In 1999, Sally Clark, a British solicitor, was convicted of murdering her two infant sons after both were found dead in their cribs — deaths she consistently maintained were the result of natural cot death, and for which no clear physical evidence of smothering or abuse was ever established. The prosecution’s expert witness testified that the chance of two natural cot deaths in one family was one in 73 million, a number so astronomically small that it seemed to prove guilt beyond any reasonable doubt. The jury convicted Clark within hours.
However, the statistic was fraudulent, as it resulted from squaring the probability of one sudden infant death syndrome event, treating these two occurrences as independent when genetic and environmental factors make siblings’ deaths linked. More accurate estimates placed the probability of two such deaths in the same family anywhere from 1 in 214 to 1 in 8500: rare, but not nearly as implausible as the prosecution suggested. Yet no one in the courtroom possessed the mathematical literacy to challenge this fundamental error.
Clark’s wrongful conviction foreshadowed a crisis that has only intensified in the digital age. As statistical and algorithmic evidence become central to criminal trials, the legal system’s mathematical illiteracy creates a fundamental due process violation. When one side presents flawed evidence under the guise of scientific rigor and another lacks the tools to challenge it, the judicial process erodes.
The adversarial system rests on a fundamental premise: if each side presents its case, truth will emerge through rigorous cross-examination. But when neither attorneys nor judges understand the mathematics involved, this foundational assumption collapses entirely. Yet, legal education emphasizes textual analysis, not quantitative reasoning. Law students master precedent but not probability; judges preside over cases involving Bayes’ theorem or algorithmic inference without ever having studied these tools. The result is institutional blindness that comes at a profound human cost: Clark died by suicide in 2007, four years after her conviction was overturned — an outcome that, while not reducible to a single cause, highlights the enduring toll of wrongful conviction.
The Clark case is part of a broader pattern of prosecutorial misconduct enabled by institutional mathematical blindness. Consider the 2003 case of Lucia de Berk, a Dutch nurse convicted of murdering patients after prosecutors claimed the odds were “one in 342 million” that so many patients died during her shifts. The calculation systematically ignored her assignment to high-risk cases and engaged in selective data sampling, cherry-picking suspicious incidents while omitting shifts where no deaths occurred. The statistics gave an illusion of scientific rigor to what was ultimately circumstantial evidence. De Berk was later exonerated after over six years in prison, when independent statisticians demonstrated that the prosecution had cherry-picked data and ignored nurse-to-nurse variation in patient mortality rates. These factors, when properly accounted for, recalculated the odds to a far more plausible 1 in 46.
Today’s prosecutorial statistical abuse has worsened with technological innovation. For instance, over 140 U.S. cities rely on ShotSpotter, a gunshot detection system that uses microphones and artificial intelligence to identify gunfire. However, a 2021 investigation revealed that human reviewers can overrule the AI’s conclusions and do so 10% of the time, a step ShotSpotter says improves accuracy by filtering out sounds like fireworks or car backfires. But the system’s design means that what is ultimately presented in court is not a fact, but a conclusion reached through layers of statistical inference and human discretion. When neither judges nor defense counsel can meaningfully examine how that conclusion was produced, the evidence takes on the force of authority without the possibility of challenge, which is precisely the condition under which statistical complexity begins to undermine due process.
Another troubling evolution involves proprietary AI systems that courts treat as infallible despite their opaque methodologies. In 2022, Adarus Black was convicted of murder in Akron, Ohio, based largely on an AI-powered tool called Cybercheck, which claimed it could place his cellphone near the crime scene with over 90% accuracy. Because the system was proprietary, the defense was denied access to its methodology, error rates, and validation studies, making independent review impossible. The court nonetheless allowed the evidence. In such cases, statistical claims are not merely difficult to challenge; they are structurally shielded from scrutiny. This raises a constitutional dilemma: the Sixth Amendment guarantees defendants the right to confront and challenge evidence. However, when the state is permitted to rely on probabilistic conclusions that the defense is barred from examining, the right to confront evidence becomes illusory, allowing statistical proof to operate as an unearned authority within the courtroom.
Meanwhile, facial recognition technology also presents emerging concerns. A 2024 Washington Post investigation revealed that U.S. police have made at least eight confirmed wrongful arrests based on facial recognition matches, with officers often treating AI matches as definitive despite internal guidance labeling them only as investigative leads. While the numbers remain small, the trend suggests growing reliance on probabilistic tools without proper statistical oversight. Though not a case of direct mathematical miscalculation, this situation reflects a growing tendency to treat algorithmic outputs as definitive, despite their probabilistic design. As a pattern, statistical complexity and proprietary design have shielded fragile methodologies from scrutiny, leaving courts ill-equipped to assess the evidence placed before them.
Thus, real reform must address the root cause: mathematical illiteracy within the legal system itself. Consider how patent law evolved with the computer revolution. In the 1980s, federal judges routinely struggled with software patents they could not understand, leading to inconsistent rulings that threatened innovation. The legal system responded by creating specialized patent courts staffed with technically-trained judges, often former patent attorneys or engineers, and requiring technical advisors in complex cases. What seemed impossible—teaching judges to understand technology—became routine institutional practice within a generation. Furthermore, other high-stakes areas within the legal practice already demand statistical literacy. For instance, malpractice law hinges on an understanding of base rates, risk ratios, and outcome modeling. That the entire legal profession does not meet similar standards represents institutional failure of the highest order.
The statistical literacy crisis in the legal profession demands similar institutional courage, though the challenge is arguably greater. Mandatory law school coursework in probability, conditional inference, and scientific reasoning is a start. However, transforming a profession requires sustained commitment; therefore, judges need ongoing training. Furthermore, judges involved in high-stakes criminal trials that concern algorithmic evidence should have access to expert statistical review boards modeled on how technical advisors are used in patent litigation. The precedent exists; what is missing is the institutional will to treat probabilistic and algorithmic evidence with the gravity it now demands.
These reforms may sound idealistic. They require the legal profession to admit its own epistemic limitations and invest in competencies it has historically avoided. But the alternative is accepting that statistical complexity can systematically undermine due process, leaving courts unable to distinguish sound evidence from flawed or manipulative claims.
Algorithms and probabilistic analysis are reshaping how justice is administered. But when these tools enter mathematically illiterate courtrooms, we run the risk of replacing the rule of law with the rule of likelihood. Due process means more than a fair-sounding trial; it means the ability to meaningfully confront evidence. If that evidence is statistical, then the legal system must develop statistical fluency. Until then, courtroom science will remain performative theater, and justice dangerously approximate.
