Abstract
This paper explores the qualitative differences between human decision-making and early heuristic programming models in Artificial Intelligence. By analyzing psychological experiments on perception and operative thinking, the authors argue that human cognition is a "productive" process involving the active construction of mental models and dynamic "sliding" standards. The study identifies distinct cognitive levels—from raw perception to complex speech-thinking—and provides empirical evidence for the operational lag between understanding a principle and executing a plan.
Introduction
In the early 1970s, the field of Artificial Intelligence was largely defined by the "Information Processing Psychology" approach, which viewed thinking as symbolic manipulation and heuristic search. This paper, authored by Zavalishina, Lomov, and Rubachin, critiques this paradigm for its inherent "passivity." The authors contend that human decision-making (DM) is not merely a selection from pre-defined options but a generative process where the subject actively extracts and transforms environmental data to build new relational systems.
The Critique of Passive AI
The core of the authors' argument lies in the distinction between "mechanical sorting" and "productive bailing." Early AI models, such as Newell and Simon’s Logic Theorist, operated on fixed problem spaces. In contrast, human DM is characterized by its ability to redefine the problem space dynamically. The authors propose that DM occurs across multiple hierarchical levels, where higher-level "speech-thinking" processes guide lower-level "sensory-perceptive" actions, creating a flexible, non-linear cognitive architecture.
Experimental Methodology
To map the micro-structure of DM, the researchers conducted two primary types of experiments:
- Perceptive-Identificative Tasks: Subjects were presented with a "Perception Alphabet" consisting of 18 stimuli at varying resolutions and noise levels. Eye-tracking technology was employed to distinguish between the initial "perceptive studying" phase and the final "identification" phase.
- Operative Thinking Tasks: A spatial reasoning puzzle required subjects to arrange six objects based on logical constraints. This task aimed to observe how individuals transition from grasping a general "action principle" to developing a concrete "operational scheme."
Quantitative Results and Eye Movement Dynamics
The empirical data gathered from eye-tracking revealed significant differences in cognitive effort between the stages of perception. As shown in Table 1, the initial phase of "perceptive studying" is characterized by a high density of information gathering, requiring nearly three times the duration and number of fixations compared to the final identification of the object.
The researchers also identified three distinct search strategies:
- Chaotic: Unplanned movements driven by stimulus intensity.
- Choosing: Systematic comparison against internal or external samples.
- Semantic: Hypothesis-driven search based on cause-and-effect associations.
Cognitive Mechanisms: The "Sliding" Standard
A pivotal finding of the study is the concept of the "sliding" standard. Humans do not rely on static templates for recognition. Instead, they utilize a dynamic internal alphabet that narrows as the image formation progresses through four phases: from general attributes (size, color) to sharp contours, fine details, and finally, a corrected mental representation. This allows for high flexibility in noisy or ambiguous environments.
Furthermore, the study highlighted an "operational lag" in spatial tasks. Subjects often identified the underlying logic of a solution (the "Action Principle") well before they could formulate a viable execution plan. This suggests that "knowing what" and "knowing how" are distinct, though interrelated, cognitive functions.
Conclusion and Implications for AI
The authors conclude that for AI to achieve true intelligence, it must move beyond heuristic search and toward "selective actualization." Future systems should mimic the human ability to dynamically adjust internal standards and actively construct new information systems based on context. While the simulation of these "productive" traits remains a significant challenge, this research provides a foundational psychological framework for developing more sophisticated, human-like decision-making algorithms.
Summary updated