Medicine has developed elaborate ways of learning new things.
There are clinical trials, systematic reviews, regulatory pathways, professional societies, guideline committees, continuing education requirements, and entire industries dedicated to getting new discoveries into clinical practice.
We have organizations whose job is to produce evidence, others whose job is to evaluate it, and still others whose job is to persuade clinicians to act on it.
But what happens when the evidence changes in the other direction?
What happens when a treatment once considered effective turns out to be useless? When a procedure’s benefits prove too small to justify its risks? When a diagnostic threshold, screening recommendation, or standard intervention no longer deserves the authority it has accumulated?
We tend to imagine that medicine updates itself. The evidence changes, the guidelines change, and eventually the practice changes.
Sometimes it does.
But a change in what medicine knows does not automatically produce a change in what medicine does. By the time a clinical practice becomes routine, it may no longer depend on the scientific proposition that originally justified it.
It has acquired a life of its own.
And unlike scientific knowledge, institutional routines do not become obsolete simply because someone proves them wrong.
The evidence has changed. So what?
Consider bed rest during pregnancy.
For decades, restricting physical activity was recommended to women at risk of preterm delivery. The reasoning was intuitively appealing: if physical activity might place stress on a pregnancy, reducing activity might reduce the risk.
It also sounded like a comparatively conservative intervention. Doing less, after all, hardly sounds like an aggressive medical treatment.
But intuition isn’t evidence.
Research failed to establish that routine bed rest prevents preterm birth. Meanwhile, prolonged inactivity carries risks, including blood clots, physical deconditioning, and psychological and economic consequences.
The American College of Obstetricians and Gynecologists has explicitly recommended against routinely prescribing activity restriction to prevent preterm birth.
The point isn’t that every obstetrician continued prescribing bed rest after the recommendation changed, or that all activity restrictions are inappropriate. Individual clinical circumstances matter.
The point is that a practice could become familiar, professionally legitimate, and embedded in expectations well before its benefits had been adequately established.
Correcting the evidence is only one part of withdrawing that legitimacy.
Or consider arthroscopic surgery for degenerative knee disease.
A 2017 BMJ clinical practice guideline strongly recommended against arthroscopy in nearly all patients with degenerative knee disease, following evidence that its benefits were generally too limited to justify the burdens and risks.
A procedure that had become familiar in orthopedic practice was now being challenged by the very machinery of evidence-based medicine that had helped establish clinical authority in the first place.
But a recommendation isn’t an instruction that automatically executes across every clinic, hospital, insurance arrangement, operating room, and physician-patient conversation.
It has to travel.
And everything it encounters along the way has its own history.
A practice is never just a practice
Suppose a hospital has been performing a particular procedure for twenty years.
During that time, people have been trained to perform it. Equipment has been purchased. Billing codes have been established. Departments have developed workflows around it. Patients have been taught to expect it. Referrals have become routine. Quality measures may have incorporated related assumptions.
Entire professional identities can become entangled with a particular way of practicing medicine.
Now imagine a new study calls the procedure’s value into question.
What, exactly, has changed?
View source
%% title: What a study updates and what a practice is made of
%% caption: A new study weakens the justification for a procedure. The chain beneath it — training, equipment, billing codes, order sets, patient expectations, quality measures — is untouched, and those are what a patient actually encounters.
flowchart TD
classDef actor stroke-width:1.6px;
classDef system stroke-width:1.2px;
classDef gate stroke-width:1.6px;
classDef repair stroke-width:2.4px,font-weight:bold;
S(["New evidence weakens<br/>the justification"]):::repair --> J{"Does the practice stop?"}:::gate
J -- "the study reaches this" --> E["The reason written<br/>in the guideline"]:::system
J -- "these would have to change too" --> P1["People trained to perform it"]:::actor
P1 --> P2["Equipment already purchased"]:::actor
P2 --> P3["Billing codes and coverage"]:::actor
P3 --> P4["Order sets and referral pathways"]:::actor
P4 --> P5["Patients taught to expect it"]:::actor
P5 --> P6["Quality measures built on it"]:::actor
P6 --> P7(["What the patient actually meets"]):::actor
The evidentiary basis for the procedure may have weakened. But the surgeon’s training hasn’t disappeared. The hospital’s equipment hasn’t vanished. The reimbursement arrangements haven’t spontaneously reorganized themselves. Patients haven’t necessarily changed their expectations.
