Write the rule, and let the AI check itself against it
The Plain-Language AI Policy Card
This week in AI — governance stopped being a document nobody reads and became a step you can run at the moment of work.
What happened
On 4 August, Mistral released Shieldstral, a small open-weights model (3B parameters, Apache-2.0) that does one narrow job: you hand it a rule written in plain English, and it scores whether a piece of text or an image breaks that rule — no retraining, decided the moment you ask. Mistral's performance and hardware figures are its own vendor claims; the part worth teaching is the method: the policy is written in your words, at the time you use it. [Mistral, 2026-08-04]
Why it matters to you
Most people assume "AI safety" is something the big labs handle for you. This flips it. The rule is yours. For an advisor, that is the difference between hoping the AI behaves and defining, in one line, what "off-limits" means for your own drafts:
- "Flag any message that promises or implies a specific return."
- "Flag any draft that names a specific client."
- "Flag anything that reads as personal financial advice."
The check becomes a step you own — not a black box you trust.
How to do it — this week, on your own work (never on client files)
- Write your rule in one plain sentence. Start with the thing that would actually get you in trouble: "This draft must not promise or imply a specific investment return."
- Test the rule, not a client. Run three of your own past drafts through it — two clean, one you know bends the rule. Does the flag land where you expect?
- Set your line. What score means "stop and review," and what means "fine to send"?
- Write your escalation. When it flags, what do you do — rewrite it yourself, or send it to a colleague?
- Keep the record. One line per flagged item, so you can show your process later.
The lesson
AI governance used to mean a policy document that sat in a drawer. The shift this week is that a rule you can write in a single sentence can now be enforced at the moment of work. The skill was never the model — it's writing the rule clearly enough that a machine, or a new hire, could apply it the same way every time.