| Takeaway | Detail |
|---|---|
| In a U.S.-focused 2026 claim audit, model output plus prompts plus selection alone will not support human authorship for generated passages. | The claim fails when the record lacks identifiable human-authored expression or a human-created arrangement mapping to the final passage. |
| Claim only expression that maps to a human creative contribution in the final work; separate or exclude unsupported model-generated material. | The audit check is passage-by-passage: if no human creative contribution is identifiable, the expression is excluded from the claim. |
| A contributorship-based authorship test requires an original idea or development of a work that disseminates intellectual content. | Per the authorship and plagiarism guidelines, mere model output does not satisfy this original-idea or development threshold. |
| If Model B transforms Model A’s output, constraints from Model A’s terms may still apply to the downstream claim. | This output-laundering check follows the quiet dependency chain in AI ownership: transforming another model’s output does not necessarily clear its constraints. |
This guide delivers a U.S.-focused 2026 audit framework for AI copyright claims, testing whether prompts and model output alone support human authorship.
It maps each claimed passage to identifiable human creative contribution, separates unsupported model material, and tracks model-dependency and license constraints.
Trace the human contribution in the final work
The audit’s unit of analysis is the specific expression in the submitted work, paired with a traceable human contribution to that expression. For every passage, image, or other claimed element, create a record identifying who supplied its expressive content, what the model generated, and what a person changed before publication. The relevant question is not whether a person initiated a request or approved an entire result, but whether the final work contains identifiable expression attributable to a human creative contribution. This applies the focus of Writing at a Distance on the human component in human-machine interactions and its relationship to the final text.
Use separate log fields for the prompt, model output, output selection, and substantive revision. A prompt records instructions or source material; it does not by itself establish authorship of the resulting language. Selecting an output identifies which generated version entered the workflow, but selection alone does not show how the selected expression originated. A substantive revision, by contrast, should be documented at the level of the changed words, images, arrangement, or other expressive features. Record the person responsible, the specific change, the draft or output it affected, and the corresponding text or media in the final work.
Maintain a version trail from source materials and prompts through generated drafts, human edits, and the submitted version. Then map each claimed human contribution to the exact portion it shaped. A useful threshold is documentary specificity: if the log cannot identify both the person’s creative act and the final expression affected by it, the record does not support treating that portion as human-authored expression. Descriptive labels such as “edited by,” “curated by,” or “based on a prompt” are not substitutes for that mapping.
Check the current U.S. Copyright Office guidance relevant to the type of material being claimed, and compare the mapped contribution with any contractual, platform, or publication standard governing the record. The FAO’s Authorship and Plagiarism Guidelines describes authorship in terms of creating an original idea or developing work that disseminates intellectual content, which is why the audit should connect contribution to the particular work rather than rely on a general credit label. Preserve drafts, revision histories, source files, permissions, and license information for the exact human contribution. This combination makes the claim reviewable: the reader can locate the contribution, follow it into the final work, and evaluate it under the applicable standard.
Separate authorship evidence from inference
The named materials support questions about attribution and contribution, not a numerical U.S. legal test. The arXiv discussion of “The human-authorship halo” uses Raymond Queneau’s Exercices de style, which contains 99 retellings; that figure describes the book, not a copyright threshold or study result established here. Hannes Bajohr’s Writing at a Distance characterizes causal authorship as a heuristic, so use it to frame contribution tracing rather than as a legal rule or measured pass rate. The FAO Authorship and Plagiarism Guidelines discusses original-idea creation and contributorship, not a fixed word-count or percentage standard for U.S. copyright claims.
To trace contribution, compare the final expression against the human-supplied input that shaped it. If a human wrote or revised the wording, structure, or arrangement of a passage, that contribution is traceable and supports a claim. If the human only provided a prompt and accepted the model’s output without further expressive editing, the output lacks a human-authored expression and should be excluded from the authorship claim. This comparison must be documented per element, not assumed across the work.
| Element | Human Input | Expression Source | Claim Supported? |
|---|---|---|---|
| Passage A | Prompt + selection | Model output, unmodified | No |
| Passage B | Prompt + line edits | Model output, human-revised | Yes |
| Passage C | Full draft | Human-authored | Yes |
Choose the narrowest supportable claim: claim human-authored material separately from unsupported model output. Do not aggregate model-generated text into a human-authored total based on prompting or curation alone. Each claimed element must map to a documented human creative contribution in the final expression. This separation preserves the integrity of the authorship record and aligns with the heuristic framing of causal authorship as a distance-measured contribution, not a binary or statistical threshold.
