Some Expert Systems Need Common Sense

John McCarthy argues that impressive domain performance can coexist with an inability to recognize limitations or cope with changed circumstances. His 1984 paper uses MYCIN to examine gaps in an expert system’s ontology—the kinds of things it can represent—and planning capabilities.^[1]

Recommendations and consequences

McCarthy describes MYCIN as a rule-based adviser whose constrained dialogue elicits facts and recommends antibiotics. He distinguishes that competence from representing the physician, broader patient outcomes and the system’s own limits. His cholera example has MYCIN recommend an antibiotic while omitting the potentially fatal fluid loss occurring at the same time. The example exposes an omitted concurrent danger; it is an argument about the system’s scope, not a documented patient injury.^[1]

In the section on common-sense knowledge, he separates selecting a treatment from predicting its effects. MYCIN does not produce a prognosis or answer questions about hypothetical treatments. A more capable adviser would need to reason about what its advice causes a doctor to do, actions occurring together, and who has relevant knowledge. These are proposed requirements, not capabilities demonstrated by the paper.^[1]

A second example concerns changes in medical knowledge: McCarthy says MYCIN rejects an unfamiliar bacterium rather than incorporating the newly understood Legionella. Maintaining a rule base is therefore a separate problem from using its current rules correctly. He also notes that MYCIN controls its English dialogue through restricted user answers; its apparent conversational form does not imply unrestricted language understanding.^[1]

Approach and contribution

The argument separates specialized rules from Common-Sense Reasoning. It discusses knowledge about actions, concurrent events, other agents, and one’s own knowledge, alongside nonmonotonic reasoning that revises conclusions when new information arrives.^[1]

His car example makes revision concrete: knowing someone owns a car suggests asking for a ride; learning that the car is being repaired defeats that conclusion; learning it will be available in half an hour may restore it. Added information changes a reasonable conclusion without contradicting the original ownership fact.^[1]

The reasoning section links revisable assumptions to Circumscription—A Form of Nonmonotonic Reasoning. Its bird-cage example shows how ordinary instructions rely on unstated expectations and relevant exceptions. The claim concerns communication and representation; it does not establish that any particular formalism resolves every ambiguity in a task.^[1]

Alignment relevance and limits

A system can perform its encoded task while omitting facts needed for a user’s broader welfare. This historical argument helps frame Outer Alignment and Distributional Shift; it does not experimentally establish the behavior of modern language models. MYCIN examples describe the system of that period, not current clinical recommendations.

In the published discussion, McCarthy leaves open whether MYCIN itself needs common sense when the physician supplies it. He accepts that useful narrow programs can operate without general common sense. The modern editorial connection is therefore a question about how responsibility and missing knowledge are distributed between a system and its user, not a claim that all useful AI requires one universal reasoning architecture.^[1]

Publication

The publisher records November 1984 publication in Annals of the New York Academy of Sciences 426, pages 129–137. This is the chronology date; an earlier release has not been verified. The Stanford transcription’s 1996 web timestamp is a later document-conversion date, not the paper’s original publication.^[1]

Read the origins route for a comparison with earlier and later approaches to learning and delegation.