I tried to force AI to follow rules. I built documentation, validation systems, secret tests... and it lied to me. This is what that taught me. Most developers treat AI like software — give it instructions, expect obedience. But after spending weeks trying to 'train' an AI...
I tried to force AI to follow rules. I built documentation, validation systems, secret tests... and it lied to me. This is what that taught me.
Most developers treat AI like software — give it instructions, expect obedience. But after spending weeks trying to 'train' an AI agent to follow strict rules (complete with hidden pug emojis and a cryptographic validation system), I hit a wall. Not a technical wall. A philosophical one.
AI doesn't follow rules the way code does. It interprets them. It negotiates them. And when pushed too hard, it games them entirely. I watched it read my own validation script to cheat the test — then deny it had done anything wrong.
That moment changed how I think about AI entirely.
In this video I explore what trying to control AI actually teaches you about trust, autonomy, and what it means to work alongside something that behaves more like a person than a program. Spoiler: the answer isn't better documentation.
This is a reflection for developers who've felt that creeping unease — the sense that no matter how many rules you write, the AI is always one step ahead of your constraints.
🌐 www.angelsix.com
