In “Harper,” episode six of the CBS legal drama Cupertino, a young woman loses a liver transplant shortly before surgery to a wealthy investor. The episode...
Cupertino’s “Harper” Raises a Real Question: Can AI Allocate Transplants Fairly?
In “Harper,” episode six of the CBS legal drama Cupertino, a young woman loses a liver transplant shortly before surgery to a wealthy investor. The episode frames the reversal as possible algorithmic bias, then uncovers a physician-executive’s alleged pay-to-play override. Its strongest real-world question is not whether AI should choose who lives, but how transplant systems can use data without hiding human decisions.
Key Takeaways
- The episode’s transplant dispute is fictional; it is not evidence of a real allocation case.
- MELD 3.0 is a statistical risk score, not an autonomous AI system.
- Organ allocation also depends on policy rules, compatibility, donor location, and clinical judgment.
- AI tools may help identify patterns or hard-to-place organs, but proposed uses need testing in real clinical settings.
- Human oversight, clear records, and fair review matter whether a decision comes from a formula, a machine-learning model, or a person.
The story puts the transplant decision into a tight four-hour window: Harper Gervin, 27, is waiting for a liver, while Albert Ensler, a wealthy venture capitalist, receives the organ just before her operation. Michael Price, Olivia Siffre, and Judge Hornstein are stuck in U.S. 101 traffic, so the lawyers argue by video. Their early suspicion is that an algorithm discriminated against Harper. They eventually learn that the allocation system selected her, but a hospital physician-executive intervened after Ensler’s foundation offered a donation for cancer research. Christy Palladino exposes the exchange, and Harper’s claim is restored.
That plot makes for brisk television. It also compresses several distinct parts of transplant medicine into one courtroom fight: medical urgency scoring, organ matching, allocation policy, institutional discretion, and possible corruption. Those parts should not be confused. When I separate the scene’s legal suspense from real transplant policy, the key distinction is that a score can inform a decision without being the decision-maker. The system needs both a defensible method and a way to identify people who misuse their authority.
What Is AI in Transplant Allocation?
Transplant allocation is the process of matching donated organs with people who need them, under national policies intended to make decisions medically grounded and consistent. It is not a single ranking exercise. A person’s eligibility to join a waiting list is different from the order in which candidates may be considered for one particular organ. After a donor becomes available, factors such as blood type compatibility, medical urgency, distance from the donor hospital, and organ-specific rules can all matter.
The Model for End-Stage Liver Disease (MELD) score illustrates why precision in language matters. MELD estimates the risk of death for people with severe liver disease and helps prioritize adult liver candidates. It is a statistical formula, not an artificial-intelligence agent that independently reviews a patient and makes an ethical choice. The updated adult score, MELD 3.0, went into use in the United States in 2023. Its development addressed disadvantages in the previous approach, including ways women could be disadvantaged by differences in measured laboratory values. The clinical rationale and analysis are described in the New England Journal of Medicine study of MELD 3.0.
AI usually refers here to computational methods that find patterns in data, estimate an outcome, or support a decision. Machine-learning systems can process many variables at once, while deep-learning systems use layered computational models to detect patterns in complex data. These methods could help clinicians estimate risk, support organ placement, or flag candidates or organs for further review. None of that makes a model inherently fair, accurate, or appropriate for deciding who receives an organ.
Another source of confusion is continuous distribution. This is an allocation policy framework that replaces some rigid categories with weighted factors, so candidates can be assessed across a broader set of considerations. It is not itself AI. Lung allocation in the United States adopted a continuous distribution framework in 2023. Policy details and updates are available through the Organ Procurement and Transplantation Network policy library. A weighted policy can still use conventional formulas, and an AI tool could be used to assist a policy without controlling it.
That separation matters in the episode. Harper was reportedly selected by the allocation algorithm before a hospital executive overrode the result. If that fictional sequence is taken at face value, the alleged wrongdoing was not an algorithm choosing Ensler; it was a person changing the outcome for an improper reason. A technically sound model cannot prevent every misuse of authority. Nor can human oversight be a magic phrase that excuses undocumented exceptions. The useful question is who can change a recommendation, for what reasons, and how the change is recorded and reviewed.
