Why do language models struggle with historical legal cases?

The Supreme Court overruling benchmark (236 case pairs) reveals a failure mode in legal AI that differs from hallucination or shallow reasoning: era sensitivity. Models show systematically degraded performance on historical cases compared to modern ones. The benchmark authors interpret this as "fundamental temporal bias in their training" — the training corpus over-represents recent legal cases, creating a recency advantage that manifests as accuracy drop when reasoning about older precedent.

This is a specific form of the training data distribution problem. Legal databases heavily weight recent cases: they are more frequently cited, more thoroughly documented, more often the subject of commentary. Historical cases, even influential ones, appear less frequently and in more varied contexts across the training corpus. The result is that models have shallower and less reliable representations of historical legal reasoning than their performance on modern cases would suggest.

The practical implication for legal AI deployment is significant. Legal research is not temporally bounded — historical precedent is often decisive, and cases from the nineteenth century can be binding authority. A system that performs well on modern case identification but degrades on historical material creates a systematically misleading picture of its reliability. The practitioner can't know which queries fall into the historically degraded zone without testing each query against the temporal distribution of the relevant legal corpus.

This connects to a broader temporal pattern: Why do language models ignore information in their context? shows that training frequency shapes what models reliably retrieve, even when contrary information is present in context. Era sensitivity is the legal-domain instantiation of this — temporal frequency distribution in training determines reliability, not just factual accuracy of the training data itself.

The mechanism also suggests a partial intervention: domain pre-training on historical legal corpora, or retrieval augmentation that specifically weights historical documents, could partially correct the recency bias. But it would need to be intentional — the bias is invisible in aggregate accuracy metrics that don't break results out by case era. The architectural alternative is to avoid the temporal boundary altogether: Why do search agents beat memorized retrieval on hard questions? — real-time search escapes era sensitivity by definition, since it retrieves from current document stores rather than compressed training representations.

The anachronism problem generalizes beyond legal reasoning to historical language simulation. A separate study (Can Language Models Represent the Past without Anachronism?) shows that prompting contemporary models with period prose does not produce output consistent with period style. Fine-tuning produces results convincing enough to fool an automated judge but human evaluators still detect the anachronism. The authors tentatively conclude that pretraining on period prose is required for reliable historical simulation — fine-tuning cannot undo the temporal contamination of contemporary pretraining. This means the era sensitivity failure mode operates at two levels: factual (knowing what historical law said) and stylistic (producing text consistent with historical linguistic norms). Both require period-specific pretraining to overcome, not just fine-tuning or retrieval.