There is now a discrepancy between the justification for a practice and the institutional arrangements that sustain it.
And the latter may be considerably more durable than the former.
This is why describing the persistence of low-value care as a problem of physician education is often insufficient.
Sometimes clinicians genuinely disagree about the evidence. Sometimes the evidence is uncertain, applies poorly to the patient in front of them, or supports more than one reasonable course of action. Those are legitimate clinical problems.
But there are also cases where people understand that a practice is no longer well supported and continue operating within arrangements that make discontinuing it difficult.
They may not have the authority to change the order set, the quality measure, the referral pathway, or the payment policy. They may lack a viable replacement workflow. They may be exposed to professional or legal consequences if they depart from established routines, even when they can defend that departure scientifically.
Knowing better is not the same as being able to act differently.
The distinction seems obvious when stated plainly. Yet much of our language about evidence-based medicine treats knowledge and action as if they were separated primarily by a failure of communication.
As though the problem is that the right information hasn’t reached the right person.
Sometimes the information has arrived.
The institution simply has nowhere to put it.
Institutions remember differently than people do
We tend to think of institutional memory as a record of what an organization knows.
But institutions also remember through the arrangements they reproduce.
An electronic health record remembers through its defaults.
A hospital remembers through its purchasing decisions.
An insurer remembers through its coverage policies.
A training program remembers through what it teaches residents to recognize as normal.
A profession remembers through its standards, routines, and expectations of competence.
None of these forms of memory requires anyone to consciously endorse the original rationale for a practice.
The reasons can be forgotten while the consequences remain.
In fact, a practice may become more difficult to question as its historical justification becomes less visible.
Early adopters may remember the uncertainty, controversy, and competing alternatives that surrounded its introduction. Later generations inherit the resulting arrangement as the ordinary way things are done.
What was once a decision becomes a default.
What was once a hypothesis becomes a protocol.
What was once one possible response to a problem becomes the definition of competent practice.
Eventually, abandoning the practice may feel more radical than adopting it ever did.
This is a peculiar form of institutional forgetting: an organization can forget why it does something without forgetting how to keep doing it.
The operational knowledge survives the epistemic justification.
And because institutions often measure successful performance relative to existing workflows, the people most competent at navigating those workflows can make the arrangement appear more successful than it is.
The institution sees continuity.
It may not see the cost of maintaining it.
The people who keep the old system working
There’s an additional complication.
Institutional practices don’t persist only because people resist change. They also persist because people become capable of accommodating them.
A clinician learns which order-set options to ignore.
A pharmacist catches the outdated instruction before it reaches a patient.
A nurse learns which standing protocols need clarification.
A department creates an informal exception to a rule that no longer makes sense.
These adaptations can be intelligent, necessary, and protective. They are often how patients avoid the consequences of organizational shortcomings.
But they can also reduce the visibility of those shortcomings.
If the people working inside an institution consistently repair its defects, the institution may never experience those defects as sufficiently urgent problems.
The workaround becomes part of the operating model.
This creates a difficult asymmetry.
People can become more competent at functioning under an arrangement without becoming any more powerful to change the arrangement itself.
Their competence preserves the system’s apparent functionality while their authority remains limited.
In healthcare, where professional responsibility for patient outcomes is often more immediate than control over institutional infrastructure, this is especially consequential.
You cannot simply stop compensating for a defective system when a patient might be harmed.
So the system continues functioning partly because people absorb costs that its formal performance measures may never capture.
The result is not necessarily a conspiracy, a failure of professional ethics, or an unwillingness to accept evidence.
It may be an institution whose ability to sustain itself has become partly dependent on the very adaptations that conceal its need to change.
Deimplementation isn’t the missing discovery
None of this is an argument that researchers have failed to notice the problem.
There is an established field devoted to deimplementation: reducing or discontinuing medical practices that provide little benefit, have been superseded, or may cause harm.
Researchers have investigated medical reversals, inappropriate prescribing, unnecessary testing, and the difficulty of withdrawing established interventions.
Choosing Wisely and similar initiatives have attempted to make low-value care more visible. Implementation scientists have developed frameworks for identifying barriers, changing professional behavior, evaluating interventions, and sustaining reductions in unnecessary care.
This work has produced meaningful results.
