New Bottle, Old Wine

Why new technology running old logic is the most dangerous configuration we have

Min Wu, PhD  ·  ai-public-health.com

A student I knew did everything the system asked of her. She optimized her GPA through middle school and high school. She stacked extracurriculars the way the rubric rewarded. She practiced standardized tests until the patterns were muscle memory. She wrote essays that admissions readers were trained to recognize. The logic worked. She was admitted to a top university.

Then college began, and the rubric ended.

She kept asking the question that had carried her for twelve years: what is the assignment, and how do I optimize for it? But there was no assignment in the old sense. There was a question she was supposed to choose, a problem she was supposed to find interesting, a direction she was supposed to want. She had been trained to answer; she had not been trained to ask. She did not know that the logic she had mastered had a scope, because no one had ever told her it had a scope.

The logic that had been a ladder became a cage.

It took me longer than I would like to admit recognizing what I was watching. The logic was not wrong. It had served her well at the level it was designed for. The failure was not in the logic. The failure was in the silence about its boundaries.

The same shape, in other rooms

What happened to her is not unique to admissions, and it is not unique to students. It is a pattern that appears anywhere, a piece of logic that worked at one level gets carried, unexamined, into a level it was never designed for. Once you see the shape, you start seeing it everywhere.

In public health, this pattern is now being industrialized at speed. AI pipelines, synthetic populations, and predictive infrastructures inherit the case definitions, thresholds, and assumptions of an earlier era and scale them across populations and decisions before anyone has audited whether the underlying logic still belongs there.

New technology running old logic is the most dangerous configuration we have. Not because the technology is bad. Not because the old logic was always wrong. But because the infrastructure now scales unexamined assumptions faster than anyone can audit them.

The pattern public health already knows

Antimicrobial resistance is the example almost every public health practitioner already accepts. Each individual prescribing decision can be defended on its own terms. A clinician sees a patient, weighs the evidence, weighs the cost of acting against the cost of not acting, and prescribes. Patient by patient, visit by visit, the logic is rational and often correct.

The population-level outcome is a public health crisis no individual prescriber caused, and no individual prescriber can solve. Resistance patterns emerge from aggregation. They are properties of the system, not of any single decision.

The clinical logic was not wrong. It was correctly scoped to the clinical encounter. It became a problem when prescribing infrastructure, professional habits, and decades of guidelines scaled it, unexamined, to a level of analysis it was never designed to address.

AMR is the canonical case of valid individual-level logic producing emergent population-level harm. It is also a warning. The same shape, the same level confusion, is now being built into AI pipelines and synthetic-population modeling across surveillance, screening, and resource allocation. The shape is familiar. The speed is not.

I want to be honest about what I do not yet know. I am not sure whether the audit I am about to argue for is feasible on the scale we now operate at. It is possible that some legacy logic will always travel without examination because the cost of auditing it exceeds the cost of letting it run, and that the practical question is not whether to audit everything but where to spend the limited audit attention we have. That is a question I am still working on, not a question I am ready to close.

Why this configuration is asymmetrically dangerous

Old logic running on paper compounds slowly. It is questioned at every transcription. It is corrected at every meeting. The friction of the medium creates room for human judgment to intervene. This is not romanticism about paper. It is a description of how unexamined assumptions used to stay survivable.

Old logic running on modern infrastructure does not have that friction. Three things converge to make the configuration uniquely hazardous, and they are worth naming separately because each one acts on a different part of the system.

Speed: A flawed assumption embedded in a paper workflow propagates at the pace of clerical labor. The same assumption embedded in an AI pipeline propagates at the pace of inference. By the time everyone notices, thousands of decisions have already been made based on it. The assumption has become a fact on the ground before it has become a question in anyone's mind.

Authority: A faxed report carries no special epistemic weight. A reader knows it is the product of human judgment and reads it as such. An algorithmic output arrives wrapped in the authority of objectivity, even when it is industrializing the same human judgment that the fax carried. The authority of the new medium quietly suppresses the questioning that the old logic depended on to stay healthy. The questioning was not a bug in the old system. It was the system's immune response.

Detectability: A low-resource community running outdated logic looks low-resource. The failure mode is visible from the outside. A high-resource community running the same outdated logic, but on top of dashboards and pipelines and modern interfaces, looks state of the art. The infrastructure camouflages failure. From the outside, and often from the inside, nothing looks wrong. The dashboards are responsive. The pipelines complete. The reports generate. The logic underneath them is a decade or more out of date, and there is no surface signal that this is so.

Speed compounds the assumption. Authority suppresses the question. Detectability hides failure. Together they describe why new technology running old logic is not simply faster than the paper version. It is structurally more dangerous.

Observer's Insight

The logic that got the student into college was not wrong. It was out of scope. The failure was not the logic. The failure was the silence about its boundaries.

This is the failure we keep repeating, across domains and across technology generations. Not the logic itself, but the absence of any conversation about where it stops belonging. Paper systems carried that silence at the speed of paper. Modern infrastructure carries it at the speed of inference.

The audit we owe is not of the infrastructure. The infrastructure is doing what infrastructure does. The audit we owe is of the logic the infrastructure is about to industrialize, and of the boundary conditions that the original logic was designed for and that we have not yet checked.

New bottle, old wine. The bottle is not the problem. The problem is that no one has checked the wine for a long time, and we are about to pour it for everyone.

A question for you

Where in your work have you watched a piece of logic that was correct at one level get carried, unexamined, into a level it was never designed for? Did the new infrastructure surface the level confusion, or did it scale it?

I would like to hear from practitioners who have seen this pattern from inside their own systems. How the pattern appears across data, rules, and interface layers, and what it would mean to audit logic before infrastructure rather than after.

In upcoming posts, I'll work through two specific cases of this pattern — diagnostic bias and selective forgetting in AI applications.

As a member of the Springer Nature Author Affiliate Program, I may earn a small commission from purchases made through this affiliated link to my book. Support the author by checking out my textbook, Artificial Intelligence in Public Health: https://tidd.ly/4mH9389