Measuring Slow Work by a Fast Clock

Measuring Slow Work by a Fast Clock
Clock on the Tree's Quiet Growth

Case of the Week, Min Wu, PhD · ai-public-health.com

In the months after my textbook came out in April 2026, I caught myself checking its page visits the way some people check stock prices. There was an early bump, and then the numbers came down. Before long I was reading every dip as a verdict on the book, and on whether careful academic work in a fast-moving field was worth doing at all.

Then I recognized the habit. It is not a book problem. It is what happens whenever slow work gets judged by the fastest number available, and public health has been making this mistake, and learning from it, for a long time.

The Number That Is Easiest to See

We tend to watch whichever number updates most often. A retirement account checked every morning feels like a daily report card, even though it was built to be judged over decades. A student refreshing the grade portal after every quiz feels the semester rise and fall by the hour, though what matters is what they can still do a year later.

The danger goes beyond anxiety. A fast number can quietly become the goal. The idea is often called Goodhart's law: once a measure becomes a target, it stops being a good measure.

Public Health Lives on Slow Clocks

Public health may be the field most exposed to this trap, because its successes are slow and often invisible. Practitioners have a familiar saying: when prevention works, nothing happens.

Measles is the clearest case. After the vaccine arrived in the 1960s, the disease became rare enough in the United States that many younger parents have never seen a case. The vaccine's success removed the very evidence that justified it, and some families came to see the shot as the risk rather than the disease. Judged by what people could see that week, the vaccine looked unnecessary. Judged on the right clock, it was one of the field's great achievements.

Tobacco control ran on an even longer clock. Adult smoking in the United States began falling in the mid-1960s (PMID: 10577492), but lung cancer deaths among men did not start declining until around 1990, roughly a quarter century later (doi:10.3322/caac.21590). Anyone who judged anti-smoking work by the next year's cancer statistics would have called it a failure.

Even a well-chosen fast number pulls behavior toward itself. When England required emergency departments to admit, transfer or discharge patients within four hours, economists later found that departures bunched up in the final minutes before the deadline (doi:10.1162/rest_a_01044). The clock was steering decisions. In that case the target also appears to have saved lives, which is the honest complication: a fast number is not always wrong, but it always needs watching for what it does to the work.

My Own Slow Clock

The deeper version of this lesson comes from my research. Much of my work is building conceptual frameworks: structured ways of thinking through how a public health task actually gets done, step by step, and where AI might help.

I chose this kind of work for a reason. The AI systems now arriving do more than answer questions. They take actions inside real workflows, such as scheduling, flagging, routing, and following up. Before anyone can safely hand them those tasks, someone has to map how the work is actually done in a given domain. That mapping sits in the middle, between grand theory and a single experiment. Sociologists call it middle-range theory, and the AI era needs a great deal of it.

A new framework starts as an imperfect idea, a partial solution to a real problem. Before it can be trusted, it needs simulation and then empirical testing in real places. Until then, it looks risky to the people whose job is to manage risk: journal editors, reviewers, funders. Their caution is reasonable. An untested idea should face hard questions. A smaller, incremental study with experimental results fits their evaluation systems better, passes desk review sooner, and reaches peer review early.

So the real choice is between clocks. The incremental paper runs on a fast clock and earns early signals of acceptance. The framework runs on a slow one. I publish it as a preprint so the idea can be read and cited. My most recent example is a middle-range framework for dementia care that checks its mechanism against history and sets out a policy agenda for delivering it with AI (doi:10.2139/ssrn.7610360). I write out the simulation model as an open specification, posted publicly for anyone to copy and test against the data in their own city. Then I wait. Choosing the slow clock means accepting that the early signals will be quiet, and not mistaking that quiet for a verdict on the idea.

Over the years I have watched which signals turn out to matter. I have seen research video projects that started on small budgets with simple functions: store the videos and let people annotate them. They kept growing, because people used them and told the team what to fix. I have also seen well-funded projects with complex features that looked strong at launch and then faded, because few people used them and no feedback came back. The grant and the feature list were fast signals. Use and feedback were slow ones, and they were the ones that decided which projects lasted.

I had already written about a version of this mistake. In "Homework Is the Rehearsal," two students earned the same grade for the same AI-assisted assignment, though only one could explain her reasoning. My rubric measured the output and missed the process. Traffic numbers and early acceptance signals can do the same thing to a researcher.

Calm Is Calibration, Not Resignation

Calming down did not mean caring less. It meant matching the measure to the clock, and that takes three steps.

  1. Name the clock. Ask how long your work takes to show its effect. A vaccine program runs on years. A framework runs on its validation cycle. A course runs on what students carry into practice. Any number that updates faster than that is noise, not news.
  2. Find one slow signal. Most fast numbers have a slower, truer counterpart. For a student, it is whether they can explain their reasoning, not the quiz score. For prevention, it is whether people have quit smoking, not this year's cancer statistics. For a tool, it is whether people keep using it and telling you what to fix, not the size of the grant behind it. For a framework, it is whether a simulation holds up, not the first editor's reaction.
  3. Check at the pace of the clock. Look once a season, not once a morning. In between, put your attention back on the process, which is the only part you control.

What I Am Still Unsure About

A slow clock carries its own risk. Work that is simply not useful also looks quiet, so patience can become a way of never hearing bad news.

I feel this most with the simulation specifications. They are waiting for someone, somewhere, to take one up and test it in their own city. I do not know when that will happen. It could be one year, two years, or never. A slow clock has no guarantee that it ever strikes.

The best answer I have is to keep the work open, keep it findable, and take silence on the slow signal seriously. If a framework fails its tests, or enough time passes on the right clock and no one finds it worth testing, that is a real verdict and deserves an honest look. A dip in the fast number is not that verdict.

Whatever you are building, whether a course, a prevention program, a research agenda, or a career, ask which clock it runs on before you decide what to watch.

The Clock Makes the Measure


What is the fast number in your work, and what slow signal would tell you more? I would love to hear. Reply or leave a comment.

Want to read the earlier essays in one place? They’re collected in AI and Public Health, Volume 1, in ebook and paperback.

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