How I Document Human Authorship When I Write With AI

I use Claude while writing books. It helps me explore structures, produce rough drafts, and test different ways to explain an idea.

That does not make every sentence Claude produces mine for copyright purposes.

The U.S. Copyright Office says that material produced by generative AI can be protected only when a human has contributed enough original expression. Human modifications and an original selection or arrangement may qualify. Prompts alone do not give the user copyright in the resulting expression. The Office explains the distinction in its 2025 report on AI and copyright.

My practical response is simple: I keep the AI draft, make the human work visible, and register only what I can honestly describe as human authorship.

Copyright can cover the human contribution

An AI-assisted book may contain several different kinds of material:

• text written directly by the author;
• AI text that the author substantially changed;
• AI text kept with little or no change;
• an original structure created by selecting and arranging material;
• research facts, quotations, and other material that copyright does not protect in the same way as original expression.

The legal question is not whether AI appeared anywhere in the process. The question is which expressive parts were created by a human.

There is no official percentage that makes a draft safe. Rewriting 40%, 60%, or 80% is not a legal test. A word-count difference may help document the process, but it cannot decide authorship by itself.

I keep four versions of the work

For every chapter, I try to preserve four things:

human outline and notes
          ↓
raw AI output
          ↓
revised working draft
          ↓
final manuscript

The outline shows the argument I intended to make. The raw output shows what the model supplied. The revised draft shows the changes. The final manuscript shows what I chose to publish.

I keep these files in one project folder:

/BookTitle/
  notes/
  prompts/
  raw-outputs/
  drafts/
  final/
  decision-log.md

This archive is useful even if no legal dispute ever happens. It prevents me from forgetting which passages still need real editorial work.

The important work happens after generation

When I revise an AI draft, I do more than replace a few words.

I may change the order of the argument, remove unsupported claims, add examples from my own work, rewrite the explanation, or reject the entire draft. Sometimes I use the model only to identify gaps and write the replacement myself.

A good test is to compare the raw output with the final chapter and ask:

1. Which ideas and examples could only have come from my experience or judgment?
2. Which sentences did I actually write?
3. Which structural choices did I make?
4. Which AI passages remain almost unchanged?

The fourth question matters most. If an important passage remains close to the raw output, I should not quietly treat it as fully human-authored text.

A decision log records choices, not just keystrokes

Version history shows that a document changed. It does not always explain why.

I keep a short decision log for important choices:

• why I rejected one chapter structure;
• why I kept one example and removed another;
• where I replaced an AI explanation with my own analysis;
• which sections still contain AI-generated material;
• which sources support a factual claim.

The log does not guarantee copyright protection. It creates a clearer record of the human decisions behind the final work.

Registration requires an honest description

When a work contains an appreciable amount of AI-generated material, the Copyright Office asks applicants to disclose and exclude that material from the claim. Its registration guidance explains the current process.

In practice, the application should distinguish between:

• the new material created by the human author; and
• AI-generated material that is not being claimed.

The description depends on the actual manuscript. A standard sentence copied from somebody else’s application may be inaccurate for your book.

I built an open-source copyright-shield skill to keep raw outputs, drafts, decisions, and a proposed registration description together. It helps with documentation. It does not decide whether a passage is copyrightable, and it is not a substitute for legal advice.

Start with one chapter

Choose one chapter that used AI heavily. Save the original AI output beside the published version. Mark which passages you wrote, which you reorganized, and which remain close to the model’s text.

If you cannot explain the human contribution clearly, revise the chapter or narrow what you claim. Documentation is most useful when it changes the writing process before publication.

This article describes my workflow and current U.S. Copyright Office guidance. It is not legal advice, and rules in other countries may be different.