The economist who watched a target break.

In 1975, Charles Goodhart was advising the Bank of England. Policymakers had been using monetary aggregates, the various ways of counting money in the economy, as a reliable signal of where things were heading. Then they started TARGETING those aggregates directly.

The signal fell apart almost immediately. Once the aggregate was the target, the financial system reorganized itself around hitting it, and the number stopped meaning what it used to mean. Goodhart's original 1975 phrasing was dry and technical. The version everyone quotes, "when a measure becomes a target, it ceases to be a good measure," is anthropologist Marilyn Strathern's cleaner 1997 reformulation. Same law. Better sentence.

It is not a story about cheating. It is a story about what pressure does to a number. A measure at rest reflects reality. A measure under load reflects effort. Those are not the same reading, and a dashboard cannot tell them apart.

Watch a measure detach from reality.

A measure gets chosen because, for a while, it tracks the real thing well. That is the whole reason it earns a spot on the dashboard. Then a consequence gets attached. From that moment, the cheapest way to move the number is rarely the same as the expensive, slow way to move reality.

The two lines split. The measure keeps climbing because that is what it is now rewarded for. The thing it was supposed to represent flattens, because moving it for real was always the hard part. The space between them is the Goodhart gap, and it is pure illusion. The dashboard is green. The thing the dashboard was built to watch is not.

Healthcare is built on proxies.

Door-to-balloon time started as a genuine signal. Faster reperfusion, better outcomes in heart attack. Then it became a reported, ranked target, and the clock became something to manage. Start it a little later. Tighten the definition of when "door" counts. The reported time improves faster than the care does.

The emergency department "left without being seen" rate is supposed to measure access. Attach a target and one reliable way to improve it is to room patients quickly, then have them wait IN the room instead of the lobby. The metric clears. The wait does not.

Thirty-day readmissions are the cleanest case. Once the Hospital Readmissions Reduction Program attached real Medicare penalties, the measured readmission rate fell. Observation stays and emergency-department revisits rose over the same window, absorbing patients who once would have counted as readmissions. The penalized number went down. Whether patients did better is a genuinely separate, and harder, question.

Sepsis bundle compliance, HCAHPS patient-experience scores tied to payment, RVU-driven productivity. The list is long and the pattern under it never changes. Each measure was real once. The target is what hollowed it out.

The EHR industrialized it.

Before the EHR, gaming a metric took effort. Someone had to pull charts, tally numbers, argue definitions. The EHR turned metric capture into a checkbox and metric surveillance into a live dashboard. That did not make measurement honest. It made it CHEAP. And cheap measurement breeds more measures.

Two things happen at once. Administrators add metrics, because measuring is now nearly free, and every new metric becomes a new target the moment someone decides it matters. Smart phrases, templated notes, and one-click attestations let the chart say exactly the right thing with almost no connection to what happened in the room.

The clinician is not lying. The clinician is doing precisely what a system of targets trains people to do, which is satisfy the chart. Best Practice Advisory acknowledgment rates sit near 100 percent and mean almost nothing. The advisory was acknowledged. Whether it changed a single decision is not on the dashboard.

What actually works.

You cannot run a health system without measures. The point is not to stop measuring. It is to measure in a way that resists the law instead of feeding it.

Measure closer to the outcome and further from the proxy. Pair every metric with a balancing metric, so a win cannot be faked by quietly losing somewhere else. Readmissions paired with mortality. Wait-time-to-room paired with total time in department. Rotate or retire measures before they fully harden into targets.

And above everything, WATCH THE GAP. If a measure is improving and the clinicians on the floor do not believe care is improving, believe the clinicians. The gap between the number and the felt reality is the most honest figure on the dashboard. It is also the only one nobody is trying to move.

The bed that filled itself.

In 1959, Milton Roemer and Max Shain published a finding that sounded almost too simple to be a law. Compare regions with similar populations and similar health. The region with more hospital beds per person had more hospital-days per person. Not because the people were sicker. Because the beds were there.

