Some Expert Systems Need Common Sense

1 Background

1.1 Definition of Expert Systems

Expert systems are a class of artificial intelligence programs designed to emulate the decision-making ability of a human expert within a specific, narrowly defined domain. They typically consist of a knowledge base—a collection of facts and rules about a particular field—and an inference engine that applies logical reasoning to the knowledge to derive conclusions or recommendations. Unlike general-purpose AI, expert systems are built for tasks such as medical diagnosis, equipment fault detection, or financial planning, where formalized expertise can be encoded as explicit rules. Their strength lies in consistency and the ability to explain reasoning steps, but their scope is deliberately restricted to the domain for which they were constructed.

1.2 The Common Sense Problem

The common sense problem refers to the difficulty of equipping AI systems with the vast, implicit, and often unstated knowledge about everyday life that humans take for granted. Expert systems, which rely on formalized domain rules, frequently fail when faced with situations that require an understanding of basic physical, social, or temporal realities—knowledge that is rarely captured in specialized rule sets.

1.2.1 Tacit Knowledge vs. Formal Rules

Tacit knowledge encompasses the informal, experiential understanding that humans acquire through living in the world—for example, knowing that a cup may spill if knocked, that water flows downhill, or that people generally do not schedule meetings at 3 a.m. Formal rules, by contrast, are explicit, logical statements that can be written in a programming language. The mismatch between these two types of knowledge means that expert systems can reason correctly within their formal framework but may produce absurd or dangerous results when the input deviates from expected patterns. Bridging this gap requires representing the countless unspoken assumptions that underlie human expertise.

1.2.2 Scope and Importance

The scope of common sense extends far beyond any single domain. It includes spatial reasoning, temporal ordering, causality, social norms, and basic physics. For expert systems that operate in real-world environments—such as medical decision support tools, autonomous robots, or scheduling software—the absence of common sense can lead to brittleness: the system may work flawlessly in controlled settings but fail catastrophically when presented with an unexpected or ambiguous input. Addressing common sense is therefore crucial for deploying expert systems in practical, dynamic contexts.

2 Historical Context

2.1 Early Expert Systems

The development of expert systems began in the 1960s and 1970s, driven by the goal of capturing specialized human knowledge in a machine-readable form. Two landmark systems illustrate both the promise and the limitations of this approach.

2.1.1 MYCIN (Medical Diagnosis)

MYCIN, developed at Stanford University in the early 1970s, was designed to diagnose bacterial infections and recommend antibiotic therapy. It employed a rule-based inference engine that could explain its reasoning. MYCIN achieved a diagnostic accuracy comparable to that of human experts in its narrow domain. However, it lacked any understanding of anatomy, physiology, or patient context beyond its rule set. For instance, MYCIN might recommend a treatment that is correct for the infection but inappropriate for a patient with a known allergy—simply because the allergy information was not encoded in its rules unless explicitly added.

2.1.2 DENDRAL (Chemical Analysis)

DENDRAL, created in the mid-1960s, applied heuristic rules to interpret mass spectrometry data and deduce the molecular structure of organic compounds. It was one of the first successful expert systems and demonstrated the power of formalized domain knowledge. Yet DENDRAL, like MYCIN, had no concept of the physical properties of molecules beyond its chemical rules. It could not reason about why a given structure might be unstable under laboratory conditions or why certain measurements might be anomalous due to trivial environmental factors—knowledge that a human chemist would apply intuitively.

2.2 The Cyc Project

In the mid-1980s, Douglas Lenat launched the Cyc project with the ambitious goal of building a comprehensive common sense knowledge base. Instead of restricting expertise to a narrow domain, Cyc aimed to encode millions of everyday facts and rules that would allow AI systems to reason about the world as humans do.

2.2.1 Foundational Knowledge Base

The Cyc team manually entered a vast collection of concepts and axioms covering time, space, causality, objects, events, and social interactions. For example, Cyc includes representations of ideas such as “if you drop something, it will fall to the ground” or “people cannot be in two places at once.” The knowledge base was structured using a formal ontology and a specialized representation language called CycL. Over decades, Cyc grew to contain tens of millions of assertions.

2.2.2 Limitations Encountered

Despite its scale, Cyc encountered significant obstacles. First, the manual encoding of common sense proved to be a monumental and never-ending task—every new domain required adding thousands of new axioms. Second, the knowledge base suffered from brittleness: a single missing axiom could cause the system to fail in unexpected ways. Third, Cyc’s formal representation struggled with the vagueness and context-dependence of everyday knowledge. For instance, “a cup holds water” is generally true but not when the cup is upside down or has a hole. Handling such exceptions required an explosion of additional rules. Cyc demonstrated that common sense is not a finite corpus but a deeply interconnected web of knowledge.

