The same tool that hollows out learning can build it.

It would be convenient if AI were simply the villain of this newsletter. It is not. A Harvard randomized controlled trial found that an AI tutor matched or beat in-class active learning, on the condition that the interaction was active: the tool questioned and prompted the learner instead of handing over answers. The study.

It is a preprint, so hold the result lightly. But what makes the tutor work is the learner doing the retrieving. Take that away and the identical tool becomes the crutch this newsletter has been describing. The tool did not change; the role of the human did.

Comprehension debt: the gap between what gets shipped and what the team can still reason about.

This is the optimistic turn, and it is also the actionable one. The lever is not a policy that bans the tools, which trades away real speed and breeds resentment. The lever is how the team uses them: active and questioning, versus passive and accepting. You do not have to choose between AI's speed and the team's depth, but the tool will not choose the active pattern for you. Making it the default is what keeps the depth your estimates are already spending against.

One for your next review:

A billing feature stores money as a floating-point number of dollars. Tests pass. Finance reports that some invoices are off by a cent, and a few totals do not reconcile.

Watch what gets blamed first: the type, or the logic. An engineer who has lost a day to this reaches straight for the storage type, knowing decimals and binary floats do not mix. Someone who hasn't reaches for another rounding patch. The penny that will not reconcile is almost never a logic error, and hearing someone say so before they go hunting tells you the model is in the room.