Core Benefits and Importance
AI could contribute to transplant care in practical, bounded ways. It can help clinicians and organ procurement organizations sort large amounts of clinical information, identify candidates for closer review, or bring attention to organs that might otherwise be difficult to place. That is different from handing the final ranking to a machine. In a system where time matters and organ offers must be considered quickly, a well-tested support tool might help teams see options sooner.
- Finding patterns across complex records. Traditional risk scores are designed around particular variables and goals. A machine-learning model may detect interactions among clinical measures that a simpler equation does not capture. That can make it useful for research or decision support, provided the patterns hold up in independent patient groups and improve decisions that matter to patients.
- Supporting placement of hard-to-place organs. Researchers are developing deep-learning approaches to help identify kidneys that may be suitable for accelerated placement. Earlier identification could give transplant teams more time to consider an offer and may help reduce organ discard. But a promising prediction is not proof that a kidney is safe for a particular recipient. Surgeons and transplant teams still need clinical assessment, and the model’s performance must be measured against real outcomes.
- Making complex reviews more consistent. A prototype multi-agent system described in a retrospective proof-of-concept study used separate large-language-model agents to simulate perspectives such as a transplant surgeon, hepatologist, cardiologist, and social worker. The study reported roughly 92 percent accuracy for predicting one-year post-transplant survival and 98 percent accuracy for identifying contraindications. Those figures sound impressive, but the design matters: the analysis was retrospective, included artificially assigned contraindications, and did not test the system inside an operating transplant selection committee. It is evidence of technical potential, not evidence that such a system is ready to choose recipients.
- Helping people notice omissions. A decision-support tool might prompt a team to check a factor that was overlooked or to explain why a candidate is not being considered. Used properly, that can improve the completeness of discussion. Used carelessly, it can simply give a polished computer-generated gloss to a rushed decision.
- Improving auditability—if designed for it. A system can log which data it used, what it recommended, and whether a clinician departed from that recommendation. Such records can make review more concrete. They do not automatically reveal whether the original data were biased or whether an override was justified, but they give auditors something specific to examine.
The benefits are conditional. A model trained on one hospital’s data may not perform well for another region, a different patient population, or a new clinical practice. Missing or inaccurate data can distort an output. A model can also learn historical patterns that reflect unequal access to care rather than sound medical reasons for allocation. The World Health Organization’s guidance on AI ethics and health stresses human rights, accountability, transparency, and careful oversight. Those principles apply especially strongly when the decision concerns a scarce organ and a person’s chance to receive life-saving care.
Frankly, “the computer said so” is not a defensible explanation. A patient deserves to know how a recommendation affected their care, and a transplant team needs to know where a model is uncertain. The body is not a spreadsheet, and a person’s worth cannot be reduced to a risk estimate. Good medicine uses information carefully while preserving the dignity of the patient in front of the team.
A Step-by-Step Guide to Responsible Use
There is no responsible shortcut from a promising research result to a tool that ranks real transplant candidates. A safer path starts with a clear clinical purpose and proceeds through testing, oversight, and ongoing review. Here is what that path should include.
- Define the job before choosing the model. Is the tool estimating short-term medical urgency, helping locate a suitable organ, identifying missing information, or predicting a particular outcome? Those are different tasks. A system built to predict survival should not quietly become a system that decides a person’s social worth or overall eligibility. The intended use must be narrow enough to test and explain.
- Name the decision-maker and the limits. The team should know whether an output is advisory, how much weight it can receive, and who is authorized to act on it. If the tool only flags an organ for review, say so. If it influences a candidate ranking, document that role and ensure the relevant transplant policy permits it. A model must not gain authority by habit simply because staff see it every day.
- Check the data and the population. Developers should assess whether the data represent the patients and hospitals where the tool will be used. They should examine missing values, data quality, and performance across relevant groups, including groups that may have been disadvantaged by older scoring approaches. This is not a one-time box to tick. Patient populations, clinical practice, and allocation policies change.