3 Practical Limitations from Lack of Common Sense

3.1 Brittleness in Unexpected Inputs

Expert systems that lack common sense reasoning are said to be “brittle”: they operate correctly only within the narrow range of inputs anticipated by their designers. When confronted with a novel or slightly off-norm situation, they may produce nonsensical or erroneous outputs. For example, a scheduling expert system might suggest a meeting at 2 a.m. if it has no rule that people typically sleep at night, or a medical diagnostic system might ignore the patient’s stated age if age ranges are not explicitly linked to treatment contraindications.

3.2 Inability to Handle Ambiguity

Natural language and real-world scenarios are rife with ambiguity. Expert systems often rely on precise, unambiguous formal statements, so they fail when input is vague or metaphorical. A medical system might read “the patient has a cold” and, lacking common sense about idioms, interpret it as hypothermia. Similarly, a robotic system tasked with “clearing the table” might sweep everything onto the floor if it does not understand that fragile items should be handled carefully. Common sense provides the implicit context that disambiguates such expressions.

3.3 Case Studies

3.3.1 Medical Misdiagnoses

Several documented examples illustrate the consequences of missing common sense in medical expert systems. One case involved a system that correctly diagnosed a rare disease but recommended a treatment that required the patient to fast. The system did not account for the fact that the patient was an infant who could not safely fast for long hours—a piece of common knowledge that any human pediatrician would apply. Another example occurred when a diagnostic system interpreted lab results with decimal points misplaced due to a data entry error, and, lacking common sense about plausible ranges, confidently recommended surgery for a nonexistent condition.

3.3.2 Scheduling Failures

Scheduling expert systems have also suffered from a lack of common sense. In a famous anecdote, a system for university course timetabling assigned a lecture to a room that was locked on weekends, ignoring the obvious fact that rooms may have access restrictions. Another system optimized a factory production schedule by staggering breaks so that workers never overlapped—saving theoretical seconds but ignoring the social reality that people prefer to take breaks together. These failures arise because expert systems treat the world as a set of logical constraints without the underlying understanding of human behavior, physical access, or social conventions.

4 Approaches to Integrating Common Sense

4.1 Rule-Based Extensions

One straightforward approach is to extend an expert system’s rule set with explicitly encoded common sense axioms. This method builds directly on the traditional symbolic paradigm.

4.1.1 Handcrafted Axioms

Handcrafted axioms are formal logical statements that capture common sense knowledge. Engineers create these axioms by analyzing the gaps in the system’s reasoning and adding rules that reflect general world understanding. For instance, a medical expert system might be augmented with axioms such as “A patient who is an infant should not be prescribed medications with a known risk for infants” or “All patients have a weight, and medications in doses by weight must reference that weight.”

4.1.1.1 Example: Time and Space Primitives

A concrete example is the addition of time and space primitives. An expert system for logistics might be given axioms that encode basic temporal facts: “An event cannot occur before its start time,” “A person cannot attend two meetings at the same time,” or “An object cannot be in two locations simultaneously.” Spatial primitives might include “An object inside a container is smaller than the container’s interior” or “A chair can support a human’s weight.” While effective for limited scenarios, handcrafting axioms is labor-intensive and prone to incompleteness—an infinite number of potential common sense facts exist.

4.2 Machine Learning Methods

Recent advances in machine learning offer alternative approaches to capturing common sense without explicit manual encoding.

4.2.1 Neural-Symbolic Integration

Neural-symbolic integration combines the pattern recognition strength of neural networks with the logical reasoning of symbolic systems. In this paradigm, a neural network is trained on large corpora of text or images to learn implicit common sense associations (e.g., “objects fall when dropped”). The output of the neural network is then fed as soft constraints or probabilities into a symbolic reasoning engine. This allows the expert system to flexibly incorporate common sense, but the quality depends on the training data and the network’s ability to generalize. A key challenge is ensuring that the learned common sense is reliable and does not introduce logical contradictions.

4.2.2 Large Language Models as Common Sense Sources

Large language models (LLMs) such as GPT-4 have been trained on massive text datasets and exhibit a remarkable ability to answer common sense questions. Researchers have explored using LLMs as a “common sense oracle” for expert systems: when the system encounters a situation that requires everyday reasoning, it queries the LLM. For example, if a medical expert system needs to know whether a patient who is a toddler can swallow a large pill, it could ask the LLM for a plausible answer. This approach avoids explicit knowledge base construction but introduces issues of output reliability, consistency, and the potential for generating incorrect or biased facts.

4.3 Hybrid Architectures

Hybrid architectures combine multiple techniques to leverage the strengths of each while mitigating their weaknesses.

