Toward a New Science of Common Sense

Ronald J. Brachman, Hector J. LevesqueAAAI 2022 · DOI

Introduction

Despite decades of rapid progress in artificial intelligence, modern systems remain fundamentally "brittle." While models like GPT-4 or advanced autonomous vehicles show flashes of brilliance, they frequently fail when faced with simple, real-world surprises. A self-driving car might navigate a complex intersection perfectly but fail to interpret a simple, unusual gesture from a pedestrian. This paper argues that the root of this fragility is a lack of common sense—not as a collection of facts, but as a functional ability to apply everyday knowledge to practical goals.

The Fallacy of the "Fact Calculator"

Historically, AI research has conflated common sense with commonsense knowledge. Projects like Cyc spent decades cataloging millions of assertions (e.g., "humans need air to breathe"), yet having a database of facts is not the same as knowing when and how to use them to solve a problem in the moment.

The authors argue that we are missing a global architecture that allows an agent to break out of routine behavior and use background knowledge to interpret unexpected experiences. Logic-based systems often work in isolation from goals and context; knowing that a wheelbarrow weighs less than a grizzly bear is mere trivia unless an agent can use that knowledge to decide whether to move an obstacle out of a driveway.

A Framework for a New Science

As a position paper, this work proposes a multi-pillar framework to move AI beyond narrow expertise and toward a more robust "New Science of Common Sense." This framework suggests that a truly commonsense system requires four key components:

  • Experiential Knowledge: Moving beyond static facts to systems that remember, generalize, and organize past experiences to inform future actions.
  • Plausible Reasoning: Instead of exhaustive logical deduction, common sense requires "rapid, plausibly sound inference" that can be executed quickly enough to be useful in real-time environments.
  • Cognitive Architecture: A "metareasoning" layer is necessary to decide when to use routine behavior, when to trigger common sense, and when to escalate to heavy-duty analytical thinking.
  • Advice-Taking: The ability for an autonomous system to be "incorrigible"—to take and act on human advice to correct its behavior dynamically.

The Cognitive Continuum and the Expertise Trap

The authors suggest that common sense sits in a "quasi-rational" space between Daniel Kahneman’s "System 1" (fast, intuitive, effortless) and "System 2" (slow, analytical, effortful). In humans, much of our day is routine. Common sense is our "break-out" mechanism; it is what we invoke when something unusual happens. It is the bridge that prevents us from failing when our routine meets its limit.

Paradoxically, AI's biggest successes—such as Go, facial recognition, and medical diagnosis—are forms of narrow expertise that actually contribute to brittleness. Because these systems lack a broad foundation of common sense to fall back on, they have no safety net when they encounter scenarios outside their training data. Despite being a goal since John McCarthy’s 1958 papers, no AI system currently possesses the common sense of a six-year-old.

Future Directions

The authors acknowledge that they do not yet have a final prescription for building these systems. Instead, they pose several core research questions for the field:

  1. How can we represent experiential knowledge so it is easily accessible via analogy?
  2. What are the appropriate tests for common sense, given that standard benchmarks often just test for "knowledge fragments"?
  3. How can we build an architecture that smoothly transitions between routine, common sense, and specialized expertise?

The paper concludes that until the AI community treats common sense as a "first-class subject of study"—rather than a side effect of more data—AI will remain confined to narrow, fragile domains.

Summary updated

BibTeX
@inproceedings{brachman2022toward,
  author = {Ronald J. Brachman and Hector J. Levesque},
  title = {{Toward a New Science of Common Sense}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21485},
}
Figure