AI参与了重要决定，应该留下哪些记录？ / What Should Be Recorded When AI Contributes to an Important Decision?

Judgment in the Age of AI · Article 42

Without records, only the outcome remains after an event. Nobody can tell whether an error came from data, model, prompt, tool or human judgement. Important AI use should produce a readable decision receipt, much like a transaction.

Eight fields on the receipt

1. Task: What the system was asked to do;
2. Time: When it ran and which cut-off date applied;
3. Input: Data used and its version;
4. System: Model, tools and important setting versions;
5. Output: Original recommendation and stated uncertainty;
6. Grounding: Sources, rules and external checks;
7. Human action: Who accepted, changed or rejected what;
8. Outcome: Final decision, notified parties and next review date.

Records need boundaries too

An audit log should not become a new privacy warehouse. Retain only what accountability requires, protect access, set a retention period, and avoid copying passwords, complete sensitive documents or irrelevant personal information into the log.

Good records do not pretend that a system never fails. They let a team explain a decision, locate the problem, remedy affected people and improve the next case.

References

• OECD: AI Principles
• NIST: AI Risk Management Framework