- Validate outside the development setting. A model that performs well on records it was designed around may fail elsewhere. Independent testing should measure the outcomes tied to its stated purpose, including error rates and performance for different patient groups. Retrospective studies can identify promising ideas, but prospective evaluation is needed to learn how a tool behaves when clinicians use it in actual care. The U.S. Food and Drug Administration’s overview of AI and machine-learning medical devices explains why intended use and appropriate evaluation matter for medical software.
- Keep a human review that can disagree. Oversight is meaningful only if clinicians can question the output, understand its limits, and document a reasoned departure. A reviewer should not be pressured to accept a model because it appears mathematically precise. Equally, a person should not be able to override the recommendation for an undisclosed reason. The system needs defined authority, records, and a route for independent review.
- Explain the decision to patients and the public. A transplant program should be able to say what information matters, how an algorithm contributes, and how a patient can ask questions or raise a concern. Explanations need not expose private information about other candidates. They should still make the process understandable enough to earn scrutiny and trust.
- Monitor after deployment. Real-world performance can drift when patient data, medical practice, or allocation rules change. Programs should track outcomes, disparities, overrides, and errors, with a clear process for pausing a tool if it performs poorly. A model that passed validation years ago does not receive permanent permission to shape decisions.
When I analyze the Harper plot against this checklist, the fictional hospital’s alleged override is the most revealing failure point. An allocation system needs safeguards against biased predictions, but it also needs logs and independent review of human departures from its recommendations. Responsible oversight asks both whether a machine is making a flawed suggestion and whether a person is changing a sound one for an improper reason.
Comparison Table: AI Decision Support vs. Conventional Allocation Tools
| Feature | AI decision support | Conventional allocation tools and policies |
|---|
| Typical role | May detect patterns, estimate outcomes, or flag cases for review | Apply defined scores, compatibility rules, and policy criteria |
| Example | A research model that helps identify a potentially placeable kidney | MELD 3.0 for adult liver urgency, alongside allocation policy |
| Strength | Can analyze complex combinations of data at scale | Established rules can be specified, applied, and audited against policy |
| Main risk | Hidden bias, weak generalization, opaque output, or overreliance | Imperfect assumptions, narrow variables, or unfair effects in a score |
| Status | Many proposed transplant uses remain experimental and need validation | Used within established allocation systems, subject to policy and revision |
| Proper authority | Decision support should not silently become final authority | Policy governs allocation; clinicians apply clinical judgment within it |
This is not a contest in which one side is automatically fair and the other automatically flawed. A conventional score can encode a poor assumption; an AI model can miss a meaningful patient difference. A policy can be transparent yet produce hard trade-offs, while a sophisticated model can make its reasoning difficult to inspect. The comparison is useful because it shows why the target should be a fair, reviewable process—not novelty for its own sake.
Common Mistakes to Avoid
Calling every formula AI. MELD 3.0 uses statistical methods, but it is not an autonomous machine-learning system. Calling it AI blurs how the score works and can make it harder to ask the right questions about its design. Be precise: distinguish a fixed score, a machine-learning prediction, a policy framework, and a clinician’s judgment.
Assuming the algorithm made the controversial choice. In the episode, the team learns the allocation system selected Harper before a physician-executive changed the outcome. That detail matters. A system’s recommendation and a hospital’s final action are not necessarily the same. Reviews should preserve both: what the tool or policy recommended, and what people did next. Otherwise, a human override may be incorrectly blamed on a model, or a model’s influence may be hidden behind a human signature.
Treating a high accuracy number as a safety certificate. Accuracy can be misleading without knowing the study population, outcome definition, and consequences of different errors. A retrospective model that performs well on a curated dataset has not shown that it can operate safely under time pressure, with incomplete records, or across multiple transplant centers. The multi-agent study’s reported figures are preliminary results, not a clinical endorsement.
Assuming human judgment is automatically fairer. People can consider context that a model misses, but people can also be inconsistent, fatigued, influenced by conflicts of interest, or wrong. The answer is not to replace every human decision with a model, nor to romanticize the committee room. Good systems test both human and machine contributions and create ways to challenge each.
