Introduction
In the late 1960s, the field of Artificial Intelligence was predominantly focused on formal logic and theorem proving. However, human belief systems rarely operate on strict logical principles; instead, they are built upon "psychological implications"—subjective expectations such as "if person A likes person B, then person A will likely help person B." This 1969 study by Lawrence Tesler, Horace Enea, and Kenneth Mark Colby explores a computational approach to modeling this messy, non-logical evaluation of "truth."
The authors address the problem of credibility estimation: determining how a person decides if a new statement is true, false, or uncertain based on their existing worldview. By constructing a model that navigates a web of subjective beliefs, the researchers aimed to move AI beyond simple fact retrieval toward a more nuanced simulation of human cognition.
Modeling the Human Belief System
The researchers developed their model using MLISP on a PDP-6/10 system. The core of the model was a Conceptual Graph representing the beliefs of a single informant—a 30-year-old mother—regarding parent-child relations. This "intensive" design allowed for a deep, coherent modeling of a specific domain rather than a shallow survey of many individuals.
The database consisted of approximately 700 entries, categorized into facts, rules, definitions, and classifications. Definitions were particularly important for establishing semantic links between concepts.
To evaluate a new proposition, the system calculated a credibility score based on two primary factors: Foundation (direct evidence for or against the statement) and Consistency (how well the statement aligns with its own implications). These numerical scores were then mapped to qualitative judgments to reflect human-like certainty.
Resource-Bounded Search
A critical innovation in this paper was the introduction of a "work" factor to control the search algorithm. To prevent the system from searching the graph indefinitely, the algorithm was allotted a finite budget of "units." It would "spend" these units on different search paths—for instance, 3 units might be spent finding a similar verb, while 2 units might be spent exploring an analogy.
This deterministic approach allowed the researchers to measure the "effort" required to reach a conclusion, providing a more controlled alternative to the random probability models used in contemporary research.
Key Insights and Results
The study yielded several significant findings regarding the structure of human beliefs and the effectiveness of search strategies:
- The Power of Supersets: Searching "up" the hierarchy (toward general categories or supersets) proved far more effective for judging credibility than searching "down" (toward specific instances). A general rule about "people" was often more informative for evaluating a statement about a specific individual than searching for that individual's past behaviors.
- Search Effectiveness: The inclusion of "similar" and "opposite" relations significantly expanded the algorithm's ability to find relevant evidence that human experimenters also deemed important.
- Mathematical Linearity: The authors found that scaling the initial credibility of all beliefs by a factor x resulted in a final computed credibility also scaled by x. This demonstrated that their complex search and quotient formulas maintained mathematical consistency.
- Belief Stability: Despite the inherent complexity of human thought, the informant's core beliefs remained remarkably stable over several months, suggesting that intensive modeling of a specific domain is a viable path for AI development.
Limitations and Future Outlook
The primary bottleneck identified was the manual translation of natural language into the model's LISP-like format, a process described as "cumbersome and time-consuming." The authors noted that a "conceptual parser" would be essential for scaling such a system. Additionally, the 700-entry database was relatively "sparse," often leading to trivial results unless the propositions were closely related to existing entries.
While this paper focused on the algorithm's mechanics, it laid the groundwork for future validation studies comparing computer-generated scores directly against human judgments, pushing AI toward a more sophisticated understanding of subjective human belief.
Summary updated