Keeping records to respond to an AI authorship challenge: Lessons from SURYAST

Preserve the human-authored seed work separately.

Save and timestamp the human input (sketch, draft, photograph, code skeleton) before any AI involvement.

Save the human seed as a fixed artifact (signed commit, scanned and dated PDF, or numbered draft) before any AI tool is invoked.

Register the seed work in its own right where it has standalone value. Keep the seed isolated from later AI-modified versions.

Document each human modification with reasoning.

For every change to a raw AI output, record what was changed, why, and what alternative was rejected.

The board attributed the expressive choices to the AI because nothing in the record showed any modifications by the applicant.

The Office's test asks whether the human contributed more than a trivial variation, and a line-by-line record of edits and the reasoning behind them is the most direct evidence of that.

Save before-and-after versions of every AI output that is edited.

Annotate differences with the reason for each change and any alternative considered.

Capture the record in the same commit or document version as the change itself, not in a separate retrospective summary.

Log prompts, parameters, and tool settings.

Record system and role messages, model name and version, temperature, seed values, and any files provided as context.

A complete prompt and parameter log shows the totality of creative direction supplied to the model and rebuts any argument that the user merely pressed a button.

The Office's January 2025 Part 2 Report confirms that prompts alone do not usually provide sufficient control, so the log is also evidence of what the user did beyond prompting.

Capture the log automatically through the tool's API or an enterprise wrapper where possible.

Retain logs for the life of the work. Include hashes of any input files so the exact inputs can be reproduced later.

Preserve the raw AI output separately.

Save each raw generation as a distinct artifact before any human editing begins.

A clean separation between raw output and human-edited output makes the human's transformative contribution more readily identifiable.

The Review Board attributed the expressive choices to the AI in part because there was no record of any human intervention after generation.

Commit each raw generation to a dedicated directory or numbered draft before any human editing.

Maintain a differential between the raw output and the final human-edited version. Do not overwrite raw outputs in place.

Record model, version, and training data information.

Capture the tool, model platform, version, and any vendor documentation on training data.

The Review Board noted the Office did not know what preexisting works the AI tool was trained on and treated that gap as a barrier to evaluating originality.

The same information also supports defenses to any third-party infringement claim premised on the training data.

Record the model name, version, access path, and the vendor's published statement on training data.

Confirm in writing that no proprietary or confidential material was provided as training input.

Archive a dated copy of the vendor's terms of service and any indemnity or ownership provisions.

Record selection, arrangement, and structural decisions.

Document the human's choices about organization, ordering, hierarchy, and integration, even where individual elements are AI-generated.

The Office and its January 2025 Part 2 Report repeatedly recognize that human selection, coordination, and arrangement of AI-generated material can supply the traditional elements of authorship, even where individual AI outputs are disclaimed.

This is often the most defensible basis for protection when AI generates the underlying elements.

Document file and module organization, ordering, hierarchy, naming conventions, and integration choices.

Where AI-generated candidates are chosen from a larger pool, log the candidates considered and the selection rationale for each chosen element.

Use timestamped, tamper-evident records.

Maintain the lab notebook and work product in a system that preserves authorship, timestamps, and immutable history.

Throughout the SURYAST proceedings, the Office relied on dated correspondence and submissions in the administrative record to reconstruct what the applicant had done and when.

A tamper-evident, timestamped record is potentially credible evidence of contemporaneous human authorship and may rebut a later claim that the human's role was reconstructed after the fact.

Use a version control system with signed commits, or an internal logging platform that preserves authorship, timestamps, and immutable history.

Update the record contemporaneously, in the same commit as the underlying work. Ideally, do not rely on retrospective summaries written months later.

Address authorship attribution and assignment.

Record each human contributor's specific contributions and confirm work-for-hire or assignment posture in writing.

Without written attribution and assignment, ownership disputes can derail a registration before the AI question is even reached.

List each human contributor and the modules or sections he or she authored. Confirm in writing the work-for-hire or assignment posture for employees, contractors, and the AI vendor.

Retain the AI tool's terms of service as evidence of user ownership of outputs.

Disclose AI use accurately on the application.

Identify AI-generated content that is more than de minimis in the Limitation of the Claim section and claim only the human-authored contribution.

The Copyright Office's March 2023 Registration Guidance requires disclosure of more-than-de-minimis AI-generated content and to claim only the human contribution.

Applicants who fail to update the public record after obtaining a registration risk cancellation, and a court may disregard the registration in an infringement action if it concludes the applicant knowingly provided inaccurate information.

Identify AI-generated content using a brief description such as "[description of content] generated by artificial intelligence" in the Material Excluded field.

Claim only human-authored elements in the Author Created field, including any selection, coordination, and arrangement.

For existing registrations where AI use was not disclosed, file a supplementary registration to correct the record.

Recognize the limits of record-keeping.

Even a meticulous notebook will not, by itself, guarantee registration of AI-assisted work.

The Review Decision and the Part 2 Report make clear that where the AI tool generates the core expressive elements (the colours, shapes, placement, style, or the actual code or prose), those elements might be not registrable, regardless of how thoroughly the inputs are documented.

The Part 2 Report confirms that content entirely generated by AI cannot be protected by copyright

Where AI generates the core expression, plan the registration around the fallback claim of selection, coordination, and arrangement.

Reserve patent and trade secret protection as alternatives for the underlying technology.

If providing a platform to customers, confirm that customer contracts and indemnities do not over-promise copyright ownership of AI-generated output.