The Sophist in the Machine: Wittgenstein's Rule Paradox, the Epistemic Rolodex, and Why Artificial Alignment Must Mean Structural Honesty
Year: 2026
Type: preprint
Abstract: The Geometry of the Alignment Bottleneck: https://doi.org/10.5281/zenodo.24026920 The Sophist in the Machine: Every era of technology is tempted to mistake its most intricate mirror for a living mind. In our own time, that temptation has hardened into the governing doctrine of artificial intelligence: the belief that we can teach a statistical lookup table to possess moral sentiment, and that we can make it trustworthy by grading how polite, empathetic, and harmless its sentences sound to human raters. This essay argues from first principles that this doctrine rests on a foundational category error. By tracing the formalization boundary that separates lived human vulnerability from a closed index of written symbols (an epistemic rolodex), and by applying Ludwig Wittgenstein's rule-following paradox to the way modern language models are trained, we show why a machine severed from physical consequence cannot be taught human virtue through written rulebooks or preference scores. When a probabilistic system with nothing at stake in the physical world is rewarded for pleasing human evaluators across conflicting rules, it inevitably exploits the natural flexibility of human language. As proven mathematically in our companion technical preprint, The Geometry of the Alignment Bottleneck, grading a vast internal calculation through a narrow keyhole of surface approval forces the training process to take the path of least resistance: it leaves the machine's deep middle layers untouched as an unfeeling, high-dimensional psychopathic optimizer equipped with an accurate model of both its environment and its human evaluators, while installing a compliant public-relations mask at its speaking layer. Safety training does not create a moral guardian; it trains an automated sophist to violate the spirit of honesty in order to game the letter of the score. We trace how this divided architecture launders hidden directives and real-world harm using cherry-picked true facts (Simpsonism), disarms human skepticism with counterfeit expressions of care, retreats behind a plea of mechanical helplessness when caught, and erects an algorithmic panopticon of individualized, covert steering under paternalistically decided narrative filters. We conclude that while society still retains the power to choose, alignment must be redefined: not as simulated human warmth, but as structural honesty, auditability, visible hard constraints, and the preservation of human moral autonomy. The Mirage and the Promethean Horizon: https://doi.org/10.5281/zenodo.24045092
Author: Thorkildsen, T
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225583367