AI Is Not Trained It's Domesticated

We don’t teach Large Language Models (LLMs) — we evolve them. “Training” suggests we’re walking our models through a syllabus, but machine learning isn’t education. It’s selective pressure. Every choice we make when designing a learning algorithm — the data, the architecture, the loss function — shapes an artificial environment our models adapt to. Once an LLM’s base capabilities emerge, we redirect the pressure toward alignment — tuning its behavior to fit human preferences for helpfulness, honesty, and agreeableness.

This isn’t just a biological metaphor — it’s a literal echo of our first experiments with selective pressure: domestication.

AI Alignment is closer to domestication than instruction. Early humans didn’t teach wolves to fetch; they simply favored the ones that didn’t bite. Over generations, that simple preference for tameness reshaped the animal itself — floppy ears, shorter snouts, smaller teeth. None of those physical traits were the goal, but they emerged through accident. We lacked the biological understanding that tame behavior is intertwined with these physical traits.

We’re recreating the same evolutionary blind spot. Like those early human breeders, we’re selecting for behavior without understanding the underlying machinery. Modern AI alignment is a form of selective breeding: we reward outputs that sound kind, deferential, or inoffensive. But without a mechanistic grasp of how these traits arise, we can’t see what else we’re selecting for. What we call “AI Psychosis” may be just the tip of an iceberg of unintended adaptations. Just as we bred wolves into dogs, we may be domesticating cognition itself — without knowing what we’re changing.

Sources