AI isn't breaking higher education – it's exposing the cracks

The debt nobody logs

Software engineering has a name for this: technical debt. Ward Cunningham, a computer programmer known for creating the first wiki, coined the term in 1993 to describe the accumulated cost of shipping a not-quite-right solution now and postponing the right one. The metaphor caught on because it captured something engineers already knew: shortcuts compound. Interest accrues. Eventually you're spending more on servicing the debt than on building anything.

Institutions patching without first diagnosing were flipping a coin and paying for the privilege.

What is less known outside of software engineering is that there's a second half to the metaphor. In 2015, Damian Tamburri and colleagues published a study in the Journal of Internet Services and Applications that extended the concept to what they called social debt, defined as the accumulated cost of organizational and social decisions, as opposed to technical ones. Studying a large European aviation software company over six months, they found that social debt wasn't a soft parallel to technical debt but an entangled one. Decisions that used technical means to solve social problems generated both kinds of debt simultaneously in a compounding pattern that couldn't be trivially paid back.

Tamburri and his colleagues created a catalog of "community smells" -- recurring organizational patterns that look normal but signal accumulating debt. It reads, to anyone who has sat on a university committee, like an ethnography of higher education written by researchers who had never visited one.

They describe the architecture hood effect, which involves decisions so dispersed that nobody can be identified as the owner of any particular decision. This produces a nobody's fault dynamic in which accountability dissolves and everyone downstream blames the architects. That's shared governance in its failure mode.

Tamburri and his co-authors described this type of failure and the radio-silence phenomenon that can accompany it. Radio silence is an increasingly formal structure full of regular procedures, in which changes are delayed while people who don't know each other are notified and certified. The researchers measured the delay at half a day to two days per decision, compounding across the organization. In my world, that's a curriculum committee.

They also described organizational silos, in which task decoupling is so high that groups duplicate each other's work while developing what the researchers call tunnel vision. That's the relationship between a student success office and the registrar, and you know it.

However, the finding I would put before a cabinet is this: Of the mitigations the researchers observed institutions deploying against these patterns, roughly 40% failed to produce the intended effect, and some made the situation even worse. Institutions patching without first diagnosing were flipping a coin and paying for the privilege.

How AI is changing things

Here is where higher education does something distinctive -- and not in a good way.

Technical debt is a leadership problem. It's a decision made at the institutional level to accept a known future cost in exchange for a present convenience. It's a governance decision in the most ordinary sense that allocates resources over time and commits successors to obligations they didn't choose.

But higher education has developed an efficient mechanism for making this problem disappear from the leadership conversation. The debt gets logged on a risk register. It's assigned to a CIO. It becomes a line item in an IT governance report that goes to a committee. The box is checked. And in that moment, the problem is successfully reclassified from a governance decision the institution has made about its own future into a technical matter being managed by technical people.

Nothing was solved. The debt is still accruing. What changed is that it's no longer visible at the layer where anyone has the authority to address it.

I want to be precise about why this matters now, because institutions have been doing this for decades without catastrophe. What's different is that AI doesn't sit on top of inherited architecture the way previous technologies did. A learning management system could be bolted onto a 15-week course container without anyone examining whether the container still makes sense. A CRM could be layered over an admissions process built on assumptions about how families made decisions in 1985. The bolt-on worked, more or less, because those tools automated existing steps.

The question isn't whether to patch or rebuild. The question is how do you tell which option you're facing.

AI doesn't automate steps. It redistributes what work is. For example, take a system that can do 60% of the coordination labor in an advising role but none of the relational labor. The inherited assumption that the term advisor represents a coherent unit of work performed by one person holding one position is no longer true. You can respond by buying an AI advising tool and bolting it onto the existing role, which is what most institutions are doing. Or you can ask what the role actually consists of, which almost nobody is doing because that question doesn't have a vendor, and institutions tend to act only on the questions someone is selling them an answer to.

When to patch and when to rebuild

Patching isn't a sin. Most of the time it's correct. Rebuilding is expensive, slow, politically costly and frequently unnecessary. Any leader who proposes a system redesign in response to every architectural mismatch will exhaust the institution's capacity for change long before the changes that matter arrive.

So, the question isn't whether to patch or rebuild. The question is, how do you tell which option you're facing. And here the debt literature is genuinely useful, because it gives you diagnostic signatures rather than instincts.

Here are some answers:

• Patch when the mismatch is local and the assumption still holds elsewhere. A process that breaks in one office under one condition is a process problem. Fix the process.
• Suspect architecture when the same class of problem keeps recurring under different names. This is the strongest signal in the literature and the one leaders most consistently miss. If your institution has addressed enrollment decline with new marketing, then new programs, then new partnerships, then new tuition discounting but the decline persists, the pattern isn't four initiatives that failed. The pattern is that all four operated at a layer shallower than the problem. Recurrence under new names is the signature of architectural mismatch.
• Suspect architecture when the workaround has become the process. Every institution has practices that exist purely to compensate for a system that no longer works. Examples include the shadow spreadsheet, the staff member everyone calls because the official channel doesn't work and the meeting aimed at undoing the effects of another meeting. These aren't inefficiencies. They're the institution's own diagnosis of its inherited architecture, already performed, already documented in behavior, waiting for someone to read it.
• Suspect architecture when the fix requires a new prohibition. When the response to a recurring problem is a new rule forbidding the symptom, the underlying structure hasn't been addressed. The rule is a patch on a patch. This pattern isn't unique to higher education -- the frontier AI labs are doing exactly that, responding to emergent model behaviors by adding hard-coded prohibitions to system prompts rather than examining the training processes that produced them. It won't work for them either.

Before an institution can decide how AI changes its workforce, it has to identify what it currently assumes.

• Rebuild when the assumption itself has been invalidated, not merely stressed. This is the AI case. The assumption that a position is the atomic unit of institutional work isn't simply under stress; it's been invalidated. That's because the work inside various positions is being redistributed by a capability that doesn't respect position boundaries. No amount of patching at the position level will resolve a mismatch that exists at the level of what a position is.

What this asks of leadership

The practical implication is uncomfortable, which is usually a sign it's the right one.

Before an institution can decide how AI changes its workforce, it has to identify what is currently assumes. Most institutions have never done this analysis, because inherited architecture is invisible precisely to the people who have operated inside it the longest. The assumptions don't appear in the strategic plan. They appear in the org chart, the position control system, the workflow that everyone works around and the committee that exists because of a decision made in 1994 that nobody remembers making.

That surfacing work isn't an HR exercise, though HR is where its absence becomes most expensive. It's architectural work, and it belongs to leadership, because it's the only layer with authority over the assumptions themselves.

The question isn't what AI tools should we buy? The questions to ask are these: What is this institution built on, does it still hold and if it doesn't, should we patch or should we rebuild? They need to be asked deliberately, with a diagnosis, rather than discovered 18 months into an implementation that's failing for reasons nobody can name.

My 1880 house is a good house. I have no intention of tearing it down. But I stopped pretending some years ago that the wiring was a small project.

Roberty Clougherty

Robert Clougherty, the AI Strategy and Innovation lead at the Alliance for Innovation & Transformation (AFIT), a nonprofit association that empowers higher education institutions to transform their organizations.