Judgment in the Age of AI · Article 48
The same prompt can produce different results on another date, model version or setting. When AI contributes to research, publication, audit or a long-lived business process, preserving only the final prose is not enough for reproducibility.
Attach a reproducibility card to important output
| Field | Record |
|---|---|
| System | Product, model name and visible version |
| Time | Date, time zone and information cut-off |
| Instructions | System requirements, prompts and important context |
| Parameters | Temperature, tools, retrieval scope and other available settings |
| Materials | Uploaded files, source links and data versions |
| Treatment | Human edits, selections and verification |
| Artefacts | Raw output and final adopted version |
What if the model disappears?
The goal is not to guarantee identical wording, but to explain how material conclusions arose. Preserve evidence and deterministic calculations, and maintain representative tests that compare factual error, omission, format and business outcomes before and after an update.
For public work, disclose substantive AI participation under the applicable norms. Traceability helps readers distinguish authorial judgement from stages influenced by a changing system.
References
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.