Harmless AI Error

Abstract

The proliferation of artificial intelligence (“AI”) in criminal proceedings has brought the technology’s propensity to err into sharp focus. Whether AI is more or less accurate than human judgment is a favorite subject of recent scholarly debate, but what is certain is that AI methods generate particular kinds of errors that are distinct from those produced by humans. Indeed, the data science techniques that underlie many AI applications can involve processes so distinct from human reasoning and investigation—for instance, large language model (“LLM”)’s practice of converting words to numbers and back to words—that the resulting errors can be difficult to anticipate or altogether detect.

This poses a special challenge for constitutional criminal procedure, and
particularly the harmless error doctrine. Harmless error, which has been describedby Landes and Posner as “probably the most cited rule in modern criminal appeals,” provides that appellate courts can uphold criminal sentences even where trial courts made errors—including constitutional errors—so long as said errors would not be dispositive of the outcome. As AI-generated errors may not align with traditional notions of harmlessness, they complicate courts’ ability to assess the impact on defendants’ rights. Moreover, the technological valence of these errors may obscure constitutional violations. Courts thus need more particularized standards for evaluating the harms associated with AI error.

This Article takes the first step towards addressing this challenge by proposing a procedural adjustment: the creation of an “AI error threshold” to distinguish errors amenable to traditional harmless error analysis from those demanding heightened scrutiny akin to structural constitutional errors. By recognizing arbitrariness as a doctrinal fault line, courts can better uphold defendants’ rights amidst technological evolution, ensuring fair adjudication in the AI age.