The advent of new AI tools has caused many to worry about the displacement effect of automation. The common reaction to this problem follows a Pigouvian intuition: automating capital harms workers, so it should be taxed. This paper argues that this Pigouvian intuition is misguided. Capital harms workers at the extensive margin of automation, but at the intensive margin, more capital in a task that has already been automated raises wages via capital deepening. I show that if the Planner can stipulate how much more expensive labor must be than capital before automation can occur, then capital taxes should be zero, recovering the celebrated result of Atkinson and Stiglitz (1976) in spite of the dependence of relative wages on equilibrium automation. Calibrating occupation-level exposure to automation to match aggregate substitution between capital and labor, I find that the optimal response to the US tax system is to require labor to be 31.8% more expensive than capital before automation can occur, but this reduces to 10.8% with an optimized non-linear income tax. Simulating the introduction of AI, I find that the optimal threshold rule is sharply increasing in the expected productivity growth of capital.