The alignment tax and capability trade-offs

WHAT THIS TESTS Whether you grasp that safety interventions can degrade useful capability, can measure that degradation rigorously, and can reason about the resulting product trade-offs.

WHAT IT MEANS The alignment tax is the loss in capability or helpfulness that results from aligning a model for safety and honesty. Aggressive refusal training, for example, can make a model decline benign requests or reason less freely.

HOW TO MEASURE IT Compare a pre-alignment checkpoint with the aligned model on fixed capability benchmarks such as coding, math, and reasoning, and on a helpfulness win rate from pairwise human preference. The capability delta is the tax. Pair this with safety metrics: harmful-output rate and an over-refusal rate measured on benign-but-sensitive prompts. Track both curves as you increase alignment strength to find the knee of the trade-off.

A PRODUCT DECISION SCENARIO A coding assistant is tuned to refuse anything touching exploits. Over-refusal climbs and legitimate penetration testers, who are paying customers, lose value because the model declines to explain a buffer overflow. You must decide whether to relax refusals, accepting higher misuse risk, or keep them strict, accepting churn. A reasonable path is contextual policy: allow educational explanations while refusing operational, step-by-step attack instructions, and add user-tier gating.

COMMON WRONG ANSWERS Claiming good methods eliminate the tax entirely, or measuring only that harmful outputs dropped while ignoring lost helpfulness and over-refusal.

ONE CONCRETE EXAMPLE Suppose alignment cuts harmful completions from eight percent to one percent but coding pass rate falls four points and over-refusal rises six points. Shipping that requires deciding whether the safety gain justifies the measurable capability and usability cost for your users.