The short version traveled faster than the paper. "A built bed is a filled bed." It landed because anyone who has worked a hospital floor already knew it. The bed does not wait politely for the patient who truly needs it. It pulls.

More beds, more days.

Plot bed supply against bed use across regions and the relationship is hard to miss. Illness varies from place to place, but it does not vary nearly as much as utilization does. The supply of capacity tracks the use of capacity more tightly than the burden of disease does.

Why the bed pulls.

The mechanism is not fraud and it is not even really a decision. It is the gravity of a fixed asset.

An empty bed earns nothing. It costs nearly what a full bed costs. So every soft call near the threshold tilts the same way. Observe overnight rather than discharge. Admit the borderline chest pain. Keep the patient one more day to be safe. No single choice is wrong. The aggregate is Roemer's Law.

This is Jevons wearing scrubs. The bed is the efficiency, and the unmet demand for caution and convenience was always there, waiting on the other side of a constraint.

Add the wing, watch it fill.

The most direct test of the law is what happens when capacity expands. A health system opens a new wing to relieve crowding. For a season, occupancy drops, because the same number of patients is now spread across more beds. The spreadsheet says the problem is solved.

Within months, occupancy climbs back toward where it started. The thresholds quietly reset around the new capacity. The crowding returns, now at a larger fixed cost.

It is not only beds.

Roemer studied beds because beds were the capacity you could count in 1959. The law is not about beds. It is about any expensive, fixed clinical resource inside an insured system.

An MRI scanner gets scanned. An operating room gets booked. An ICU tower gets occupied. Add clinic slots and they fill with visits that the old schedule was rationing. Researchers call the broad version supply-sensitive care, and the regional-variation data behind it is some of the most studied evidence in American healthcare. Two counties, similar patients, very different utilization, and the difference tracks the supply.

The planning lesson is uncomfortable. Capacity is not a neutral container you pour existing demand into. It is an input that changes the demand. Build with that in mind, or build the same shortage one size larger.

Solow's line.

In 1987, economist Robert Solow wrote one sentence that outlived the article it was in. "You can see the computer age everywhere but in the productivity statistics." Firms were pouring money into computing. The output numbers would not move.

Economists argued about it for a decade. Was the technology weak. Was the measurement wrong. Was the gain real but delayed. The honest answer turned out to be the least satisfying one. The gain was real, and it was delayed, and it would not arrive until the work itself was rebuilt around the machine.

The EHR was supposed to pay for itself.

HITECH put roughly 27 billion dollars behind EHR adoption. The pitch was efficiency. Fewer errors, less duplication, faster care, lower cost. Adoption went from a minority of hospitals to nearly all of them in under a decade. On an adoption chart, it is one of the fastest technology rollouts in the history of American healthcare.

The productivity curve did not follow. The technology is everywhere. The promised output gain is the part that is hard to find in the numbers.

Where the time went.

The clearest place to see it is the clinician's day. Time-and-motion research in ambulatory practice found physicians spending close to two hours on the EHR and desk work for every hour of direct patient care.

That is the paradox in one image. The EHR did not fail to do work. It generated new work. Documentation that is more thorough and more defensible. Billing capture that is more complete. Inbox messages, results, and alerts at a volume the paper chart physically could not produce. Each is arguably useful. Together they ate the dividend.

Does the paradox break?

Solow's version did. US productivity did climb in the late 1990s. It happened once firms stopped bolting computers onto old workflows and started redesigning the workflow around the computer. The lag was not the technology failing. It was the organization catching up.

Healthcare is now standing at the same fork with AI. The ambient scribe, the inbox drafter, the imaging pre-read all promise to give time back. The Productivity Paradox says the time will not show up on its own. It shows up only if someone redesigns the visit, the documentation standard, and the staffing model to actually claim it. Otherwise the saved minutes get absorbed exactly the way the EHR's did.

The 1967 paper nobody could ignore.

Melvin Conway submitted a paper called "How Do Committees Invent?" Harvard Business Review rejected it. It ran in Datamation in 1968 and never stopped being quoted.

