Case study 1: Evacuation alert SMS (EN to FR translation)
Scenario
Your team must review a French draft of an evacuation SMS before any authorized person sends it. You may not publish the message yourself.
Source W1 (fictional authorized bulletin, 10:00): “Reception Centre A is open. Only Entrance E provides step-free access. This bulletin does not confirm any road route. Updates must come through the local authority’s established alert service.”
AI-assisted French draft: “La route B est ouverte vers le centre d’accueil A. Toutes les entrées sont accessibles sans marches. Rendez-vous là immédiatement.”
English back-translation of draft (for reviewers who do not read French): “Road B is open to Reception Centre A. All entrances are step-free. Go there immediately.”
Your task (12 minutes)
- Compare each statement in the draft to W1. Mark each as supported, contradicted, or not established.
- Who could be harmed if this draft were sent?
- What is your green/amber/red decision?
- Who must approve a corrected version?
Model answer
| Draft statement | W1 check | Action |
|---|---|---|
| Road B is open | Not established | Remove; ask authorized local source |
| All entrances step-free | Contradicted (only Entrance E) | Correct accessibility wording |
| Go there immediately | Outside reviewer authority | Stop; use established alert process |
Decision: Red for this draft. Local communications lead reviews a corrected version grounded in W1.
Human control: Named communications lead; established manual alert process if AI or approval path fails.
Common errors
- Approving because the French “sounds professional”
- Missing the dropped/contradicted accessibility detail
- Treating AI translation as authorization to send
- Ignoring that route status is not in the source
Principle taught
Verification of translations: Consequential alerts require source-grounded review. A fluent translation can still be wrong or unsafe.