How to Document Human Authorship in AI-Assisted Work

The US Copyright Office will register work made with generative AI. It will register only the human parts, though, and the application asks you to say which parts those are. Most creative teams could not answer that today, because the decisions that made the work theirs were never written down.

The Office’s registration guidance says applicants must disclose AI-generated material that is more than de minimis, exclude it from the claim, and briefly describe what the human author contributed. That description is only as good as the record behind it. Records made while the work happens are worth more than a narrative written at filing time.

This post covers what the Office asks for, what a successful filing looked like, and what to record while you work. It also covers the tension at the heart of the problem: you want the fullest possible record, and you want to control who sees it. The answer is to record everything and publish selectively. At the end, we show how to keep that record in Numonic, including where it falls short today.

What the Copyright Office Actually Asks For

Two documents set the rules. The Office’s March 2023 registration guidance says applicants must disclose AI-generated material that is more than de minimis, exclude it from the claim, and briefly describe what the human author contributed. Its January 2025 report on copyrightability (Part 2 of its series on copyright and AI) explains where the line falls.

The short version:

• Prompts alone aren’t authorship. The report concludes that, given current generally available technology, prompts do not give a user enough control over the output to make them its author. The Office left room to revisit this as tools change.
• Human contributions are protectable. Human-made material that is perceptible in the output, human modifications of AI material, and the human selection and arrangement of AI material can all be protected.
• Courts back the human requirement. In Thaler v. Perlmutter, the D.C. Circuit held in March 2025 that copyright requires a human author, and the Supreme Court declined to hear the case in March 2026. That case involved a work credited solely to an AI, so it doesn’t settle how much human input is enough.
• The hard cases are still open. Allen v. Perlmutter, the challenge over the prize-winning image Théâtre D’opéra Spatial (made from hundreds of Midjourney prompts plus edits), was fully briefed in early 2026 and, as far as we can find, still awaits a ruling.

The clearest early example is Zarya of the Dawn (February 2023). The Office registered the comic’s text and the selection and arrangement of its text and images, and excluded the Midjourney images themselves.

What a Successful Filing Looked Like

The animated short Critterz shows the other side. It was made in 2023 with DALL-E character and background images, then edited, assembled, scripted, and voiced by people. Its filing lawyer, Lisa Oratz, has described how it was registered as a whole work. On the Technically Creative podcast in October 2026, she said her team believes it was the first registration of an animated short made this way.

The human contributions she listed are the useful part:

• Background panels stitched together and modified by hand.
• Character images edited with tools a person controlled.
• A human-written script and human voice performances.
• Selection and arrangement of all of it into the finished film.

The filing excluded the raw AI images “to the extent not copyrightable.” That’s her team’s wording, chosen to leave room in case courts later protect more than the Office does now. It isn’t a phrase the Office prescribes. A feature-length Critterz was announced in September 2025.

What to Record While You Work

No statute or Office rule requires you to keep records of your creative process. Practitioners recommend it anyway, because the application asks you to describe the human contribution, and the Office or a court can ask for more. Oratz’s advice is to document every human touch point, since nobody yet knows which ones will matter most.

A practical record covers six things:

1. Human-made inputs. Sketches, storyboards, reference photos, and mood boards that shaped the output. Oratz notes that protection for a human sketch carries into output that reflects it.
2. Prompts and iterations. These are context, not authorship by themselves, but they show intent and direction.
3. Selection. Which generations you kept and, just as important, which you rejected.
4. Modifications. The before-and-after versions of anything a person changed.
5. Arrangement. How the pieces were combined, sequenced, and edited into the final work.
6. Who and when. Which person made each decision, and the date.

Records made while the work happens are worth more than a narrative written at filing time. That’s our reasoning rather than settled law: a contemporaneous log is harder to dispute, and it captures choices people forget.

Record Everything, Publish Selectively

Oratz’s second point is less obvious. She advises against making that documentation public, because it shows which parts came straight out of the tool and might be unprotectable. That’s a strategic judgment rather than a rule, but it’s a sound one, and it creates a real tension: you want the fullest possible record, and you want to control who sees it.

That tension is growing. Provenance standards such as C2PA, and AI-disclosure rules such as Article 50 of the EU AI Act, push more files to carry a public signal that generation was involved. A public label will say “AI was used.” Only your private record will show how much human authorship went in. We explored the same trade-off in The Metadata Stripping Paradox.

