Find your tool below. The short answer is the same for all of them: copyright protects human authorship, so what you can protect depends on what you did, not which tool you used.
Copyright protects human authorship: expression a person controlled. With AI-assisted work, you author through three recognized acts, selection (choosing this output over others, what to keep and discard), arrangement (ordering and composing material), and modification (changing the generated material beyond the trivial). Do these meaningfully and the result carries your authorship.
Prompts alone are insufficient. A prompt conveys your idea; it does not give you control over how the model renders the expression, and detail does not change that. There is also no percentage: the law refuses a quantitative threshold and looks at the kind of control you exercised, case by case. The tool you used does not decide whether your work is copyrightable. What you did with it does.
For the full standard, read the authorship methodology and the FAQ. To see how this plays out for your craft, see developers and writers.
AI-suggested code you accept unchanged sits on the same footing as any unmodified AI output. The architecture you chose, the parts you edited, the components you selected and wired together, and the direction you gave across iterations are human authorship. Most real software built with an assistant carries substantial human creative control.
You set the architecture, accept or reject each change, and redirect the work as it goes. That control over what gets built and how is the protectable layer; raw accepted suggestions are not.
Verbatim AI completions are raw output. Your refactors, the structure you imposed, and the edits you made to suggestions are where your authorship shows.
A completion you accept untouched is thin. Selecting among suggestions, editing them, and arranging them into a coherent design is the human contribution that can be protected.
The more the agent runs unsupervised, the thinner your claim on what it produces. Your review, correction, and architectural direction are what carry authorship.
Your commit-by-commit decisions, the scope you set, and the edits you keep or revert document the human control that makes the result yours.
Same test as any assistant: accepted output alone is weak, but your selection, arrangement, and modification of what it generates is protectable.
Generated scaffolding is raw material. The features you designed, the logic you wrote or reshaped, and the choices you made about how it fits together are the human layer.
A generated component pasted as-is is weak. Composing components into a real interface, restyling, and rewriting the parts that matter is authorship.
The prompt that spins up a starter is an idea, not expression you controlled. Your edits, structure, and selection across the build are what get protected.
Generating an app from a description does not by itself make the output yours. Shaping it, choosing what to keep, and modifying the result is the protectable work.
Hand a task to an autonomous agent and walk away, and you have controlled the idea, not the expression. Your specification, review, and the changes you direct are the human contribution.
Ideas are never protected by copyright; expression is. Text you paste verbatim with no meaningful change is raw output. Text you substantially rewrote, or passages you selected and arranged into a larger work, can be protected. The more your expressive control shows, the stronger the claim.
Pasted-verbatim text is not yours. Rewriting it substantially, or selecting and arranging output into a work you shaped, is the human authorship that can be protected.
The model behind the output does not change the test. What you select, arrange, and modify into your own work is protectable; raw responses are not.
Same standard as any chat tool. Your editing and arrangement of the generated material is the protectable layer, not the unmodified reply.
Synthesized answers are raw material. Facts are not copyrightable anyway; how you write up, select, and arrange the findings into your own work is.
Generated drafts in a document are a starting point. Your rewrites, structure, and selections are what carry authorship into the final file.
Templated output you ship unchanged is thin. Editing for voice, selecting among variations, and arranging copy into a campaign is the human contribution.
Generated prose dropped in verbatim is raw output. Your line edits, the passages you keep, and the way you compose the manuscript are protectable.
Continuations taken as-is face the same problem as any raw output. Substantial editing and your selection and arrangement of the narrative are where authorship lives.
The raw image out of a text-to-image generator is the hard case. The Copyright Office has consistently found that prompting alone, even very detailed prompting, does not make the output human-authored, because you do not control how the model renders the expression. Meaningful modification afterward, or selecting and arranging generated elements into a larger composition, can be protected.
A raw generation, however detailed the prompt, is generally not human-authored. Editing it meaningfully, or selecting and arranging images into a larger work, can be protected; the unedited output on its own is not.
Prompting conveys your idea, not control over the expression. The protectable contribution is what you do after: modify, composite, select, and arrange.
Even with custom models and elaborate pipelines, the raw output faces the prompting-is-insufficient rule. Your post-generation editing and arrangement are the human layer.
Generative fill and edits inside a composition you authored sit alongside your own work. The human-made composition and substantial editing are protectable; the raw generated fill alone is not.
Generated images, including text-in-image, are raw output. Selection, arrangement into a design, and meaningful modification are what carry authorship.
Asset generations taken as-is are weak. Editing, compositing, and arranging them into a coherent work is the protectable human contribution.
Higher fidelity does not change the test. The raw output is not human-authored by the prompt; your modification, selection, and arrangement are.
The same human-authorship test applies to sound. Lyrics you wrote are yours. Arrangement, selection, and substantial editing of generated audio can be protectable. Audio generated from a prompt, taken as-is, faces the same problem as raw image output: the prompt conveys an idea, but you did not control the expression.
Lyrics you wrote are protectable. A full track generated from a prompt, taken as-is, is raw output; your editing, arrangement, and selection of the audio are the human layer.
Same standard as any music AI. Composition you contribute, substantial editing, and arrangement of generated audio carry authorship; the unmodified generation does not.
Synthesized narration of your script: the script is yours. The raw voice output is generated, but your writing, direction, and editing of the audio into a work can be protected.
Generated video is the newest frontier of the same rule. A clip produced from a prompt is raw output you did not control at the level of expression. Editing, sequencing, compositing, and arranging generated footage into an edited piece is where human authorship enters.
A prompt-generated clip is raw output. Your editing, sequencing, and arrangement of footage into a finished piece is the protectable human contribution.
Detailed prompting still conveys an idea, not control over how the video is rendered. What you edit, select, and arrange afterward carries authorship.
Generated clips taken as-is are weak. Compositing, sequencing, and meaningful editing into an edited work are where your authorship shows.
Same test as any video AI. The raw generation is not authored by the prompt; your modification and arrangement of the output are.
Higher quality output does not change the standard. Your editing, selection, and arrangement of the generated footage are the protectable layer.
Different tool, same test. Whether you are writing, coding, generating images, scoring music, or cutting video, the raw output of the model is not what you own. Your selection, your arrangement, and your modification are the protectable layer. That layer is real, and in most serious creative work it is substantial. The problem is never that the authorship is missing. The problem is proving it after the fact.
So capture it as you create. Copyrightable records the human decisions behind a filing, the prompts, the edits, the rounds of selection and revision, and ties them to a tamper-evident evidence file. When someone asks whether a human authored it, you have an answer instead of an assertion. Read the full methodology, browse the blog for worked examples, or check the FAQ.
Already made something in ChatGPT, Claude, or Gemini? Paste your conversation transcript and watch the evidence build from the session you already had.
Copyrightable is evidence of human authorship, not legal advice and not a grant of copyright.