4.3.1 Modular Design

A modular hybrid system separates the domain-specific expert system from a common sense module. The domain expert system handles the core reasoning using its formal rules, while the common sense module—possibly implemented as a separate knowledge base, a neural model, or an LLM—is consulted only when the expert system encounters ambiguous or missing information. This design preserves the explainability and reliability of the symbolic component while adding flexibility. However, it requires careful engineering to define the interface and to decide when to invoke the common sense module.

Dialogue-based refinement involves an interactive loop where the expert system, when uncertain, asks a human user for clarification or common sense guidance. This approach acknowledges that some common sense knowledge is best supplied by a human in context. For example, a scheduling system might ask, “Is it acceptable for employees to work overnight?” if it lacks that information. Over time, the system can store the answer and reuse it. Dialogue-based refinement reduces the need for a prebuilt common sense repository but relies on human availability and may be impractical for real-time autonomous systems.

5 Contemporary Applications

5.1 Medical Decision Support

Contemporary medical decision support systems increasingly incorporate common sense constraints. For example, systems used for drug interaction checking now include rules about patient demographics (e.g., pediatric dosing, pregnancy) and lifestyle (e.g., alcohol consumption) that were absent in early expert systems. Some use large language models to parse clinical notes and flag inconsistencies—such as a recommended procedure that contradicts the patient’s documented allergies. However, the need for rigorous validation remains, as incorrect common sense inferences could have serious consequences.

5.2 Autonomous Robotics

Autonomous robots, especially those used in domestic or service settings, require a robust understanding of common sense physics and social norms.

A robot tasked with navigating a cluttered home must know that a lamp on a table can be knocked over, that doors can be opened, and that floors may be slippery. Modern robots use a combination of simulation, sensor data, and learned models to approximate such knowledge. For instance, a robot may be trained on a dataset of household objects to predict whether a given item is likely to break if bumped. Expert system rules are still used for planning high-level tasks, but common sense is increasingly provided by neural networks that have been trained on physical interaction data.

5.3 Personal Digital Assistants

Personal digital assistants such as Siri, Alexa, and Google Assistant have evolved from simple command responders to systems that attempt to understand user intent in context.

5.3.1 Contextual Understanding

These assistants rely on large language models and cloud-based knowledge graphs to infer common sense. For example, if a user says “Remind me to buy milk when I’m near the store,” the assistant must understand that “near the store” means within a certain geographic proximity, that stores have operating hours, and that milk is perishable. Although these systems are not pure expert systems, they incorporate expert system components for tasks like scheduling. The integration of common sense via machine learning has dramatically improved their ability to handle ambiguous requests, though they still occasionally fail on tasks that require deep reasoning about social or physical contexts.

6 Future Directions

6.1 Open Technical Challenges

Despite progress, significant obstacles remain in equipping expert systems with reliable common sense.

6.1.1 Scalability of Knowledge Acquisition

Both manual and automated approaches to acquiring common sense knowledge face scaling issues. Manual encoding, as exemplified by Cyc, cannot keep pace with the breadth of human knowledge. Machine learning methods require enormous, diverse, and carefully curated datasets; they often capture statistical correlations rather than genuine understanding. Scaling to cover all domains of everyday experience—from cooking to traffic rules to etiquette—remains a daunting task.

6.1.2 Validation of Common Sense Rules

Common sense is not always universal; it varies across cultures, contexts, and individuals. Validating that a particular piece of common sense holds in the intended deployment environment is a nontrivial challenge. For example, the rule “People shake hands when greeting” is true in many Western cultures but not in others. An expert system deployed globally must either tailor its common sense knowledge to local norms or use a fallback that avoids errors. Formal verification of common sense rules in a logical framework is also difficult because many common sense facts are defaults (true in most cases but with exceptions).

6.2 Potential Paradigm Shifts

Emerging research directions may fundamentally change how common sense is integrated into AI systems.

6.2.1 Embodied AI and Learning from Interaction

Embodied AI refers to systems that learn by physically interacting with the world, much like humans do. A robot that explores a house can learn that glass breaks, that doors require force to open, and that sunlight creates shadows—without being told explicitly. By combining reinforcement learning with symbolic reasoning, such systems can develop a grounded common sense that is more robust than knowledge extracted from text. This paradigm shift emphasizes learning from experience rather than from precompiled facts, promising systems that can adapt to novel environments.

6.2.2 Crowdsourced Common Sense Repositories

Another avenue is the construction of large, publicly editable common sense knowledge bases through crowdsourcing. Projects like Open Mind Common Sense, ConceptNet, and Wikidata have collected millions of facts contributed by volunteers. While these repositories are noisy and incomplete, they provide a foundation for expert systems to draw upon. Future work may focus on automated methods to cross-validate assertions, resolve contradictions, and integrate with formal ontologies. If successful, crowdsourced repositories could serve as a living, evolving common sense resource for all AI systems.