Judgment in the Age of AI · Article 35
Writing a message and sending it are different permissions. The first creates an inspectable draft; the second changes relationships, commits resources and may form a formal record. The closer convenience comes to acting in your name, the higher the confirmation threshold should be.
Create a risk-sorted outbox
| Message | Reasonable setting |
|---|---|
| Meeting confirmations and low-risk internal reminders | Automatic drafting; sampled review or rule-based approval before sending |
| Client replies, public statements, cost or commitments | Human confirmation of recipient, facts and tone for every message |
| Dismissal, discipline, legal, medical, safety or intimate matters | Never auto-send; the responsible person writes or approves it personally |
A four-part pre-send check
- People: Are the recipient, copies and attachments correct?
- Matter: Do facts, dates, amounts and links have sources?
- Authority: Are you entitled to make the commitments in the message?
- Consequence: Could you tolerate publication or accidental delivery?
Use a send delay, daily limit and complete log. A new recipient, unusual attachment, sensitive term or out-of-scope request should automatically return the message to draft.
The safest default is not “AI may send unless I stop it”, but “AI prepares until I authorise”.
References
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.