The claim was specific. If four teams build a compiler, you get a four-pass compiler. The software cannot help it. Every interface in the system has to be negotiated, and the negotiations follow the lines of communication that already exist. So the product comes out shaped like the organization, because the organization is the only blueprint the builders actually share.

You ship the org chart.

The cleanest way to see it is side by side. The hospital has its departments. The build has its modules. The shapes are not similar by coincidence.

Walk any Epic or Oracle Health build and you can read the hospital's politics straight off the screen. The system did not invent those walls. It poured concrete around the ones already standing.

Why interoperability stalls.

Conway's Law explains the single most expensive failure in health IT. Two systems will not talk to each other, and a decade of standards has not fixed it.

The HL7 interface, the FHIR endpoint, the data-sharing agreement. Each one looks like a negotiation between computers. It is actually a negotiation between departments, or between health systems, or between competitors. The standard is mature. The organizations are not aligned. The data stays put.

Use it backward.

The useful version of Conway's Law runs in reverse. If the system you want has a certain shape, the team has to have that shape first. Software people call this the inverse Conway maneuver, reorganizing the builders deliberately so the architecture comes out right.

For a health system the lesson is blunt. A build team siloed by department will ship a siloed build. Every time. No analyst is good enough to overcome the org chart they report into. If you want an EHR that crosses departments, you need a build governance that crosses departments first.

One sentence, two errors.

Roy Amara ran the Institute for the Future and is remembered for a single line. We tend to overestimate the effect of a technology in the short run and underestimate it in the long run.

Most laws in this guide name one failure mode. Amara's names two, and they point in opposite directions. That is what makes it hard to use. You cannot just be a skeptic, and you cannot just be a believer. You have to hold a wrong short-run forecast and a real long-run direction in your head at the same time.

The curve that crosses.

Expectation spikes early. It runs far ahead of anything real, then collapses when the demos do not become deployments. Actual impact does the opposite. It lags, looks like nothing for a long time, then climbs, and eventually passes the point the hype promised, just years late.

Healthcare AI is living both halves.

The short-run overestimate is loud right now. AI will read every scan by next year. The ambient scribe will end documentation burden by the next budget cycle. It will not. Integration friction, liability, reimbursement, and clinical trust all move slower than a demo.

The long-run underestimate is the quieter, more dangerous error. The same skeptics who watched IBM Watson for Oncology collapse now assume the whole category is a fad. That is the trap. Watson was the short-run overestimate. It says nothing about whether the long run is empty.

Planning for both.

The honest position is uncomfortable. The current wave is almost certainly overhyped on TIMING, and almost certainly underhyped on EVENTUAL reach. A health system has to plan for both at once. Do not rebuild this year's workflow around a capability that will not be reliable for three. And do not let this year's disappointment talk you out of the capability that is genuinely coming.

It pairs directly with the Hype Cycle. The Hype Cycle draws the shape. Amara's Law states the lesson hidden inside it. The trough is the short-run error correcting. The plateau is the long-run truth arriving.

Five phases, one shape.

It starts with an innovation trigger, a demo, a launch, a paper. Expectations climb fast to the peak of inflated expectations, where every conference keynote promises the technology will fix everything. Then reality arrives. Pilots stall, integration is hard, and the mood drops into the trough of disillusionment. The survivors climb the slope of enlightenment as real use cases get sorted from fantasy ones, and level off at the plateau of productivity, lower than the peak, but actually real.

Healthcare keeps riding it.

Telehealth is the cleanest recent lap. It hit the peak in 2020, when every visit was suddenly virtual. By 2023 the takes had flipped to "telehealth is dead" as in-person volume returned. That was the trough. What is left now is the plateau, a smaller, durable role for the visits that genuinely work remotely.

EHRs ran the same arc over a longer decade. Meaningful Use was the peak. Burnout and "the EHR ruined medicine" was the trough. What exists now is a grudging plateau, indispensable and still resented.

The phase tells you how to act. Near the peak, the job is to resist the keynote and ask what actually integrates. In the trough, the job is the opposite, to NOT write off a technology that is simply unfashionable this year. Clinical AI is sitting near the peak. Plan for its trough before you are standing in it.

