that “content-focused writers (e.g., academics) prioritize ownership during planning, while form-focused writers (e.g., creatives) value control over translating and reviewing” (Reza and colleagues, 2025). In a study with 30 writers, AI was found more helpful for turning ideas into text and for reviewing than for planning (Chakrabarty and colleagues, 2024). Planning — deciding what you actually think — is the phase to protect.

What goes wrong when AI writes for you?

The problems are not random. They cluster along six fault lines.

1. Thinking. In an MIT study, 15 of 18 writers who used ChatGPT (83%) could not quote a sentence from the essay they had just written, and their brain connectivity was the weakest of the three groups. The authors describe cognitive debt as something that “defers mental effort in the short term but results in long-term costs, such as diminished critical inquiry” (Kosmyna and colleagues, 2025, a preprint). In a randomised study of 117 students, ChatGPT improved essay scores more but not learning (Fan and colleagues, 2025).

2. Meaning. Researchers using AI to help write literature reviews reported that “the primary bottleneck is not writing quality but maintaining semantic and citation integrity across iterative revisions”: citations moved to the wrong sentence, and their voice was “flattened into a neutral, generic style” (Nogueira and colleagues, 2025, a preprint). Another study found that even edits meant only to fix grammar can significantly change what a text means (Abdulhai and colleagues, 2026, a preprint).

3. Authorship. People do not feel they own text the AI wrote, yet still put their names on it — “the AI Ghostwriter Effect” (Draxler and colleagues, 2024). In a 2026 study, writing by hand produced higher ownership than every AI-assisted condition (Pennanen and colleagues, 2026).

4. Trust. In 636 citations generated by ChatGPT, 55% of GPT-3.5’s and 18% of GPT-4’s were fabricated, and many real ones had errors (Walters and Wilder, 2023). In a 2025 study of mental-health reviews written by GPT-4o, 20% of citations were fabricated and 45% of the real ones contained errors (Linardon and colleagues, 2025).

5. Voice. Across more than 880,000 texts, AI polishing reduced the variety of writing complexity by 21–50% (Sourati and colleagues, 2026). AI suggestions led Indian writers “to adopt Western writing styles” (Agarwal and colleagues, 2025), and AI-assisted stories were “more similar to each other than stories by humans alone” (Doshi and Hauser, 2024).

6. Dependence. In two pre-registered experiments with 1,100 people, AI help boosted creativity during the task but “may inadvertently hinder independent creative performance when users work without assistance” (Kumar and colleagues, 2025).

The pattern behind all six

Look at the six together and one pattern emerges, in the words of Carlos Vettorazzi, who built Wretify: AI writing tools optimize for the wrong thing. They optimize for output — fluent, fast, polished text — while every one of these problems is about process: the thinking, meaning, ownership and skill that writing exists to build. Fluent text is not proof that the thinking happened.

Should I tell my publisher I used AI?

Yes. The ghostwriting guidelines published by a working group convened by Gotham Ghostwriters (2026) say writers “should disclose the nature of their AI use to clients and publishers”, because doing so “creates mutual understanding, which benefits everyone in the publishing ecosystem.” The Authors Guild warns that “AI-generated text is not copyrightable” and that including it may breach a book contract’s warranty. On Amazon KDP, AI-generated text must be disclosed; using AI to brainstorm or check your own writing counts as AI-assisted. Can I publish a book if I used AI? →

How does Wretify draw these boundaries?

Wretify is built on them. Its AI takes the mechanical work: it searches 250 million real papers for sources and turns them into footnotes, builds your reference list, finds what you have already written on a theme, and runs a grammar pass that fixes only what is broken. It refuses the work where understanding forms: it will not draft your chapters. Instead it asks — explain this passage, find its blind spot, test the logic — and the sentences stay yours.

Why Wretify’s AI refuses to write →