His subject is video review in football, the system that freezes a frame and draws lines across the pitch to rule a goal offside by the width of an armpit. His observation is that the technology did not end the arguments about offside; it shrank them and moved them. Officials used to argue about whether a player was level with the defender. Now they argue about which frame shows the instant the ball left the passer’s boot, and whether a shoulder counts as a scoring part of the body. Technology, he writes, seems to make errors fractal: the same mistakes, reproduced at ever-finer scales. The judgment is still there. It has been relocated somewhere harder to see, and the people who relocated it have declared victory.

Michael’s own conclusion is that the answer is not more precision but what he calls system stewardship: accept that judgment cannot be engineered out, decide in advance which calls belong to people, and build rules that can absorb the residual disagreement rather than pretending it is gone. He names artificial intelligence as the most urgent place to apply that thinking. I would narrow it further. The most urgent place is the clinical AI your health system bought last year.

Here is the fractal pattern, as I now see it, in three tools most large systems already run.

Ambient documentation removed the typing and kept the judgment about what the visit was about, but moved that judgment from the act of composing the note to the act of reviewing a draft. Those are different mental tasks. Composing forces you to decide what mattered because you cannot write the note without deciding. Reviewing asks only whether anything on the screen looks wrong, and a draft that has quietly left out the finding you would have led with does not look wrong. It looks finished. The decision did not disappear; it moved to a moment where it is easy to skip.

Predictive alerts did the same thing to the decision to treat, converting it into a smaller, later, more frequent decision about whether this particular alert deserves attention. And imaging AI turned the radiologist’s job from generating a read into second-guessing one that arrived first.

What happens to human performance on the relocated decision? The radiology evidence is uncomfortable. When German researchers gave 27 radiologists mammograms accompanied by purported AI suggestions, a dozen of which were deliberately wrong, even the most experienced readers, who were right 82% of the time with good suggestions, were right only 45% of the time with bad ones. Less experienced readers fell to 20%. Twenty years of practice were worth roughly 25 points of resistance, and the expert still lost the coin flip.

And the relocated skill does not hold its value. A study published last year followed endoscopists at four Polish centers after AI polyp detection was switched on, looking only at the colonoscopies they still did without it. Their unassisted adenoma detection rate dropped from 28% to 22% within three months. The reps went to the machine, and the skill followed the reps.

Subscribe now

This is where Michael’s stewardship idea becomes a concrete agenda for a governance committee, and where I think most health systems are doing the fractal thing without noticing. They are measuring the machine with increasing precision and the relocated human decision not at all. The number everyone can quote is the override rate, which counts how often a clinician pushed back against the tool and is silent on whether the pushback was correct, how much time the clinician had, and how the patient fared. That is attendance, not oversight. Stewardship would mean sampling the overrides and the acceptances, tracing the outcomes, checking unassisted performance periodically, and writing down, before deployment, where a model is permitted to participate and where a human keeps the call by policy rather than by accident.

Football, to its credit, wrote that list before its first video review. For eight years it held at four kinds of decision. This summer it grew, which Michael would recognize as the pattern operating on the rulebook itself.

I have laid out the full argument on my site this week, with the evidence and a proposed replacement for the override-rate dashboard.

Also worth your time

Michael Hallsworth, “Clear and obvious errors,” The Judgment Gap, June 2026. Start here. The football is the hook; the argument about relocated judgment and system stewardship is the payload, and it transfers to any domain where a precise tool meets an imprecise task.

Goddard, Roudsari, and Wyatt, “Automation bias: a systematic review of frequency, effect mediators, and mitigators,” JAMIA 2012. The foundational review of why people over-trust automated advice. One finding worth pinning above every decision-support build desk: a recommendation biases the user more than the same information presented without one. Open access.

Jassim and colleagues, “Performance of artificial intelligence in breast cancer screening programmes: a systematic review,” BMJ Open 2025. Thirty-one studies and more than two million examinations. The sentence to read twice says triage works “when thresholds were conservatively calibrated.” Somebody picks the threshold. That somebody is the frame. Open access.

That 2023 piece admired a group at the Medical University of South Carolina who wanted emergency physicians to send at-risk patients home with naloxone. Rather than throwing a modal window across the screen, they watched for signals already in the chart (e.g., a triage reason of drug overdose, or the words naloxone or Narcan typed into the free-text complaint) and then quietly dropped a suggestion into the history section of the note the physician was already writing. Take-home kits went from 25.7% of eligible encounters to about half. Three-quarters of the physicians surveyed rated the thing good or excellent, which is not a sentence anyone gets to write about a best-practice alert.

I credited the format and the timing. Interrupt less, arrive at a receptive moment. That reading is not wrong, and it is incomplete in a way that took me three years and somebody else’s data to see.

Look at what that intervention actually asked a physician to do. Hand the patient a box. That is the entire request, and it finishes inside the visit. No result comes back two days later. Nobody has to decide whether a borderline number means anything. No parent needs a phone call. No threshold for referral has to be weighed by a person who has eleven other charts open. The work terminates.

Without meaning to, I had chosen the friendliest possible case for non-interruptive decision support: one where the task closes on contact. Design the prompt well and the loop shuts by itself.

Most clinical work is not like that, and I think this is where a decade of decision support has quietly gone to die.

What the new data show

A study out in Pediatrics this summer looked at what children actually receive after a first diagnosis of high blood pressure. The guideline is not ambiguous: urinalysis and a chemistry panel for everyone, at the front. What happened instead is that the echocardiogram finished first, at 21.1%, ahead of the urinalysis at 19.9% and the chemistry panel at 17.7%.