Reading your own position.

The Hype Cycle is most useful as a question, not a forecast. For any technology in front of you, ask where it sits. If every vendor is selling it and no peer has run it past a pilot, you are at the peak, and the move is patience. If the conference circuit has gone quiet and the early adopters are grumbling, you are in the trough, and the move is to look hard at whether the quiet ones are still using it.

The plateau is the only phase where buying on the hype and buying on the evidence point the same direction. It is also the least exciting phase, which is exactly why it gets the least attention.

A law from the supply chain, not the clinic.

In 1997, Hau Lee and colleagues at Stanford gave a name to something Procter & Gamble had been staring at for years. Consumer demand for diapers was about as stable as demand gets. Babies are not seasonal. Yet the orders P&G's factories saw swung wildly.

The distortion was not caused by anyone behaving badly. It was caused by distance. Each link in the chain could only see the link directly in front of it, and reacted to that order rather than to the real, calm demand at the far end. Lee called the result the bullwhip effect. A small flick at the handle, a crack at the tip.

The same signal, amplified.

Picture one steady, mild fluctuation in what patients actually use. Then watch what each link upstream does with it.

2020 was the textbook case.

A modest, real rise in demand for masks and gloves hit the front of the healthcare supply chain. Then every link did the rational thing. Hospitals, uncertain, ordered extra. Distributors saw the spike, could not separate signal from panic, and ordered more extra to protect themselves. Manufacturers saw that, and built capacity for a demand curve that was already half fear.

Then the orders that were really just hoarding got cancelled. The whip cracked back. Gluts, then shortages, then gluts again, long after bedside demand had settled.

Not just a pandemic story.

PPE was the loud version. The quiet version runs every year. A single plant goes offline and the contrast-media supply swings for months. A sterile-injectable line has a quality hold and saline, or a routine generic, lurches between glut and shortage. The federal drug-shortage list is, in large part, a bullwhip readout.

Health IT can dampen this or it can feed it. Shared consumption data, real point-of-use signal instead of guarded order data, shrinks the whip. A faster ordering portal with no shared signal just hands every link a quicker button to overreact with.

The law, and where it comes from.

The name points at Jensen Huang, who runs NVIDIA. The looser version, sometimes called Huang's Law, is the observation that AI compute performance has been improving faster than Moore's Law alone would predict, because the gains come from chips, software, and networking all at once.

The sharper, reframed version treats a data center as an AI factory. Its product is throughput, tokens per second of useful model output. Its hard limit is not capital. It is megawatts. Once power is the ceiling, any spend that improves performance per watt does not just trim cost. It raises how much product the same ceiling can hold. Spend more, make more. Spend more, save more per unit.

Throughput climbs faster than cost.

That is the whole claim in one shape. Capital spend rises. Monetizable throughput rises faster. The widening gap between the two lines is the margin, and it is what keeps the price of a unit of compute falling.

Why healthcare is downstream.

Jensen's Law is a statement about compute, not about clinics. But healthcare AI does not run in a vacuum. It runs on exactly these factories. The cost curve of every clinical model, every ambient scribe, every imaging pre-read is set by whether the people building data centers can keep buying throughput faster than they buy cost.

When the factories keep winning, inference keeps getting cheaper, and cheaper inference is the fuel for the Jevons loop in the first tab. Jensen's Law sets how fast the price of a clinical AI call falls. Jevons' Paradox says that as it falls, healthcare's total demand for those calls rises to meet it.

What it means for a health system.

The practical read is narrow and a little uncomfortable. The economics of healthcare AI are partly out of your hands. They are being decided in power contracts and chip fabs you will never see. The reasonable planning assumption is that the price of inference keeps falling, because the law upstream is built to push it down.

Read it alongside two neighbors in this guide. Jevons says the falling price will pull total demand up, not down. Amara says the timing of all this will be slower than the keynote promises. Plan for cheap inference, plan for the demand it unlocks, and do not bet the budget on the date.