This is Part 3 of a 7-part series exploring how the classic trolley problem manifests in modern technology.

A hospital has one ventilator left. Two patients need it to survive. Patient A is 75 years old with underlying health conditions. Patient B is 40 years old and otherwise healthy. The algorithm must decide who gets the ventilator. The other patient will almost certainly die.

Who should the system choose?

This isn't hypothetical. During the COVID-19 pandemic, hospitals worldwide faced exactly this scenario thousands of times. And increasingly, they turned to algorithms to make the decision.

The trolley problem has moved from the railroad tracks to the ICU.

The COVID-19 Triage Crisis

In March 2020, hospitals in Italy, Spain, and New York faced a nightmare: more patients needed ventilators than ventilators existed. Doctors accustomed to saving everyone suddenly had to choose who would get life-saving treatment and who wouldn't.

The ethical weight was crushing. Individual doctors making life-or-death decisions faced moral injury and burnout. So hospitals turned to triage protocols—systematic rules for allocating scarce resources.

Many of these protocols incorporated algorithmic scoring systems. Patients received points based on age, comorbidities, likelihood of survival, and expected years of life remaining. The algorithm ranked patients. The highest scores got ventilators. The lowest scores got palliative care.

The trolley problem had been automated.

The Pittsburgh Protocol

The University of Pittsburgh's triage guidelines became influential during the pandemic. The protocol used a multi-principle allocation framework that prioritized:

  1. Saving the most lives (utilitarian principle)
  2. Saving the most life-years (prioritizing younger patients)
  3. Giving everyone a fair chance (lottery for tied scores)

Sounds reasonable. But consider what it means in practice: a 75-year-old gets fewer points than a 40-year-old, even if both have equal survival chances. The algorithm doesn't just predict who will survive—it decides who deserves the chance to try.

Critics called it ageism encoded into medical care. Defenders called it rational resource allocation during crisis. Both were right.

Quality-Adjusted Life Years: The Metric That Measures Worth

Many medical allocation algorithms use QALYs—Quality-Adjusted Life Years. The concept is straightforward: one year of perfect health equals one QALY. A year with reduced quality of life (chronic pain, disability, limited mobility) equals less than one QALY.

QALYs let healthcare systems compare interventions: Is it better to give one person ten extra years or ten people one extra year each? Should we fund a treatment that extends life or one that improves quality of life?

But QALYs have a dark side. They systematically devalue the lives of people with disabilities. A person with paraplegia might be assigned 0.6 QALYs per year—their life literally counted as worth 60% of an able-bodied person's life.

During COVID-19, some triage protocols explicitly deprioritized patients with disabilities, reasoning that they had fewer QALYs to save. Disability rights advocates called it eugenics by algorithm. They weren't entirely wrong.

The trolley problem asks: should you save five people or one? QALYs ask: what if the five people are "worth" less than the one?

The Organ Transplant Algorithm

Medical AI makes trolley problem decisions every day, not just during pandemics. The organ transplant system is a permanent triage algorithm.

UNOS (United Network for Organ Sharing) manages the U.S. transplant waiting list using a complex scoring system. For kidneys, the algorithm considers:

  • Medical urgency
  • Tissue match quality
  • Time on waiting list
  • Geographic proximity to donor
  • Pediatric status
  • Prior living donor

The system tries to balance utility (save the most lives), equity (give everyone a fair chance), and efficiency (don't waste organs on poor matches). But these goals conflict.

In 2014, UNOS changed the kidney allocation algorithm to prioritize younger recipients and better tissue matches. The goal was to maximize "graft years"—how long the transplanted organ would function.

The result: older patients and those with rare tissue types waited longer. Some died waiting. The algorithm saved more total life-years but distributed them less equally.

Is that ethical? It depends on whether you prioritize aggregate outcomes or individual fairness. The algorithm can't answer that question—it can only execute whatever values we program into it.

The UK's Controversial Algorithm

In 2020, the UK's National Health Service developed an algorithm to predict which COVID-19 patients would need intensive care. The goal was to allocate limited ICU beds efficiently.

The algorithm used age, sex, and underlying conditions to generate risk scores. High-risk patients got priority for ICU admission. Low-risk patients were treated in general wards.

But the algorithm had a problem: it was trained on historical data that reflected existing healthcare disparities. Minority patients, who had worse health outcomes due to systemic inequities, received lower priority scores. The algorithm didn't cause the disparity—it learned it from data and perpetuated it.