The workable answer is to separate the record from the deliverable:

• Keep the full history internally: inputs, prompts, versions, selects, and edits, linked to each output.
• Decide per export what travels with the file. A portfolio piece might carry only your name and an AI-disclosure label, a client deliverable might add licensing terms, and the archive keeps everything.
• Never rebuild the record from memory at filing time if you can capture it as you go.

This is the argument we made from the German rulings in AI Copyright Rulings Test One Thing: the question courts keep asking is whether you can reconstruct what you did.

How to Keep This Record in Numonic

Numonic is the asset library we build for AI-generated work, so here’s how the six records map onto it today, including where it falls short. The aim is a record that builds itself while you work, so there’s nothing to reconstruct when you file.

1. Bring your human-made inputs in first. Upload sketches, storyboards, reference photos, and mood boards through Import, and keep them in the same project collection as the work they guide. If you generate in ComfyUI and feed a sketch in through a LoadImage node, Numonic links the output to that sketch on import by matching the image’s content, as long as the sketch is already in your library. That link is stored with the output and available to agents through Numonic’s MCP tools. The asset page doesn’t display it yet, so keep the sketch beside the output in the collection as well.

2. Let capture run while you generate. Add the Save to Numonic node to your ComfyUI workflows (it’s listed in ComfyUI Manager as Numonic Workflow Recovery), or point Connect Folder at your output folder in Chrome or Edge. Numonic reads the prompt, seed, model, LoRAs, and full workflow graph from each file as it arrives, so the iteration history exists without anyone typing it up. Midjourney work comes in through the Midjourney importer, which rebuilds each image’s evolution chain and lets you compare any two steps.

3. Write down why you kept what you kept. Every asset has notes: threaded comments you can pin to a spot on the image, each stamped with who wrote it and when. A one-line note on a select (“kept for the hand pose; rejected the other three for the lighting”) is exactly the selection evidence an application can draw on. For Midjourney chains, you can also mark each branch Selected, Rejected, or Undecided and record a reason.

4. Keep edits connected to their originals. When you rework an approved image in ComfyUI (re-lighting it, or removing an object) and load the original through LoadImage, the new version is linked to its source the same way as in step 1. For edits made in other tools, import the before and after versions together and note what changed. Numonic doesn’t detect outside edits on its own.

5. Let the collection show the arrangement. Nest collections by project, scene, or deliverable, drag assets into the order they appear in the final work, and add text cards for the script beats or the reasoning behind a sequence. The collection then documents the selection and arrangement, which is the part of an AI-assisted work the Copyright Office most clearly recognizes.

6. Export with the right preset. Each export preset decides what travels with the file:

• Share strips the workflow, prompt, model names, seeds, location, and personal metadata.
• Portfolio strips the same process data and location but keeps your name and copyright.
• Client Delivery also strips the process, and keeps creator, copyright, title, keywords, job reference, and licensing terms.
• Full Archive keeps everything and, for a single asset, offers the workflow graph as a download. That’s the copy for your own records.

If a client does need to see your method, the Custom option lets you choose each field. Exports of AI-generated work also get the industry-standard IPTC disclosure for AI-generated media automatically, so the public signal is consistent while the detailed record stays with you.

What Numonic doesn’t do yet. There’s no one-click authorship report: at filing time you assemble the description from the project collection, its notes, and a Full Archive export. Notes and uploads show who did what and when, but there’s no single per-asset history of every change. And Numonic keeps the record. It doesn’t file applications or give legal advice.

A Note on What This Isn’t

This article describes the Copyright Office’s published guidance and one practitioner’s approach. It isn’t legal advice. Registration strategy depends on the work, the jurisdiction, and facts a blog post can’t see, so talk to a copyright lawyer before you file.

Key Takeaways

• The Copyright Office registers AI-assisted work, but only the human contribution, and you have to describe it.
• Prompts alone aren’t authorship under current guidance. Selection, arrangement, modification, and human-made inputs can be.
• Record-keeping isn’t legally required, but it’s how you answer the application’s questions, so capture it while you work.
• Keep the full record private and choose what each export reveals.
• In Numonic: import your own inputs, let the ComfyUI node or Connect Folder capture each generation, note why you kept each select, arrange the project collection in its final order, and export public work with Share or Portfolio while keeping a Full Archive copy for your records.