# Alignment Tax The performance cost — in capability, speed, flexibility, or user experience — incurred by making an AI model safe, helpful, and aligned with human values. Safety training ([[RLHF]], [[Constitutional AI]]) can make models more cautious, more likely to refuse requests, slower to respond (if using extended safety checks), and less capable on certain tasks where safety restrictions prevent full expression of the model's knowledge. ## The rhetorical function "Alignment tax" frames safety as a *cost* rather than a *feature* — something you pay rather than something you gain. This framing has political consequences: **In the AI arms race:** If alignment carries a capability tax, then the safest lab produces the weakest model. This creates a competitive disincentive for safety investment and fuels the argument that the U.S. can't afford to be more cautious than China: safety-first means second-place. The "alignment tax" framing directly supports the [[AI Arms Race]] logic. **In the open-weight debate:** The alignment tax is what gets removed when open-weight models are fine-tuned to strip safety training. An "uncensored" model is marketed as more capable precisely because it doesn't pay the alignment tax — it will answer any question, generate any content, follow any instruction. **In safety research:** Reducing the alignment tax — making safety training impose less capability cost — is a core research goal. If alignment were free (no capability cost), the competitive argument against safety would collapse. Anthropic and others invest heavily in techniques that improve safety without reducing capability, explicitly to drive the alignment tax toward zero. ## See Also - [[Alignment]] — the technical challenge - [[AI Arms Race]] — the competitive logic that makes the tax matter - [[Guardrails]] — the user-facing expression of the alignment tax