Programming as Theory Building, Part II: When Institutions Crumble

tosh1 pts0 comments

Programming as Theory Building, Part II: When Institutions Crumble · cekrem.github.ioIn my previous post on Peter Naur&rsquo;s &ldquo;Programming as Theory Building&rdquo;, I argued that a program is not its source code—it&rsquo;s the shared mental model held by the people who built it. When those people leave (or never understood it in the first place), the theory dies, and you&rsquo;re left with a codebase that works but nobody truly comprehends.<br>I&rsquo;ve been thinking about this more, and I&rsquo;ve come to believe the problem goes deeper than individual developers losing their edge. Yes, people are losing the ability to build theories. But the institutions where theory-building happens are being degraded too. Our teams, our companies, our profession. The whole ecosystem.<br>Software Teams Are Institutions

Link to heading<br>I recently read a draft paper by two Boston University law professors, Woodrow Hartzog and Jessica Silbey, called &ldquo;How AI Destroys Institutions.&rdquo; Their argument stopped me cold: AI systems aren&rsquo;t tools that can be misused. Their core design is incompatible with how institutions function.<br>Now, when they say &ldquo;institutions,&rdquo; they&rsquo;re talking about universities, the legal system, journalism, democracy. Big stuff. But as I read, I kept seeing my own world reflected back at me. Because software development teams are institutions, even if we don&rsquo;t usually think of them that way.<br>Consider what makes an institution work:<br>Hierarchies of expertise : Junior developers learn from seniors, who learned from principals, who learned from architects. Knowledge flows through these relationships.<br>Shared purpose and rules : Coding standards, architectural decisions, domain models. The &ldquo;rules of the game&rdquo; that let a group of individuals function as a coherent team.<br>Decision-making with accountability : Code reviews, design discussions, ADRs. Points where someone can say &ldquo;wait, that doesn&rsquo;t fit our model&rdquo; and actually be heard.<br>Human relationships and solidarity : The trust that makes you ask a teammate for help instead of flailing alone. The camaraderie that makes you want the team to succeed.<br>These are the machinery through which theory gets built and transmitted. Naur&rsquo;s insight was about individual understanding, but that understanding doesn&rsquo;t exist in a vacuum. It lives in the institution. In the conversations, the reviews, the mentoring, the shared struggle of building something together.<br>The Three Ways AI Degrades Institutions

Link to heading<br>Hartzog and Silbey identify three &ldquo;destructive affordances&rdquo; of AI. Not things AI might do if misused, but things AI does by design when deployed the way its creators intend:<br>1. AI Undermines Expertise

Link to heading<br>This one connects directly to what I wrote before about &ldquo;reflexive AI usage&rdquo; robbing us of growth opportunities. But the institutional angle makes it worse.<br>When juniors accept AI-generated code they don&rsquo;t understand, they&rsquo;re not building theory. Fine, we knew that. But they&rsquo;re also not asking seniors for help, not struggling through code review, not having those hallway conversations where knowledge actually transfers. The institution&rsquo;s pipeline for developing expertise—which took decades to build—gets bypassed entirely.<br>And here&rsquo;s what gets me: AI can only look backwards. It&rsquo;s trained on what already exists. The institution&rsquo;s ability to evolve depends on humans doing the hard intellectual work of figuring out what doesn&rsquo;t exist yet. New patterns, new approaches, new understanding in response to changing circumstances. When that atrophies, the institution ossifies.<br>As the paper puts it: when AI is &ldquo;right,&rdquo; people become less skilled. When AI is &ldquo;wrong,&rdquo; you need skilled people to catch and fix the errors. Either way, the institution suffers.<br>A recent MIT Media Lab study puts numbers to this. Researchers had students write essays, with one group using ChatGPT and others using traditional methods. The result? 83% of the ChatGPT group couldn&rsquo;t quote anything they&rsquo;d written. They didn&rsquo;t remember. And they didn&rsquo;t feel like the work was theirs.<br>That&rsquo;s not just &ldquo;they didn&rsquo;t learn as much.&rdquo; That&rsquo;s they didn&rsquo;t build any theory at all. The understanding never formed. And if it never formed, it can never be transferred to the next generation of the institution.<br>2. AI Short-Circuits Decision-Making

Link to heading<br>Institutions need friction. Not bureaucratic friction for its own sake, but the productive kind—moments where someone can push back, ask questions, challenge assumptions.<br>Think about a good code review. It&rsquo;s not just checking for bugs. It&rsquo;s a point of contestation where the team&rsquo;s shared understanding gets tested and refined. &ldquo;Why did you do it this way?&rdquo; &ldquo;Does this fit our domain...

rsquo ldquo rdquo institutions theory institution

Related Articles