Consider what those two tests cost to obtain. One of them needs a cup. The other needs somebody to arrange a referral, fight an insurer, find an appointment, staff a sonographer, and book cardiology reading time, and the guideline does not call for it until a medication is on the table, six to twelve months down the road. Every implementation framework I have ever worked from says the easy test wins. It lost.

Sort the tests a different way, though — by how much unfinished business each result creates — and the ranking stops being mysterious. An echocardiogram resolves. It reassures the family, or it moves the child to cardiology and the next several decisions belong to somebody else. A urinalysis showing trace protein resolves nothing at all. It produces a repeat test, a judgment call, a fuzzy threshold for phoning nephrology, and a family conversation that has to be had while the meaning of the number is still unsettled. There is nowhere for it to go. It becomes something the physician is now carrying.

Nobody skips a urinalysis because an abnormal result would make work. The mechanism is quieter than that and almost certainly below conscious awareness. Some orders close a visit. Others open a loop, and a loop can always be pushed to the next visit, defensibly, one child at a time, until several hundred thousand small reasonable choices add up to a number like 19.9%.

So the governance question I would put to any decision support proposal is no longer whether the alert fires cleanly. It is: for every result this thing can produce, what happens next, and whose name is on it? “The ordering clinician decides” is an acceptable answer. It is a terrible default, and it is what you get when the meeting ends five minutes early.

The longer version, including why your order sets may still be running a guideline that was retired in 2017, is on my site.

Also worth your time

• Menon and colleagues, in BMJ Open, on why test results get missed — they name three situations where results reliably fall through: tests ordered by trainees, alerts handed to a covering clinician, and patients with nobody assigned as their physician. Every one of them is an ownership vacuum rather than a technology failure. Free to read.

• Meyer and colleagues, in the Journal of General Internal Medicine, on catching delayed follow-up automatically — instead of nagging a clinician at the moment of ordering, they built an algorithm that goes looking for abnormal thyroid results nobody acted on within sixty days. This is software doing the work rather than assigning it, which is the whole argument above, implemented. Free to read.

• Nugent and Kaelber, in Pediatrics — the study behind this issue. Two pages, open access, and worth reading in the original if only for how flat the authors keep their voices.

This goes out every three weeks or so (sometimes “so” is doing a lot of work). If it stops being useful, the unsubscribe link is right there, and I will never know.

Over a decade ago, Robert Campbell outlined the Five “Rights” of Clinical Decision Support: the right information, to the right person, in the right intervention format, through the right channel, at the right time in the workflow. For purposes of this discussion, let me focus on the third right: the right intervention format. There are clearly times when a clinician should be stopped in their tracks. For example, if a doctor signs an order for a medicine that is likely to cause a serious allergic reaction or if a nurse is about to administer a drug to the wrong patient, a sudden stop is indicated. But often, it’s the case that a gentle push in the right direction is all that’s needed for a clinical expert.

In this month’s Journal of the American Medical Informatics Association, researchers from the Medical University of South Carolina (MUSC) wrote of their research to use non-interruptive CDS to learn if they could increase the rate at which emergency medicine physicians prescribe take-home naloxone for patients who may have had a drug overdose (OD). They specifically attempted to suggest the order at a point in the clinical workflow when the doctor might most easily be receptive to it: during documentation of the emergency department (ED) encounter.

For background, naloxone is a medicine that can reverse the effects of opioid overdose. It’s a lifesaver, but for it to work, it must be administered soon after the patient starts overdosing. Many first responders now routinely carry naloxone because it can safely be given when OD is suspected (side effects are minimal, so even if there was no opioid overdose, no serious harm will come from the drug). The goal of this research was to give a patient who might be at risk of overdosing in the future the naloxone that they or their friends or family might administer if it was needed.

Instead of popping up a modal window, the MUSC researchers embedded a reminder into the progress note that the ED physician would likely use to document the visit. This may seem simple, but it’s not obvious that overdose is the likely diagnosis when a patient is first presenting to the emergency department. Recall that a typical workflow in the ED is to start documenting a note as soon as the physician starts collecting the patient’s history. Hence, the researchers looked for clues in the chart. They settled on a field in their EHR that collects both a discrete reason for the visit and a free-text description of the chief complaint. If the reason for the visit as documented by the triage nurse was “drug overdose” or if the words “naloxone” or “Narcan” (brand name for naloxone) appeared in the free-text box, the clinical decision support reminder kicked in.

Once the encounter was identified as a possible OD, in the history of present illness (HPI) section of the routine emergency medicine note, text appeared suggesting that the physician might want to use a short documentation template (“dot phrase”) to collect the necessary information regarding the event and also might want to send the patient home with the take-home naloxone. The key from the clinician's perspective is that this is a good point in the workflow to make these suggestions, as the doctor is likely receptive to these prompts at this time.

What were the results of the CDS prompt in the documentation template? Before the intervention, naloxone kits were distributed 25.7% of the time; afterward, the rate almost doubled to 50%. Obviously, the goal is to give take-home naloxone kits every time, but this was a significant improvement. Further, they found that physicians actually appreciated the advice. The majority were net promoters, and 74% scored the CDS intervention in the excellent or good range on a survey.

Non-interruptive clinical decision support needs to become the norm, and as our EHRs get more sophisticated, I think we’ll see fewer pop-up alerts that annoy physicians. CDS that reminds physicians to do the “right thing” at the right time will be embraced by clinical staff, leading to a decrease in cranky comments and a rise in the awareness of the benefits that the EHR can bring to patients and those who care for them.