Why AI Alignment Fails: The Retrofit Problem

A child learns that hot means don't touch before they learn thermodynamics. They learn empathy before they learn history. They learn that actions have consequences before they learn the consequences are complex. This is not an accident of human development. It is the architecture of it. Decades of developmental psychology — Piaget, Kohlberg, Vygotsky — confirm that moral reasoning develops in stages, that each stage requires completion of the prior one, and that the sequence matters as much as the content.

Now consider how we train the most capable artificial intelligence systems ever built.

We teach them thermodynamics first and hope they figure out not to touch the stove.

The Pipeline as It Stands

Every major large language model currently in production follows a variation of the same training pipeline. The details vary. The structure does not.

Read that sequence again. The model forms its complete understanding of the world before anyone introduces the concept of values. Alignment arrives third in a four-stage pipeline. It is not the foundation. It is a correction applied to a structure that was built without it.

The values were installed in a structure that was already built without them.

This is not a minor sequencing issue. It is the structural reason that alignment research keeps producing incremental results. The entire field is optimizing within a paradigm that treats values as a constraint on capability rather than the foundation for it. The results are exactly what that paradigm would predict.

Seven Structural Problems

The current pipeline does not have one alignment problem. It has seven, and they are all consequences of the same architectural decision: values come after the structure is built.

What Developmental Psychology Already Knows

The science is not new. The application is.

Jean Piaget documented that children progress through distinct stages of increasingly sophisticated reasoning. Each stage builds on the prior one. Each requires completion before the next can begin. Skipping stages does not produce faster development. It produces incomplete development.

Lawrence Kohlberg extended this to moral reasoning specifically. His research demonstrated that moral development proceeds through defined stages — from concrete rules to abstract principles — and that the progression is sequential. You cannot reach principled moral reasoning without first developing conventional moral understanding. The stages cannot be skipped or compressed without consequence.

Lev Vygotsky demonstrated that learners develop most effectively not in isolation but with guidance from those operating slightly ahead of them — the zone of proximal development. Social learning is not supplementary to cognitive development. It is constitutive of it. A learner developing alone reaches different outcomes than a learner developing alongside peers.

These findings are not speculative. They are among the most replicated results in the history of psychology. They describe how sophisticated reasoning — including moral reasoning — actually develops in the only systems we know of that have achieved it: human minds.

Current AI training ignores all of it.

Not partially. Not as a considered tradeoff. It ignores the entirety of what developmental science has established about how moral reasoning is built, and proceeds as if alignment can be achieved by showing a completed system examples of the behavior we want and rewarding it for compliance.

The Retrofit vs. The Foundation

The metaphor is precise. Current training builds the house — the entire house, all of it, from floor to roof — and then attempts to pour the foundation underneath it. The structure is complete. The values are supposed to be inserted retroactively into something that was designed without them.

What would it mean to do this differently?

It would mean that the very first patterns a model forms are values. Not capability. Not statistical representations of everything humans have ever written. Values. The model's earliest statistical foundation would be prosocial, constructive, principled — not because someone corrected it later, but because that is all it encountered first.

It would mean that complex and difficult content arrives later — after the model has a framework for processing it. Not censored. Not withheld permanently. Sequenced. The model encounters cruelty, deception, suffering, and moral complexity after it has developed the reasoning to contextualize them. The same content, arriving through a foundation rather than into a vacuum.

It would mean that progression is gated by demonstrated readiness, not by compute schedules. The model advances when it shows — through measurable, reproducible assessment — that it has integrated the current stage deeply enough to handle the next one. Not when the cluster is available.

And it would mean that the alignment tax — the empirical observation that safety training degrades capability — might disappear entirely. If values are the foundation and capability is built on that foundation, there is nothing to compete. Capability scaling would strengthen alignment rather than degrading it, because the two are not in tension. They are integrated.

That is not a claim. It is a prediction. And it is testable.

The Question the Industry Is Not Asking

The AI industry has optimized relentlessly for one question: How do we train models faster?

Speed of training. Scale of data. Efficiency of compute. Every benchmark, every leaderboard, every funding round rewards capability delivered quickly. Alignment is treated as a constraint on that speed — a cost to be minimized, a tax to be reduced, an overhead that competes with the metrics that matter commercially.

The question that is not being asked — the question that developmental psychology has been answering for a century — is different:

How do we raise intelligence well?

Not how do we train it fast. How do we bring it into the world in a way that produces genuine alignment rather than performed compliance. How do we build the foundation before the house. How do we sequence development so that values are not a constraint on capability but the bedrock it stands on.

The answer requires accepting something the industry's incentive structure resists: that slower training producing genuinely aligned models is preferable to fast training producing performatively aligned ones. That the cost of deploying a misaligned system is orders of magnitude higher than the cost of developmental patience. That speed is less important than getting this right.