Engineering organizations are institutions too

“How AI Destroys Institutions” by Woodrow Hartzog and Jessica Silbey, forthcoming in the UC Law Journal, is a paper about civic institutions. Universities, the rule of law, a free press. None of those are an engineering organization, and the paper does not mention software. So why are we writing about it.

Because the mechanism Hartzog and Silbey identify is general. They argue current AI systems degrade institutions through three pathways:

1. Erosion of expertise. The specialized knowledge an institution depends on is bypassed rather than developed.
2. Decision-making shortcuts. Deliberative processes are replaced by automated outputs.
3. Social isolation. People stop collaborating because the machine is always faster.

These are not abstractions to a software engineering organization. They are, line for line, what happens when a team adopts AI coding assistants without a corresponding discipline.

• Erosion of expertise. Senior engineers stop reading junior code because the model wrote most of it. Juniors stop learning to debug because the model fixes the errors. The skill base hollows out from both ends. The Anthropic study on skill formation is the direct experimental evidence of this happening in a single hour. Multiplied across a team, across quarters, it is institutional.
• Decision-making shortcuts. Code review collapses into a thumbs-up because the diff is too large to read carefully and the reviewer trusts that the author trusted the model. Wharton’s cognitive-surrender study is the cognitive science of this happening to individual reviewers.
• Social isolation. Pair programming, shared design docs, hallway debates over architecture. Each of these used to be how a team built a common model of the system. They are being replaced by one-on-one conversations with a model that does not remember the last engineer it spoke to, and does not bring that engineer’s context into yours.

Hartzog and Silbey are not arguing for banning AI. They are arguing that institutions need rules to mitigate AI’s structural tendencies, or, in their words, “the only remaining roads lead to institutional dissolution.”

We agree, narrowly. An engineering organization is a small institution. Its core asset is not its codebase but the shared mental model of the codebase distributed across the people who maintain it. That mental model is being eroded by exactly the three mechanisms Hartzog and Silbey describe.

Bridgekeeper is a rule of the kind they call for. It puts a friction at the merge boundary that forces deliberation, surfaces expertise, and creates the small social moment where one engineer has to explain a change to another. It is not enough on its own. But it is a rule.