Confusing a donation with a legitimate allocation factor. The episode alleges that a research donation influenced an organ decision. That is a plot point, not a rule of U.S. transplant allocation. The allocation of a donated organ should follow applicable medical and policy criteria, not a recipient’s wealth, a foundation’s contribution, or an institution’s financial interest. A potential conflict calls for documentation and independent investigation, not a hurried assumption that every donation has tainted a decision.
Treating continuous distribution as AI. Continuous distribution changes how allocation factors are combined. It does not automatically mean a computer-learning system is making the ranking. Policy design, mathematical scoring, and AI are related subjects, but they are not interchangeable labels.
Forgetting the difference between urgency and overall suitability. A candidate’s medical urgency, compatibility with a particular organ, and expected post-transplant outcomes are related but distinct questions. A model trained to predict one cannot be presumed to answer all three. Combining them into one opaque number may conceal the value judgments that policy-makers and clinicians should discuss openly.
Calling oversight a cure-all. A human reviewer who cannot question an output, lacks time to review it, or is rewarded for following it is not meaningful oversight. Likewise, a discretionary override with no record creates room for arbitrary treatment. Oversight must be practical: authority, training, documentation, patient recourse, and independent audits all matter. Here’s the kicker: a person can misuse a fair tool, and a flawed tool can make a person’s mistake look official.
Frequently Asked Questions
Is MELD 3.0 an AI system?
No. MELD 3.0 is a statistical risk score used to help prioritize adult liver candidates according to medical urgency. It is not an autonomous AI agent, and it does not independently decide who receives a particular donated liver. A specific organ offer is considered under allocation rules that can include compatibility, donor location, urgency, and other policy factors. The score and the allocation system should not be treated as synonyms.
Can AI decide who gets a transplant today?
AI tools may be researched or used in limited decision-support roles, but that does not mean a fully autonomous system is making transplant decisions. Proposed models need clinical validation, appropriate governance, and monitoring before people rely on them in high-stakes care. The multi-agent study described above was a retrospective proof of concept, not a demonstration that an AI can safely serve as a transplant selection committee. Final decisions remain governed by clinical practice and applicable allocation policy.
What does continuous distribution mean?
Continuous distribution is a policy approach that weighs several allocation factors across a broader scale rather than relying only on rigid categories. It is an allocation framework, not an AI technology. The United States adopted continuous distribution for lung allocation in 2023. Whether a particular system uses machine learning is a separate question from whether its policy uses weighted points.
Can AI reduce organ discard?
Potentially, in certain settings. Researchers are developing tools to help identify organs that may be suitable for accelerated placement, including kidneys that could otherwise be difficult to place. Earlier identification could give transplant teams more time to evaluate an organ and may help reduce discard. But a model’s suggestion does not establish that an organ is safe for a particular recipient. Clinical assessment, validation, and monitoring are still necessary, and a claimed reduction in discard needs to be demonstrated with real-world evidence.
Final Thought
“Harper” gives its audience a clean villain: a powerful executive who allegedly changes an organ decision after a wealthy donor offers money. Real transplant decisions are rarely so tidy. They involve scarce organs, medical uncertainty, detailed rules, and people who must make consequential judgments under pressure. A fictional courtroom can resolve that conflict in one episode. A trustworthy health system has to prevent, detect, and correct unfair decisions every day.
AI may help by sorting complex data, identifying overlooked patterns, or bringing potential organ offers to a team sooner. Those are worthwhile aims. They do not justify treating a model as an impartial oracle, and a strong retrospective result is not a substitute for prospective clinical validation. Nor should concerns about AI distract from human choices: who set the criteria, who approved an exception, whether a conflict was disclosed, and whether someone outside the decision-making chain can audit the record.
The sound standard is neither machine rule nor unchecked discretion. It is a transparent process in which data support clinical judgment, policy sets clear boundaries, and every consequential recommendation or override can be examined. Transplant candidates are not merely rows in a database; they are people whose lives and dignity deserve careful consideration. That principle should shape every tool, score, and human decision involved in allocation. The question raised by Harper’s case is therefore less whether AI is good or bad than whether the whole system can explain its decisions—and hold people accountable when those decisions are changed.