Cognitive automation uses AI to encode and operationalize human skills and knowledge. It handles narrowly defined problems or tasks supported by many precise examples, enhancing computers without replicating human minds. section 1.10
Automation is not adaptability
A cartoon character can act out every scene it was drawn for; a living being can adapt to the unexpected. Similarly, a language model may produce human-sounding answers while remaining dependent on situations represented in training or programming. Robust automation is difficult because every relevant scenario must be anticipated. section 1.10
Use these distinctions when evaluating an AI system:
- Treat deep-learning automation as powerful local generalization: it maps familiar input situations to outputs and performs best when new cases resemble past data. section 19.2.4
- Separate cognitive automation from cognitive autonomy: automation handles trained or programmed situations, while intelligence adapts to unknown situations and plans for the long term. section 1.10 section 19.2.4
- Expect broad generalization to require handling unknown unknowns within a domain, such as a self-driving car responding safely to any situation. section 19.2.4
