Summary
- Performance in a defined task does not imply general understanding or safe transfer to a new setting.
- Operators need drift tests, refusal rules and a handoff when inputs leave the validated range.
Narrow AI is built for a limited class of predictions or actions. Its value comes from specialisation, but so does its risk: users may assume the system understands context it has never been tested against. A deployment should publish the valid input range, known failure modes and the person who receives exceptions. Monitoring must detect when data or behaviour changes. The next useful evidence is not a broader claim of intelligence, but a record showing that the system refuses, escalates and recovers at the edge of its competence.
