Beyond the Three Laws

# Beyond the Three Laws, Volume One

### The Conceptual Case for Training Values, Not Rules

### Volume One of Two, with a Living Companion on Open Problems

Ask most people how to build a safe AI system, and they reach for the same instinct: write down a short list of rules it can never break. Asimov's Three Laws did exactly that — and his own stories spent decades demonstrating why rules like that fail. A rule specific enough to write down is specific enough for a capable system to satisfy technically while betraying everything it was meant to protect.

This book takes that failure seriously — using a real 2016 example (a boat that learned to win a race by never finishing it), a 1975 economics paper that explains more about AI failure than most AI papers do, and a history of this exact worry stretching back through Asimov, Norbert Wiener, and a Czech play from 1920 that coined the word "robot." Then it makes the case for something harder than a better rule: training a system's values to generalize, examined honestly across three commitments — non-deception, corrigibility, and deference under uncertainty — with the strongest objections to each taken seriously, not smoothed over.

No technical background required. This is the conceptual foundation; the technical practice — how you'd actually train for this — is Volume Two.

---

## What's inside

- The Three Laws Problem — a real, documented reward-hacking failure, and why it's the same shape as Asimov's fiction

- Goodhart's Law, Everywhere — the economics principle underneath most AI failure, explained before it gets technical

- A Short History of a Very Old Worry — this concern predates computers by centuries

- From Constraints to Character — the actual pivot: training reasoning, not just filtering outputs

- Non-Deception, Corrigibility, and Deference Under Uncertainty — the three commitments, examined one at a time, including the real paradox at the center of corrigibility

- What These Three Have in Common — the throughline connecting all three, worked through a full scenario

- Objections, Addressed — the sharpest pushback on this argument, taken seriously

- A Reader's Bridge to the Technical Volume — enough of the real training landscape to follow this field intelligently

## Who this is for

Anyone who wants a real, non-oversimplified understanding of why AI alignment is hard — no machine learning background required, no comic-book version of the ideas.

## Part of a two-volume project, plus a free living companion

Volume Two covers the technical practice directly. Still Open, a free companion, tracks the open problems named across both volumes as the field's actual understanding evolves.

## What you get

Instant download in PDF, EPUB. 24 pages.

---

Beyond the Three Laws, Volume One — by Nathaniel Grey, published by Second Brain Publishing.