Tariffs, AI, and the Art of Changing Your Mind

Trade policy is boring until it isn't. Three pieces this week make it interesting by actually doing the math.

The textbook case

EconTalk's episode with Doug Irwin on the economics of tariffs and trade is the clearest explanation of what trade deficits actually mean that you'll find anywhere. Irwin, a Dartmouth economist, walks through a point most people miss: dollars spent on imports don't vanish — they come back as asset purchases. The deficit is an accounting identity, not a scorecard. Russ Roberts pushes back on whether this technicality matters when manufacturing jobs are actually disappearing from specific communities, and the exchange is genuinely productive.

The contrarian case

Noah Smith published a piece arguing that tariffs on China could help the global economy — a position that puts him at odds with most of his economist peers. His argument: China's industrial overcapacity is a genuine problem, and tariffs could force a rebalancing that benefits developing nations who'd gain relocated manufacturing. Smith is careful to frame this as "tariffs can be the least bad option in a specific context," not "tariffs are good." The distinction is important and usually lost in the discourse.

The numbers nobody expected

Tyler Cowen followed up on the AI productivity question with a post on new claims about AI productivity improvements, highlighting Yale research tracking 500+ professionals using 13 different LLMs. The finding: each year of model progress reduced task completion time by about 8%. Extrapolated out, that projects to a ~20% US productivity boost over the next decade. Cowen's characteristic move — presenting the data without strong commentary and letting readers draw conclusions — works well here because the numbers genuinely speak for themselves.

The lesson

The best economic thinking this week came from people willing to hold two ideas simultaneously: tariffs are generally bad and sometimes the best available option; AI is overhyped and showing up in the productivity data. Nuance isn't a cop-out. It's the price of actually understanding something.