Critics argued the algorithm would worsen existing inequalities. Defenders argued that without the algorithm, decisions would be even more biased, based on individual doctor prejudices rather than systematic criteria.

Both were probably right. The algorithm was biased. The alternative was also biased. There was no neutral option.

The Danger of Optimizing for Metrics

Medical allocation algorithms optimize for measurable outcomes: survival rates, life-years saved, QALYs gained. But not everything that matters is measurable.

What about the value of keeping families together? Should a single parent get priority because their death would orphan children? Should a doctor or nurse get priority because they can save other lives? Should a "good person" get priority over a "bad person"?

Most triage protocols explicitly reject these considerations. They focus on medical factors, not social worth. This seems fair—we don't want algorithms judging who deserves to live based on their social contributions.

But refusing to consider social factors is itself a value judgment. It says that medical outcomes matter more than social consequences. That's not neutral—it's utilitarian.

The trolley problem assumes the five people on the track are interchangeable. Real people aren't. They have families, communities, responsibilities, relationships. Algorithms that ignore these factors aren't being objective—they're choosing to value some things and ignore others.

Who Decides What Makes a Life Worth Saving?

Perhaps the most troubling aspect of medical AI is how it forces us to quantify the unquantifiable: the value of a human life.

When an algorithm prioritizes a 40-year-old over a 75-year-old, it's making a judgment about whose life is worth more. When it uses QALYs, it's deciding how much disability reduces the value of existence. When it considers "expected life-years," it's saying that longer lives are more valuable than shorter ones.

These aren't medical judgments—they're philosophical ones. But they're being made by engineers, data scientists, and hospital administrators, often without explicit ethical deliberation.

The trolley problem was designed to reveal our moral intuitions. Medical AI reveals something else: that we're making these decisions constantly, systematically, at scale, often without realizing we're making them at all.

The Illusion of Objectivity

Hospitals adopted algorithmic triage partly to remove human bias from life-or-death decisions. Algorithms seem objective—they follow rules, they don't play favorites, they treat everyone the same.

But this is an illusion. Algorithms don't eliminate bias—they encode it. Every decision about what factors to include, how to weight them, and what outcomes to optimize reflects human values and priorities.

An algorithm that prioritizes younger patients isn't being objective—it's implementing a specific ethical framework that values life-years over equality. An algorithm that uses QALYs isn't being neutral—it's accepting a utilitarian calculus that many ethicists reject.

The danger isn't that algorithms are biased. It's that they appear objective while executing deeply controversial ethical judgments.

Medical Trolley Problems Are Different

The original trolley problem involves a single decision: pull the lever or don't. Medical allocation algorithms make thousands of decisions, continuously, with compounding effects.

Each individual decision might seem reasonable: this patient has a slightly better prognosis, so they get the ventilator. But aggregate those decisions across thousands of patients, and patterns emerge: older patients systematically deprioritized, disabled patients systematically devalued, minority patients systematically disadvantaged.

The trolley problem asks about one choice. Medical AI makes that choice thousands of times per day, and the cumulative effect is a systematic reordering of whose lives matter most.

What This Reveals About Algorithmic Ethics

Medical AI shows us that the trolley problem isn't just about impossible choices in crisis situations. It's about the everyday decisions that healthcare systems make about resource allocation, priority, and value.

Every medical algorithm embodies answers to questions we haven't fully debated: Should we maximize total life-years or give everyone equal chances? Should we consider quality of life or only quantity? Should we account for social factors or only medical ones? Should we prioritize those most likely to benefit or those most in need?

These aren't technical questions with technical answers. They're ethical questions that require democratic deliberation, not just algorithmic optimization.

Tomorrow, we'll see how similar dilemmas play out in content moderation, where algorithms must decide which harms to prevent and which to allow. The stakes shift from physical survival to psychological safety, but the underlying tension remains: someone must decide, and every decision reflects values that not everyone shares.

The medical trolley problem shows us that we're not just building healthcare systems. We're building systems that make systematic judgments about whose lives are worth saving—and we're doing it with algorithms that hide those judgments behind a veneer of objectivity.

Series Navigation

  • Part 1: The Original Trolley Problem (Sunday, Feb 8)
  • Part 2: Self-Driving Cars (Monday, Feb 9)
  • Part 3: Medical AI (You are here)
  • Part 4: Content Moderation (Wednesday, Feb 11)
  • Part 5: AI Hiring (Thursday, Feb 12)
  • Part 6: Predictive Policing (Friday, Feb 13)
  • Part 7: Synthesis and Frameworks (Saturday, Feb 14)