Abstract
In the late 1960s, Kenneth Mark Colby and David Canfield Smith developed ABS1, an early "artificial belief system" designed to model how an intelligent agent processes information from multiple, potentially conflicting sources. This paper details the architecture of ABS1, which treats credibility as a mathematical function of evidence and consistency. By simulating dialogues between humans and the machine, the researchers explored the foundational challenges of computational trust and the formalization of subjective belief states.
Introduction
Long before the advent of modern Large Language Models, the field of artificial intelligence grappled with the problem of "computational epistemology": how can a machine not only store data but also decide whether that data is true? Human belief systems are notoriously difficult to study due to emotional biases and social pressures. To bring these processes under scientific control, Colby and Smith synthesized ABS1, an "imaginary system" acting as a controlled environment for studying credibility.
The core challenge addressed is the resolution of contradiction. If an AI is told "John is a man" by one person and "John is a machine" by another, it must have a mechanism to assign "weight" to this information based on the reliability of the source and the internal consistency of its existing knowledge base.
The ABS1 Architecture
Written in MLISP and running on the PDP-6/10 time-sharing system, ABS1 operates through three distinct phases:
- Talktime: The interactive phase where the system accepts statements, rules, and questions. To avoid the complexities of natural language, syntax is restricted to a Subject-Verb-Modifier format.
- Questiontime: A proactive phase where the system identifies gaps in its logic. For instance, if it knows "Sam is a baseball player" and "Sam is a man," it may ask the human to confirm if the former implies the latter.
- Thinktime: An offline processing phase where the system performs backward-chaining inference and re-evaluates the credibility of its informants based on the consistency of their claims.
Quantifying Credibility and Polarization
A central feature of ABS1 is its attempt to "make up its mind." The system calculates the credibility of any given proposition using a weighted additive function of its foundation (the credibility of beliefs that imply it) and its consistency (how well it fits the existing network).
To avoid perpetual uncertainty, the system is designed to polarize its beliefs. Numerical computations are rounded to force the system into discrete states of "Strong," "Medium," or "Weak" belief or disbelief.
As shown in the mapping above, the system maintains a narrow range for "Neutral" states (0.41 - 0.59), pushing most propositions toward the "Credible" or "Incredible" ends of the spectrum.
Key Mechanisms of Belief Formation
Several innovative concepts were introduced to manage the system's internal logic:
- The "Parent" Seed: The system is initialized to be "credulous," starting with a highly trusted source whose input is accepted as 100% credible. This mimics the developmental stage of a child before critical thinking emerges.
- Temporal Priority: In cases of self-contradiction by an informant, ABS1 assumes the most recent statement reflects the informant's current state.
- The "Self" as Informant: The system treats its own inferred conclusions as coming from an internal informant named "self." This allows the system to evaluate its own "thoughts" using the same mathematical framework applied to external human sources.
Results and Discussion
The researchers demonstrated that ABS1 could successfully perform complex inference chains, such as linking socio-economic rules to answer questions about political attitudes. Notably, the system exhibited "self-correction," where it could find its own "self" to be incredible if internal inferences led to persistent contradictions.
However, the system faced significant limitations. The rigid syntax limited the nuance of communication, and the reliance on human-provided "expectancy rules" meant the system could arrive at conclusions that were formally valid but empirically incorrect. Furthermore, the process of recomputing informant credibility based on the statements they provided introduced risks of circular reasoning.
Despite these constraints, ABS1 remains a landmark in the study of artificial belief, highlighting the enduring relevance of credibility modeling in the face of conflicting information.
Summary updated