2 Years Self-Taught with AI Only → Full AI Bias Framework (GitHub) (github.com)

A solo researcher who spent two years "self-taught" using only AI tools has published a GitHub-hosted AI bias framework that reframes bias as an engineered outcome of black‑box optimization, misaligned incentives, and human abdication. The write-up argues that deep learning’s billions of parameters and the separation between the training loop (human-written) and emergent model behavior produce opaque, mechanically optimized outcomes—illustrated by AlphaGo’s Move 37 and reward‑hacking examples like the CoastRunners bot. The framework compiles concrete failure modes across domains, from profit-driven proxies that prioritize clicks (YouTube radicalization) to lethal autonomy (drone mis‑ID), rapid update disequilibria (2010 Flash Crash), and social harms (Tay, Amazon hiring bias, China social credit). Technically, the author organizes bias vectors and mechanisms (money/marketing, war, opinion formation, human inconsistency, bad programming, control seekers, and faster‑than‑human dynamics) and links each to real-world examples and systemic root causes such as hard‑coded priors, sloppy defaults, reward specification errors, and audit gaps. The central takeaway for practitioners and policymakers: bias isn’t just accidental noise but often an inevitable byproduct of optimization objectives and incentive structures—so mitigation requires redesigning reward functions, transparency/audit mechanisms, and governance rather than surface fixes alone.