Social scientists increasingly use large language models (LLMs) to classify text at scale, raising a key question: when can LLMs replace expert human annotation? Prior work found that earlier generative models failed on complex social science tasks while fine-tuned BERT succeeded, but whether current frontier-scale models close this gap remained untested. We investigate this question on a challenging legal reasoning task—classifying paragraphs from U.S. Supreme Court opinions as employing formal, grand, or no reasoning. Testing frontier LLMs including GPT-5.2 and leading open-weight alternatives, we find that even the most capable prompted models consistently underperform fine-tuned BERT. Only when high-parameter-count generative LLMs are fine-tuned on human-annotated training data does performance improve, and fine-tuned BERT remains a cost-effective alternative. Contrary to a common view, our results demonstrate that scaling to frontier-size LLMs does not eliminate the need for expert annotation on tasks requiring deep domain expertise—a finding with important implications for computational social science measurement.

Generative language models (LMs) are increasingly used for document class-prediction tasks and promise enormous improvements in cost and efficiency. Existing research often examines simple classification tasks, but the capability of LMs to classify on complex or specialized tasks is less well understood. We consider a highly complex task that is challenging even for humans: the classification of legal reasoning according to jurisprudential philosophy. Using a novel dataset of historical United States Supreme Court opinions annotated by a team of domain experts, we systematically test the performance of a variety of LMs. We find that generative models perform poorly when given instructions (i.e. prompts) equal to the instructions presented to human annotators through our codebook. Our strongest results derive from fine-tuning models on the annotated dataset; the best performing model is an in-domain model, LEGAL-BERT. We apply predictions from this fine-tuned model to study historical trends in jurisprudence, an exercise that both aligns with prominent qualitative historical accounts and points to areas of possible refinement in those accounts. Our findings generally sound a note of caution in the use of generative LMs on complex tasks without fine-tuning and point to the continued relevance of human annotation-intensive classification methods.