I use Claude to write books. I feed it outlines, generate chapter drafts, iterate on structure, produce full manuscripts. Two of my books went through this process.
In March 2026, the Supreme Court confirmed what the Copyright Office had been saying for three years: if AI wrote it, you can’t copyright it. No human author, no protection. Thaler v. Perlmutter, cert denied.
But I copyright my books. Because the law draws a line between “AI-generated” and “AI-assisted,” and that line is where the work happens.
The line
The U.S. Copyright Office published a report in January 2025 (Part 2: Copyrightability, based on 10,000+ public comments) that maps the territory. Three things matter for text:
Prompts don’t count. Even complex, iterative prompt engineering. The Office compared it to the “sweat of the brow” argument that the Supreme Court rejected in Feist v. Rural Telephone back in 1991. You can spend 40 hours crafting prompts. You still don’t own the output.
Substantial rewriting does count. If you take AI-generated text and rewrite it in your own voice with your own analysis, you own what you created. The Copyright Office calls this “creative control over the work’s expression.”
Selection and arrangement count. Generate 50 AI outputs, pick the best 10, organize them with your own transitions and commentary. The compilation is copyrightable. The individual AI passages are not. Kashtanova got copyright on her text and her creative arrangement of AI images in Zarya of the Dawn, but not on the images themselves.
How I work
My process runs through Claude in Cowork mode. I built a copyright-shield skill (instructions that Claude follows during writing sessions) that automates the documentation side. When I say “copyright mode,” Claude sets up a project folder:
/BookTitle/
prompts/ ← every prompt I send, timestamped
raw-outputs/ ← every Claude response, timestamped
drafts/ ← my revised versions
notes/ ← outlines, research, creative decisions
decision-log.md ← why I chose what I chose
final/This folder is my evidence archive. If someone challenges my copyright, I show them every raw AI output next to my final text. The diff between those two files is my authorship.
The rewriting step
This is the part that determines whether you own your book or not.
I generate a chapter draft from Claude. I save it to raw-outputs. Then I rewrite. I restructure paragraphs, add analysis from my own experience, cut the generic language AI defaults to (“It’s important to note that…” goes first), rewrite sections in my voice. I add examples and arguments that Claude can’t produce because it doesn’t know my industry the way I do.
The Copyright Office hasn’t set a percentage threshold. No one can tell you “rewrite 60% and you’re safe.” They evaluate case-by-case. My approach: every chapter should contain material that could only come from me. If someone ran a diff between Claude’s draft and my final manuscript and the changes were trivial, I’d have a problem.
The decision log
The skill maintains a running log of creative decisions. I generate three versions of a chapter opening, pick version 2, the log records my choice and the reason. I cut a section Claude wrote and replace it with my own analysis, that goes in the log too.
The Copyright Office cares about creative judgment. Picking which AI output to use, deciding how to structure the book, choosing what to cut: authorship decisions. The log proves I made them.
Registration
When the book is done, the skill generates copy-paste text for the three fields that matter on the copyright.gov application:
Author Created: “Text as written, revised, and edited by author; original analysis, arguments, and commentary; selection, coordination, and arrangement of all content.”
Material Excluded: “Text generated by artificial intelligence (Claude AI).”
Note to CO: Two to four sentences describing the creative process, tailored to the specific project. What AI did, what I did, how I shaped the final work.
The Copyright Office has registered hundreds of works containing AI-generated material since 2023. Disclosure is mandatory and honesty is mandatory, but registration of AI-assisted work is routine at this point.
The case that should scare you
Jason Allen won a Colorado State Fair art competition with a Midjourney image, edited it in Photoshop, and applied for copyright. The Copyright Office refused because his edits were “inextricably merged” with the AI output. The examiner couldn’t tell where Midjourney ended and Allen began.
For text, the lesson is clear. If your edits blend so tightly with the AI text that no one can separate them, you risk the same outcome. Keep the raw outputs. Keep the diffs visible.
Allen sued in 2025 and the case is pending in Colorado federal court. I’m not waiting for the outcome. The evidence archive means I don’t have to.
Tools worth knowing about
The copyright-shield skill is a Claude skill I built for this workflow. It handles the archive structure, decision log, registration text generation, and an authorship report that compares raw AI outputs against my final text and produces a per-chapter percentage breakdown. What I kept unchanged, what I rewrote, what I wrote from scratch. It covers most of what Grammarly Authorship does except real-time keystroke replay. Open source, MIT license. I also have companion skills for converting manuscripts to Kindle format and generating KDP covers.
Git gives you a full version history of every file. Commit the AI’s raw output, commit your revised draft, commit the final. The diff between commits is your authorship evidence. Works offline, and the history is harder to fabricate than a Word document’s version timeline.
Grammarly Authorship tracks writing in real time inside Google Docs and Word. Categorizes text by origin: typed by the user, pasted from AI, copied from the web. The replay feature shows the document’s creation from first word to last keystroke. Free tier covers the basics, Pro is $12/month.
Google Docs version history records changes with timestamps. On by default. If you paste Claude’s output and then rewrite it, the steps are recorded. Don’t copy the document to a new file, that wipes the history.
One thing you can do today
Open your current manuscript. Find a chapter where you used AI to generate text. Save the AI’s original output in a separate file. Then look at what you published. If the two files are almost identical, you have a copyright problem.
If they’re different, you’re doing it right. Keep the evidence.
Your move.