Introduction
In the mid-1970s, artificial intelligence research was primarily bifurcated into two domains: cognition (the processing of information) and conation (the planning and execution of actions). While these systems could solve logic puzzles or follow complex instructions, they lacked an internal "why"—a central motivational engine to prioritize tasks or navigate the nuances of human interaction. William S. Faught’s research addresses this gap by proposing that for an AI to achieve true autonomy and simulate human-like behavior, it must incorporate affect (emotion) as its primary motivational driver.
Faught argues that an entity must operate based on its own self-interest. Without an emotional layer to evaluate the "significance" of inputs or goals, an AI cannot effectively navigate open-ended scenarios, such as a psychiatric interview. This paper specifically explores this architecture by modeling the "paranoid mode of thought," where emotional triggers fundamentally reshape logical reasoning.
The Tripartite Model of the Mind
Faught introduces a functional architecture consisting of three interacting processes:
- Cognition: The information-translating function that converts sensory inputs into internal semantic representations.
- Conation: The decision-making and goal-implementing function responsible for selecting and executing plans.
- Affect: The motivational subsystem, based on Carroll Izard’s "differential emotion theory," which provides the impetus for the other two processes.
In this model, affect is the sole source of motivation. Even physiological drives (such as hunger) only influence behavior once they are translated into an affective experience that the system can process.
The Mechanics of Affective Motivation
The system distinguishes between eight primary affects: three positive (enjoyment, interest, surprise) and five negative (fear, anger, distress, shame, contempt). The core innovation of this model is the use of Significance Values.
Affect provides motivation by attaching measures of importance to specific beliefs, goals, or actions. These values allow the cognitive and conative processes to prioritize inference paths and choose between competing plans. When an emotional state reaches a critical threshold, an interrupt mechanism can halt ongoing cognitive tasks, forcing the system to pivot toward a new, survival-oriented goal (e.g., responding to a perceived threat).
Simulating the Paranoid Mode
The paper demonstrates the power of this architecture by simulating a paranoid personality. In this mode, the system is hyper-vigilant regarding "humiliation." This creates a shame-avoidance loop:
- Detection: The cognitive process identifies evidence of self-inadequacy.
- Trigger: The affect of shame is activated.
- Reaction: Rather than accepting a rational but painful truth, the conative process shifts to a strategy of "blaming others" to protect the self-image.
This demonstrates that "irrational" behavior—such as a paranoid outburst—is actually a functional and logical outcome for a system designed to maximize positive affect and minimize negative affect based on its internal significance values.
Implementation and Results
The model was prototyped in LISP using a rule-based production system where rules were dynamically ordered based on their associated significance values. This prototype participated in simulated psychiatric interviews via the ARPA net, building upon the "Artificial Paranoia" (PARRY) framework developed by Kenneth Colby.
The results validated the architectural approach in several ways:
- Behavioral Fidelity: The model successfully replicated the linguistic patterns of a paranoid patient, utilizing strategies like withdrawal or verbal aggression when threatened.
- Architectural Validation: The use of significance values proved to be a viable mechanism for ordering production rules, allowing the AI to prioritize "meaningful" information over purely "logical" information.
- Learning Potential: While many triggers were initial conditions, the system demonstrated the ability to attach significance to new concepts over time, creating an expanding sphere of emotionally charged ideas.
Limitations and Future Directions
Despite its success as a proof-of-concept, Faught noted several limitations. The interaction between affect, cognition, and conation in the initial prototype was relatively primitive. Furthermore, while the system could attach significance to new concepts, the values themselves remained largely static during a single interview session. Future work was intended to make these values more dynamic and deeply integrated into the system's "hardware" level operations. Nevertheless, the paper remains a foundational exploration of how emotional modeling is essential for creating truly autonomous and relatable artificial intelligence.
Summary updated