AI code generation shatters authorship-based knowledge metrics

Software engineering has long relied on a simple proxy: the person who wrote a chunk of code must understand it. Metrics like truck factor (how many developers a team can lose before a project stalls), degree-of-authorship, and degree-of-knowledge all depend on this inference. But with AI generating entire modules, the version-control attribution still shows a human as author—yet that human may have only read and merged the code, not written it. As Brett Wheeler's new preprint argues, the same authorship footprint now supports full, partial, or zero comprehension. The metrics still return numbers, but those numbers have become uncoupled from the knowledge they were meant to estimate. This is the “substrate collapse”: the measurement surface has changed, so the instrument no longer works.

Wheelers builds on this by stating a falsifiable prediction: teams with a healthy authorship-derived truck factor but low comprehension-measured retention will suffer incident-resolution failures that the authorship metric does not predict. If true, it confirms that comprehension—not authorship—is the real signal. The paper deliberately leaves open how to build a comprehension-grounded instrument at system and team scale, calling it the field's next open measurement problem. This is not a refinement of existing metrics but a call for a new category of tools that measure actual understanding, such as knowledge surveys, artifact mapping, or active learning traces. For engineering leaders managing AI-enhanced teams, the implications are immediate: you can't trust old proxies for bus-factor or code ownership.