Humans, Machines, and the Corporate Veil

Key Takeaways

• LLMs and XGBoost classify veil‑piercing opinions with ~92% accuracy.
• Capitalization and corporate formalities emerge as key judicial factors.
• Claude 3.7 Sonnet outperforms the original naïve Bayes model.
• Automated triangulation replaces costly human coding for legal text analysis.

Pulse Analysis

The corporate veil‑piercing doctrine—used to hold shareholders personally liable—has long suffered from analytical opacity. Early scholarship, notably Macey and Mitts (2014), argued that three policy rationales explained virtually all judicial outcomes, sidelining traditional factors such as undercapitalization and formalities. Their methodology combined manual coding of 1,000 cases with a naïve Bayes classifier, achieving a 76.6% match to human judgments. While groundbreaking at the time, the study left open the question of whether more sophisticated computational tools could uncover deeper patterns.

In the latest research, Barnard and Oh harness Claude 3.7 Sonnet, a state‑of‑the‑art large language model, and an XGBoost stacked ensemble to re‑examine the same corpus and expand it to 16,202 opinions. Both models exceed the original agreement rate, reaching roughly 92% accuracy, and uniquely flag firm capitalization and adherence to corporate formalities as decisive cues—elements omitted from the earlier taxonomy. Moreover, the XGBoost model captures intricate word‑pair interactions, suggesting that judges apply conditional logic rather than simple factor weighting. This methodological leap not only refines the academic narrative around veil‑piercing but also illustrates how AI can surface latent judicial reasoning that manual coding may miss.

Beyond the substantive legal insights, the study proposes an automated triangulation framework that substitutes the traditional dual‑coder reliability check. By letting Claude Opus 4.5 and XGBoost independently classify every manually coded opinion, researchers obtain full‑sample coverage, confidence scores, and an inferred ground truth without the bottleneck of expert labor. This approach promises faster, reproducible, and more accurate empirical legal research, positioning large language models as indispensable allies for scholars, litigators, and regulators seeking to decode complex judicial behavior.