Choose the narrowest supportable claim
Choosing the narrowest supportable claim means treating model output and human-authored expression as distinct categories rather than blending them into a single authorship assertion.
Apply the passage-level mapping described above: do not broaden a claim beyond the specific human expression documented in the final work. Keep generated passages outside the human-authorship claim unless the record identifies a human contribution to those passages.
An undifferentiated mixed-work claim, which treats human-written and model-generated material as one inseparable block, should be avoided unless the human contribution to the claimed expression is clearly identifiable. Without a way to isolate which portions reflect human creative input, such a claim lacks the granularity needed for a credible audit. The risk is that the entire block may be discounted if the model-generated share cannot be separated, leaving the human-authored portions unsupported.
The delimited human-contribution claim is the preferred approach. This method requires identifying and claiming only the portions of the work where human-written or human-shaped expression can be directly traced. By scoping the claim to these delimited segments, the audit maintains a defensible position that aligns with the principle of claiming only expression mapped to a human creative contribution. This separation also supports clearer licensing and attribution decisions downstream.
For an editing claim, require a side-by-side comparison rather than relying on a change count or a general statement that a person “revised” the draft. Mark each retained, deleted, and newly written segment, and classify the change as mechanical or expressive. If the human contribution cannot be identified at the level of the final wording or structure, treat the edit as insufficient support for claiming the underlying generated passage.
For a selection claim, ask two separate questions: did the person create an expressive arrangement, and did the person create the selected material itself? If the answer to the first is yes, preserve the arrangement as its own claim and describe its components precisely. If the answer to both is no, selection evidence should not be used to convert the model’s passage into a human-authorship claim.
For a terms claim, capture the service agreement and relevant product terms at the time of use, including any terms incorporated by reference. “AI Ownership: If You Don’t Own the Model, What Do You Own?” warns that a dependency chain can carry constraints from an upstream model even when a different service produced the immediate output. If the terms are unclear, flag the material for permission review rather than treating copyright eligibility as permission to use.
Make the final filing decision conditional: if identifiable human expression or a human-created arrangement is documented in the submitted work, claim only that supported contribution; if it is not, separate or exclude the unsupported model-generated material. Keep the comparison files, arrangement notes, and applicable terms with the claim record so the decision can be checked without relying on memory.
Frequently Asked Questions
What must a claimant identify to support human authorship for a generated passage?
The claimant must identify human-authored expression or a human-created arrangement that maps to the final passage.
Do prompts, model output, and selection alone establish human authorship?
No, prompts, model output, and selection alone do not support human authorship for generated passages.
How is an AI copyright claim evaluated passage by passage?
Each passage is tested for an identifiable human creative contribution, and expression without one is excluded from the claim.
What should a claimant do with model-generated material that is not supported by human creative contribution?
The claimant should separate or exclude the unsupported model-generated material from the claim.
What authorship threshold must a human contributorship satisfy?
A human contributorship must involve an original idea or development of a work that disseminates intellectual content.
Can transforming one model’s output remove another model’s licensing constraints?
No, transforming Model A’s output does not necessarily clear constraints from Model A’s terms.
Quick answers
| What does a model-only claim lack when it relies on prompts, model output, and selection? | It lacks identifiable human-authored expression or a human-created arrangement mapping to the final passage. |
| How should unsupported model-generated material be handled? | Separate or exclude it from the claim. |
| What is the unit for auditing generated passages? | The audit is passage-by-passage, and expression is excluded when no human creative contribution is identifiable. |
| Does mere model output meet the contributorship-based authorship threshold? | No, mere model output does not satisfy the original-idea or development threshold. |
| Does one model’s transformation of another model’s output necessarily remove the first model’s constraints? | No, transforming another model’s output does not necessarily clear its constraints. |
Also worth reading: Can Prompts Qualify for Copyright?: Section 202.1 in 2026 — Length or Human Expression?: Can Prompts Qualify for Copyright?: · Why Your Product Launch Needs a Pre-Filing IP Check: Why Your Product Launch Needs · Recompute Your PTA: Why the Face Value Hides Recoverable Days: Recompute Your PTA: Why the