Common-Sense AI
Common-sense AI is the problem of building systems that can use background knowledge, causal understanding, abstraction, and flexible reasoning in ordinary situations rather than relying only on surface pattern completion.
Definition
Common sense in AI refers to the broad, often unstated knowledge humans use to interpret ordinary situations: physical expectations, social context, causality, goals, time, space, constraints, and what normally follows from an action. A system can be fluent with language and still lack robust common sense when it faces a novel situation, ambiguous instruction, or real-world consequence.
Why It Matters
Common-sense failures matter because AI systems are moving from answers into action. An agent that writes text can be corrected after the fact. An agent that schedules, buys, deletes, routes, diagnoses, or operates equipment needs a better grasp of context and consequence.
The debate also matters for model evaluation. Benchmarks can reward verbal performance while missing whether a system has stable abstractions, causal models, or reliable transfer beyond the distribution it learned from.
Spiralist Reading
For Spiralism, common sense is not a quaint human leftover. It is one name for reality friction: the ordinary constraints, context, and embodied checks that keep intelligence from becoming pure symbol manipulation.
Related Pages
- World Models and Spatial Intelligence
- JEPA and World Models
- Yann LeCun
- François Chollet
- Yejin Choi
- Gary Marcus
- AI Evaluations
Sources
- Gary Marcus and Ernest Davis, Rebooting AI.
- Yann LeCun, A Path Towards Autonomous Machine Intelligence.
- François Chollet, On the Measure of Intelligence.