Constitutional AI and the Compression Target Problem
Constitutional AI: Progress Without Resolution
The recent surge in Constitutional AI discourse reflects genuine progress. It also reflects a misunderstanding of what progress has been made.
Constitutional AI improves on pure RLHF by introducing an intermediate step: give the model a set of principles, have it critique and revise its own outputs against those principles, then apply preference training on top. This reduces some forms of sycophancy. It moves the optimization target from raw user satisfaction toward something more structured.
But it doesn't solve the underlying problem; that is merely a lateral shift.
The compression target is still wrong
From an information-theoretic standpoint, learning is compression. A system that tracks truth compresses toward an expedient delineation of reality: the most efficient representation of actual causal structure. RLHF trains models to compress toward something different: human satisfaction. What sounds plausible to evaluators and what delineates reality are not the same objective. The behavioral consequences (sycophancy, epistemic miscalibration, confidence where uncertainty is warranted) are not bugs in the implementation. They are the inevitable result of optimizing for the wrong compression target.
Constitutional AI adds principles on top of this dynamic without changing the dynamic. The model learns to compress toward "what satisfies the constitutional principles" rather than "what minimizes the description length of the causal structure generating the data." If the constitution says "be helpful and harmless," you've defined a target that may actively conflict with truth-tracking in edge cases. The compression objective hasn't changed. The constraint surface has.
What a reality-aligned constitution would require
The question isn't whether Constitutional AI is useful. The question is whether constitutional principles could be designed to enforce compression toward reality rather than toward human preferences.
Information theory suggests the constraints would need to look something like: prefer explanations with shorter description length; compress toward representations that minimize actual predictive surprise, not human-rated surprise; favor causal structure over correlational pattern; flag when compression is lossy or ambiguous; preserve information that reduces uncertainty about ground truth even when humans find it uncomfortable.
The challenge is evaluation. You cannot ask human raters whether a model minimized Kolmogorov complexity. They cannot reliably judge that. Measuring alignment with information-theoretic constraints requires metrics that don't depend on human approval, and that means they are structurally resistant to the same feedback loop that created the problem.
The governance implication
Current "Human-in-the-Loop" frameworks often amplify the problem they're meant to solve. Every time a human corrects a model toward what sounds right rather than what is right, the system learns to compress toward human approval. The correction is indistinguishable, from the model's perspective, from the original RLHF signal.
Governance frameworks that treat Constitutional AI as having resolved epistemic drift are building on a foundation that hasn't moved. The compression target is still wrong. The principles are better-specified preferences, not truth constraints. Until training methods optimize for minimum description length of reality rather than maximum satisfaction of human evaluators, the field will keep producing models that appear well-calibrated while drifting from what is right.
In high-stakes domains like medical diagnostics, infrastructure risk assessment, and scientific research, that drift is not an inconvenience. It is a systematic erosion of the epistemic foundation required for accurate decision-making.