A 2025 systematic review and meta-analysis of 140 randomized trials in primary care found that deimplementation strategies reduced the relative risk of targeted low-value care by roughly 10–35%, depending on the intervention.
That’s important evidence that change is possible.
It also demonstrates that identifying a low-value practice and issuing a recommendation against it are not sufficient.
Changing practice takes additional work.
But I think the problem extends beyond whether we have effective interventions for reducing individual low-value services.
It raises a more general question about how institutions are organized to revise themselves.
We have devoted considerable attention to the movement from evidence to implementation. What remains less settled is how responsibility, authority, and the cost of revision should be distributed once a practice becomes embedded in the machinery of care.
Who is responsible for detecting that an established practice may no longer be justified?
Who can compel reconsideration?
Who has authority to change it?
Who bears the costs of transitioning away?
Who verifies that the change actually occurred?
And who is accountable when everyone agrees that something should stop, but no single actor has both the authority and incentive to stop it?
These aren’t merely questions about clinician behavior.
They’re questions about institutional design.
Innovation has sponsors. Retirement has dependencies.
Introducing a new medical technology can mobilize an enormous amount of organized effort.
A manufacturer has incentives to secure approval, demonstrate clinical value, obtain reimbursement, persuade clinicians, and establish market access.
Hospitals may establish innovation offices. Investors fund commercialization. Professional organizations evaluate emerging evidence. Vendors supply implementation resources.
None of this guarantees that adoption will be appropriate or successful. Nor does it mean new practices spread effortlessly.
But innovation can have identifiable sponsors with a direct interest in overcoming obstacles.
The retirement of an established practice often has a different political economy.
View source
%% title: Adoption has sponsors, retirement has costs
%% caption: A new procedure arrives with a chain of actors pulling it into practice, each with an interest in overcoming obstacles. Removing it has diffuse beneficiaries and concentrated costs, and no actor owns the transition, so the default outcome is that the practice continues.
flowchart TD
classDef actor stroke-width:1.6px;
classDef system stroke-width:1.2px;
classDef gate stroke-width:1.6px;
classDef failure stroke-width:2.2px,stroke-dasharray:6 3;
classDef repair stroke-width:2.4px,font-weight:bold;
subgraph Adoption ["Adoption: someone is pulling"]
direction TB
A1["Manufacturer wants approval"]:::actor --> A2["Reimbursement secured"]:::actor
A2 --> A3["Clinicians persuaded"]:::actor
A3 --> A4(["Practice becomes routine"]):::repair
end
A4 --> R1["Evidence weakens"]:::system
R1 --> R2["Guideline recommends against"]:::system
R2 --> R3{"Who absorbs the transition?"}:::gate
R3 --> R4["Workflow rewritten"]:::failure
R4 --> R5["Order set rebuilt"]:::failure
R5 --> R6["Revenue lost"]:::failure
R6 --> R7["Mistake admitted"]:::failure
R7 --> R8(["Practice continues"]):::failure
The benefits of discontinuation may be diffuse: fewer unnecessary interventions, less patient burden, lower spending, fewer adverse events.
The costs of changing the system may be immediate and concentrated.
A department must revise workflows. Someone must rebuild an order set. A clinician must explain why a treatment previously recommended is no longer appropriate. An organization may lose revenue, confront professional disagreement, or acknowledge that its prior practices were mistaken.
No single actor necessarily captures enough of the benefit to justify absorbing all those costs.
And some actors may benefit from the practice continuing.
This does not make deimplementation universally harder than implementation. Some interventions disappear rapidly after compelling safety evidence, changes in coverage, or the arrival of superior alternatives. Others never become deeply embedded in the first place.
But it means that the organizational machinery built to introduce a practice cannot simply be assumed to work equally well in reverse.
Removing an established practice is not the inverse of adopting it.
By the time something has become routine, the conditions that allowed its introduction may no longer exist.
Its original advocates may have moved on. Its purpose may have changed. Its costs may have been redistributed. Its justifications may have been replaced by habit, convenience, or dependencies that nobody deliberately chose.
To undo it, an institution may have to confront not only the original decision but everything that has accumulated around it since.
What happens when AI learns from institutions that haven’t learned to forget?
This problem becomes more interesting as healthcare increasingly uses AI to synthesize medical knowledge, support decisions, and automate clinical workflows.
We often imagine that these systems will help medicine update itself faster.
