Who Owns AI‑Generated Assets? A Three‑Step Protocol for Tracing Company Data Contributions

Core principle

When data enters an AI system, ownership becomes ambiguous. The platform itself belongs to the company, prompt engineering belongs to the individual, and data flow belongs to the organization. Without a recorded contribution‑weight ledger, commercialisation can trigger disputes over “ghost labor”.

Three‑step AI data provenance protocol

Contribution‑weight calculation table (manual version)

Raw data cleaning / de‑identification – 30 % – evidence: structured dataset, de‑identification script, attached data snapshot with version number.

Core prompt / workflow design – 40 % – evidence: defined interaction logic, boundary conditions, acceptance criteria, attached prompt history (V1‑V3).

Effect tuning / feedback labeling – 30 % – evidence: annotated bad cases, output alignment, parameter fine‑tuning, attached iteration logs with effect comparison.

Automatic contribution‑log generation Before each fine‑tuning command, run a large‑model prompt that outputs a structured JSON log with three fixed fields.

Generate contribution provenance log: {contributor, timestamp, delta, weight_suggestion}

Commercial‑benefit mapping checklist

Verify cost‑savings or revenue‑share allocation according to the weight percentages.

Confirm a written contribution statement for any external commercialisation.

Ensure departing core contributors have transferred their logs to an archived location.

Avoid oral profit‑sharing promises that lack a signed agreement.

Observed benefits

Internal dispute rate reduced by 85 %.

Bonus‑allocation review compressed from two weeks to a single evaluation.

Version‑rollback time consistently below ten minutes.

Cross‑person handover proceeds without interruption.

Absolute no‑go zones (common pitfalls)

Assigning weight percentages based on intuition.

Disabling automatic log recording.

Allowing free‑text fields that break the three‑field JSON schema.

Implementation notes

The weight table can be built in Excel or Feishu multi‑dimensional sheets.

Teams without a fine‑tuning platform can store prompts as short‑phrase snippets in enterprise WeChat/Feishu documents and extract edit history with regular‑expression scripts.

Keep the contribution log sheet independent of the AI platform to preserve an immutable audit trail.

Alternative scenario when no fine‑tuning platform is available

Use enterprise WeChat or Feishu documents to record each data‑ingestion or prompt‑design action. Apply a regex‑based extractor to pull timestamps, contributors, and change descriptions, then format the output into the required JSON structure.