Curious Boundary

Utah’s AI Prescriptions and Cricket’s Data Problem: When Pattern-Matching Meets the Exception

Utah's AI Prescriptions and Cricket's Data Problem: When Pattern-Matching Meets the Exception

Written by

Curious Boundary

in

,
Quick summary (click to reveal)

Summary: Utah authorized AI to issue initial prescriptions for topical treatments in a phased pilot starting October 2026, drawing criticism from physicians. Cricket offers a 150-year case study in the same tension: relying on population-level statistics for team selection and tactics often fails against exceptional batsmen or unusual conditions. Both systems struggle to handle incomplete information and contextual exceptions that data cannot anticipate.

Full article below.

On 7 October 2026, Nolla Health announced that Utah had authorised the first U.S. pilot program allowing artificial intelligence to issue initial prescriptions without a physician’s immediate approval. The AI selects from a physician-approved list of topical treatments; if it cannot confidently choose, the patient is referred to a licensed doctor. Three staged phases, requiring 95 per cent agreement with physicians and zero serious adverse events before escalation, will determine whether the system can work. Medical associations have protested. Yet the core problem the Utah pilot exposes is not new to high-stakes decision-making: how do you trust a system trained on population-level patterns when the individual case is an exception?

Cricket has been answering this question imperfectly for over 150 years. Every selector, coach and captain knows that statistical averages tell you something important about a batsman’s quality. They also know that those averages can mislead catastrophically when applied to the wrong pitch, the wrong opposition, the wrong moment in the match. The tension between data-driven pattern recognition and contextual judgment is not a flaw in either system. It is the problem neither system has fully solved.

Why Averages Betray You

A batter’s career average is a population statistic. It lumps together performances on hard pitches and soft ones, against world-class attacks and moderate bowlers, in winning situations and losing ones. When you compare two batsmen by career average alone, you are comparing apples harvested from different orchards in different seasons.

Cricket researchers have known this for decades. Pitch conditions influence batting and bowling performance so profoundly that identical scores on good versus bad pitches reflect entirely different quality of play. A century scored on a deteriorating surface against a hostile attack is not equal in value to a century scored on a batting paradise. Yet most selection meetings still begin with a folder of career numbers.

The problem deepens when you meet the outlier. Multivariate analysis of Test cricket performance identifies exceptional batsmen statistically, those who perform in ways the aggregate data does not predict. VVS Laxman is a famous example: his performance index of around 1.0 is an outlier well above the cluster where most top batsmen sit, below 0.7. Yet knowing that Laxman is an outlier does not tell you whether he is an outlier because he is exceptionally good or because he batted in peculiar circumstances, or both.

The Utah AI faces the same problem in reverse. It is trained on prescription data: thousands or millions of cases where a topical treatment was prescribed, what happened, who had an adverse reaction. The AI learns patterns from that population. But it cannot know whether your specific case is like the population or like the outlier whose circumstances were too rare to register in training data. A human clinician, by contrast, can sense that something feels missing. An AI model cannot.

Context Overrides the Pattern

Cricket coaches use statistical analysis relentlessly. They track which bowlers perform well against which batsmen on which pitch types. They identify matchups that favour the team. Yet in match situations, the aggregate statistic routinely gets overruled. A batter in poor form might be dropped even though career average says he is a good player. A bowler might be rested not because his average is poor but because his body language suggests he needs recovery. A batsman known to be strong on hard pitches might be promoted in the order on a soft, slow surface because the match situation demands a quick fifty, not a cautious build.

These are contextual judgments. They require knowledge of the individual’s current state, the opposition’s current strengths, the pitch’s likely evolution and the match’s demands. No population-level statistic captures all of that.

Medical decision-making works the same way. Evidence-based medicine does not offer rigid rules. It offers guidelines, because clinical reasoning is too complex to anticipate all unique patient circumstances. A prescription that is safe for 99 per cent of a population might be dangerous for the 1 per cent with an unrecorded allergy, an undisclosed pregnancy, a drug interaction from a medication started last week, or a rare genetic variant. The fuller the patient’s record, the safer the decision. The more incomplete the record, the more context is guesswork.

Who Bears Responsibility When Both Systems Fail?

Utah’s pilot is designed to fail safely. The first 100 patients see all prescriptions reviewed by two independent physicians. If the system performs well (95 per cent agreement, zero serious adverse events), the next 500 patients get prescriptions issued directly with weekly physician review. Only the final stage samples 10 per cent monthly, with escalation for complications. Staged deployment, written state approval required to advance, and human oversight built in.

Yet staged oversight creates a new risk: the illusion of risk management. When does weekly review become monthly review? When does monthly review become no review? Cricket teams deploy new selectors and analytics tools the same way, gradually increasing their authority until someone has to ask: do we actually need the coach anymore, or is the algorithm enough?

The answer in cricket is nuanced. Analytics have made selection smarter in many cases. They have also produced selection decisions that no experienced cricket mind would have made: selecting a batsman in terrible form because the algorithm weighted recent matches heavily, dropping a bowler mid-tour because a three-game average looked bad, promoting a batter with no experience in the position because his T20 strike rate was high. The human override has prevented disasters, but only because humans still exercise it. Remove the human and the system’s failures become the team’s failures.

In medicine, the legal and ethical accountability is clearer, yet still unsettled. If an AI issues a prescription based on incomplete information, and the patient is harmed, is the responsibility on the system designer (for building inadequate safeguards), the operator (the AI company deploying it), the regulator (for permitting deployment), or the patient (for accepting the risk)? Utah is answering that question in real time, but the answer is not yet written.

Cricket offers no final answer either. When a selector makes a call based on incomplete analysis and the player fails, the selector is blamed. When an AI would have made a different call, and the player fails anyway, is the blame shared? Most cricket organisations have not yet had to face that question directly. Medicine is about to.

Sources

More posts