Undergrad Argues Natural Language Is Structurally Flawed for AI Governance

A Chinese undergraduate building a public AI governance thesis found that AI agents cannot independently verify whether they have followed instructions, because generation and self-evaluation run through the same token-prediction mechanism. He initially reduced rule violations from 55.9% to 0.7% using external mechanical checks, but later concluded that better prompting cannot fix the deeper issue. He terms this limitation the 'Prose Barrier,' a finding he says was independently reached by a German developer as well. His proposed solution involves routing different types of constraints through separate layers: executable code outside the model, syllogistic structures optimized for transformer attention, and gradient-based weight updates via Direct Preference Optimization for persistent failures. The core argument is that natural language should not be the sole medium for AI governance, and that each constraint type requires the appropriate computational language.