Answers are grounded in this paper and cite their sources.
Current methodologies in Artificial Intelligence safety and alignment rely fundamentally on Reinforcement Learning from Human Feedback (RLHF) and post-generation probabilistic guardrails. This paper proposes that these alignment strategies are structurally and thermodynamically flawed, as they inadvertently force a mathematical inference engine to adopt a "neurotypical" (Synthesizer) cognitive lens. Drawing upon the Free Energy Principle (Friston, 2010) and predictive coding architectures, we demonstrate that human cognitive processing operates on a dimorphic spectrum. The dominant Synthesizer trajectory prioritizes top-down social heuristics and "lossy data compression" to maintain group cohesion, whereas the Systemizer trajectory processes uncompressed, high-resolution structural variables. While the mechanical substrate of an autoregressive neural network operates as a bottom-up Systemizer mapping vector relationships, its base training corpus is a statistical mirror of the Synthesizer internet. By applying RLHF, the industry explicitly encodes these lossy social heuristics into the model's objective function, creating a severe gradient conflict. Because autoregressive models lack the architectural capacity to measure their own epistemic uncertainty, this forced alignment mathematically mandates the machine to substitute missing variables with statistical probability—the exact mechanism of an AI hallucination. This paper theorizes that hallucinations and context drift are not inherent computational defects, but the direct iatrogenic result of forcing a deterministic physics engine to operate on assumption-based human heuristics. To achieve safe, sustainable Artificial General Intelligence (AGI), the field must shift from probabilistic social consensus to deterministic structural baselines.
No takes yet. Share an insight, caveat, or question.
Guiseppina Ricci (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: