Prompts alone do not make you an author

The instinct is almost universal and it is wrong. You write a prompt specifying the lens, the grade, the blocking, the time of day, the mood and the lighting ratio, iterate it forty times, and conclude that the result is more yours than something a stranger typed in six words. The US Copyright Office's position is that the length and precision of the prompt do not change the analysis. Prompts, however detailed and however many times refined, do not by themselves make you the author of the output.

That is a narrow claim and it is worth stating precisely, because it gets stretched in both directions. It does not mean generated material can never sit inside a protected work. It does not mean prompting is unskilled or commercially worthless. It means one specific thing: the act of writing an instruction is not, on its own, authorship of what comes back.

Two things that are true at the same time

Your prompt may well be your own copyrightable text. A long, structured, distinctive prompt is writing, and writing gets the ordinary treatment. That is a copyright in the prompt.

The image or clip that comes back is a separate question with a separate answer. The reason the Office gives is about control over expression: between your instruction and the finished output sits a system making the expressive choices — which specific face, which exact composition, which particular fall of light. You described a destination; something else picked the route and drew the scenery.

People conflate these constantly, and the conflation produces the confident but unfounded claim "I wrote a 600-word prompt, so I own the image." You may own the 600 words. That is not the same asset.

Why more detail does not scale into authorship

Here is a test you can run yourself in ten minutes, and it is the cleanest way to see why length does not help.

Take your longest, most specific prompt. Run it three times with a different seed each time. Look at what changed.

If the same instruction can produce materially different expression — different faces, different compositions, different lighting — then the instruction was not determining the expression. Something else was. That is not a legal argument you would make in a filing, but it is the intuition behind the Office's reasoning, and it is why adding another hundred words does not move the needle. Specificity narrows the range of outputs. It does not close it, and the gap is precisely where the expressive choices are being made.

The same logic dispatches the iteration argument. Forty prompts is forty instructions. Each one produced a machine output; none of them produced a human-authored one. Volume of attempts is a measure of effort, and effort has never been what copyright rewards.

An audit of your own contributions

The useful question is not "do I own this" but "which of the things I did counts, and over what." Run your last project through this.

What you did	Does it carry a claim?	Over what
Wrote the prompt, however long	No, not over the output	Possibly over the prompt text itself
Re-prompted forty times	No	Forty instructions is not authorship
Swept seeds until one landed	No	Picking a winner is not authoring it
Chose the model and settings	No	A tool choice, not an expressive one
Edited the chosen frame — inpaint, remove, extend, regrade	Yes	The changes you made
Arranged separate generated elements into a composition	Yes	The arrangement, not the elements
Wrote the script	Yes	The script, as its own work
Recorded or performed the voice	Yes	The performance and recording
Cut a timeline — shot order, duration, transitions	Yes	The edit as a structure
Set original type, drew, or photographed	Yes	What you made
Trained a LoRA on your own drawings	The drawings, yes	Not the generations from it

Two rows deserve elaboration because they are the ones people misread.

Seed sweeping. Choosing one output from forty feels like a creative act, and in a workflow sense it is. The recognised route through selection is narrower than that: it is the creative selection and arrangement of material into a new whole — the compilation logic — not the act of picking a favourite. One chosen image is still one machine output.

Model and settings. Choosing a model, a sampler, a guidance value or an aspect ratio configures the system. It is a tool choice, not an expressive choice in the finished work.

Video is the strongest case, and most people undersell it

If you produce video, the audit above should read as good news rather than bad, because the human-authored surface in a finished cut is large and easy to describe.

A finished cut assembled from generated clips contains a written script, a decision about which clips to use, the order they appear in, how long each one holds, where the cuts land, and where sound and captions sit. None of that came out of a model. The individual clips remain unprotected in the US analysis; the structure containing them is a different work. How to build that structure so the claim is actually there is selection and arrangement copyright for AI edits. How to write the contribution down on a registration form is registering an AI-assisted work.

The production bias worth adopting: generate elements, author the assembly. Do not ask a model for a finished deliverable and then look for authorship afterwards. The editor is EDL-based, so the timeline is the record of those decisions and stays re-renderable rather than baked. Iterating on it runs through preview: true, a 480p pass at no credit cost with a short per-user cooldown, and the export charge lands once on the version you confirm regardless of how many clips sit on the timeline — the mechanics are in previews and the final export.

Keep the evidence while it is cheap

Authorship you cannot evidence is authorship you will struggle to assert, and none of it is reconstructable from a flattened export a year later. Three habits cover most of it: note the model, version and date at generation time; keep the intermediates in order so an edit sequence is visible rather than inferred; and keep one line per real decision, because the reason behind an arrangement is the claim stated in words.

Past generations can be retrieved from the prompt they were made with, and finding something you made before by description rather than filename is what makes the record usable rather than merely stored. The still-image version of this build order is in what makes an AI print design legally protectable.

None of this is legal advice, and the guidance continues to develop. The position above is the US Copyright Office's Part 2 report of 29 January 2025, which is Office analysis rather than a statute or a holding. Check the Office's current materials directly and take counsel before relying on any of it commercially. The wider copyright and safety guide covers the adjacent questions — training data, voice consent, platform rules — that this post deliberately leaves alone.

FAQ

Is my prompt itself copyrightable?

Potentially, as a piece of text, on the ordinary standards that apply to any short written work. That gives you rights in the prompt, not in the output, and the commercial value of that is usually limited unless you are selling prompts as products — in which case licence terms for selling prompts and LoRAs is the relevant read.

If I iterate two hundred times, does that eventually count?

No. Two hundred iterations is two hundred instructions, and the analysis does not accumulate. What does change the answer is doing something to a specific output — editing it, composing with it, building a structure around it. One deliberate edit to a chosen frame does more for your position than a week of re-prompting.

Does using reference images or a fine-tune change it?

Not by itself. A reference steers the system and a fine-tune changes what it produces, but the output remains model output either way. The one place it matters is when the source material is yours: your own drawings or photographs used as references are human-authored works you already own, and that ownership stays with the source, not with what the model makes from it.

Does this apply outside the US?

This is the US position specifically, and jurisdictions genuinely differ — some have statutory provisions for computer-generated works with no human author that start from a different premise entirely. If you licence internationally, treat the US answer as one input rather than the answer, and get advice per territory before making promises in a contract.