Abstract
This paper proposes a formal engineering and philosophical framework to demystify artificial intelligence by distinguishing it from traditional rule-based software. By introducing a "decidability" taxonomy (Type-I, II, and III decisions) and a specific notation for stochastic systems, the authors argue that the responsibility for verifying AI truth ultimately shifts from the developer to the user. This framework provides a grounded methodology for deciding when and how to safely deploy AI systems in real-world engineering contexts.
Introduction
The AI industry is currently caught between "enthusiastic imagination"—often manifesting as AGI hype—and "conceptual vagueness" driven by metaphorical language. From a rigorous engineering perspective, there is a critical lack of clarity regarding how AI fundamentally differs from traditional software.
Traditional software is deterministic: developers write explicit rules, and the machine executes them. AI, conversely, is stochastic; its "hallucinations" are often inherent features of its design rather than bugs. This creates a "trust gap." If we cannot specify an AI's behavior in advance or verify its outputs against a single expected result, how can we treat it as a reliable engineering artifact? This paper seeks to provide an epistemic framework to navigate this ambiguity.
Theoretical Framework
The authors move away from viewing AI as a scientific discovery akin to a law of physics, instead treating it as a Philosophical Object within an engineering context. To facilitate this, the paper introduces several key notations and roles.
Core Notations
- Program $\Phi$ (Phi): Traditional software that follows strict, inherent rules.
- Program $Я$ (Ya): AI software that mimics human cognition and exhibits adaptive, autonomous behavior. The use of the Cyrillic letter $Я$ (meaning "I" in Russian) represents the self-referential, mirroring nature of AI.
- The Epistemic Triad: A model consisting of the Epistemic Subject (ES) (the thinker), the Method (EM) (the process of thought), and the Object (EO) (the subject of thought).
To ground these concepts in a practical ecosystem, the authors define specific personas involved in the lifecycle of Program $Я$.
The Nature of AI Outputs
A central contribution of this work is the Decision Taxonomy, a three-tiered system used to evaluate AI outputs:
- Type-I: Empirical truth (Is the statement "true" in the physical world?).
- Type-II: Internal consistency (Does the output $y$ logically follow from input $x$?).
- Type-III: Predictability (Can the behavior be governed by preset rules?).
In this framework, "hallucinations" are recognized as a core mechanism of generative models. As noted by industry pioneers, the expectation that computers should only be "right or wrong" is a legacy of deterministic $\Phi$-thinking that does not apply to the stochastic nature of $Я$.
The Shift of Responsibility
The paper derives several logical conclusions (C1–C7) regarding the relationship between developers and users. A crucial insight is that neither the AI model nor the AI researcher ("David") can truly "know" if an output is Type-I decidable (empirically true). Only the user ("Jack") possesses the situational context required to capture value by verifying the output against reality.
Because the developer cannot guarantee empirical truth, the burden of verification shifts. The authors conclude that users must answer three reflective questions before deploying or relying on AI:
- Can I ignore a false result?
- Can I verify the result at an affordable cost?
- Can I bear the consequences if verification fails?
Conclusion and Future Work
This grounded view of AI suggests that human progress relies on $\Phi$-thinking (the impulse to adhere to rules), while AI ($Я$-thinking) emulates and blends these rules with non-rational patterns. While this paper provides a theoretical foundation, it does not yet offer an automated metric for calculating "decidability" in real-time.
Future research should explore whether rule-based autonomous artifacts (Program $\Phi$) can be designed as "guardrails" to monitor and intervene in the stochastic $Я$-relation, effectively building software that understands the philosophical boundaries of the AI it oversees.
Summary updated