A model can retrieve recent evidence. A knowledge system can identify new recommendations. Clinical software can surface relevant guidance at the point of care.
Those are potentially valuable capabilities.
But an AI system can be current with the literature while the institution using it remains organized around older assumptions.
What happens when the model recommends something the local workflow doesn’t support?
What happens when a practice is contradicted by emerging evidence but remains embedded in institutional guidelines, templates, or reimbursement policies?
What happens when sources disagree because they were updated at different times, for different populations, under different evidentiary standards?
And what happens when the system is trained on historical records containing practices that were once considered appropriate but are now discouraged?
The problem isn’t simply whether the AI has access to the latest information.
It is whether the system can distinguish among what has been observed, what is currently believed, what is professionally recommended, what is institutionally authorized, and what clinicians actually do.
Those are different kinds of knowledge.
They change at different speeds.
And they can contradict one another without any individual component being obviously broken.
There is also a danger in treating AI-generated summaries as a way to compress away those contradictions.
A clean answer can obscure the history of disagreement, the uncertainty surrounding a recommendation, or the institutional decisions that made one interpretation authoritative.
A system that makes knowledge easier to retrieve is not necessarily a system that makes knowledge easier to revise.
It could even make outdated assumptions easier to reproduce if the underlying institutions remain unchanged.
The deeper opportunity for clinical AI may therefore be less about generating answers than about making the relationship between evidence, authority, and practice more inspectable.
What changed?
Why?
Who accepted the change?
Where has it been implemented?
Where does the older practice persist?
And under what conditions should an established recommendation be reconsidered?
These are not questions a language model can resolve by itself. They require evidence, clinical judgment, institutional authority, and accountability.
But technology could make the discrepancies harder to ignore.
That would be a meaningful advance.
Forgetting should be a capability
Medicine should not indiscriminately abandon old practices merely because they are old.
Nor should every new study prompt a wholesale revision of clinical care. Evidence changes unevenly. Findings are contested, populations differ, and apparent reversals sometimes turn out to be narrower than their headlines suggest.
There are good reasons for caution.
The objective isn’t to make institutions maximally responsive to every new claim.
It’s to make them capable of distinguishing justified continuity from inertia.
A healthcare system that can learn but cannot reliably reconsider its established commitments is not fully evidence-based. It is evidence-accumulating.
It can produce new knowledge while retaining arrangements that no longer deserve their authority.
And that distinction has consequences.
It means scientific progress can coexist with institutional stagnation.
It means better information can coexist with unchanged practice.
It means people can recognize that something ought to change while remaining dependent on the conditions that prevent them from changing it.
The challenge is not simply to accelerate the dissemination of new discoveries. It is to build institutions in which the justification for an established practice remains open to examination, and where reconsidering that justification can lead to actual revision.
That requires more than an updated guideline.
It requires preserving the reasons for decisions, identifying the dependencies built around them, assigning authority to revisit them, and making the costs of change visible rather than leaving them to be absorbed by whoever happens to be closest to the problem.
In other words, institutions need more than memory.
They need the capacity to revise what their memory has made possible.
Medicine has become extraordinarily sophisticated at producing new knowledge.
The harder test of a learning healthcare system may be whether it can relinquish an old practice after the people, technologies, incentives, and routines built around it have made that practice seem indispensable.
The measure of an institution’s intelligence isn’t only what it can learn. It’s what it can stop doing once it has learned better.
References and further reading
- American College of Obstetricians and Gynecologists. Physical Activity and Exercise During Pregnancy and the Postpartum Period. 2020.
- Siemieniuk RAC, et al. Arthroscopic surgery for degenerative knee arthritis and meniscal tears: a clinical practice guideline. BMJ. 2017.
- De-implementing wisely: developing the evidence base to reduce low-value care. BMJ Quality & Safety. 2020.
- The Impact of Choosing Wisely Interventions on Low-Value Medical Services: A Systematic Review. Milbank Quarterly. 2021.
- What Are Effective Strategies to Reduce Low-Value Care? An Analysis of 121 Randomized Deimplementation Studies. 2023.
- Deimplementation of ineffective and harmful medical practices: a data-driven commentary. American Journal of Epidemiology. 2025.
- Effectiveness of different de-implementation strategies in primary care: systematic review and meta-analysis. BMJ Medicine. 2025.