IPHS 300: AI for Humanity
Document Type
Poster
Publication Date
Spring 2026
Abstract
This project examines whether frontier Large Language Models (LLMs) can serve as structured “expert critics” in economic and legal reasoning. I built a panel of AI judges composed of GPT-5.5, Opus 4.7, and DeepSeek V4, to evaluate three antitrust merger retrospectives: Carlton et al. (2022) on AT&T/Time Warner, Hazlett and Crandall (2024) on T-Mobile/Sprint, and Argentesi et al. (2021) on Facebook/Instagram plus Google/Waze. Using a 10-factor weighted rubric, the panel scored each paper on empirical alignment, parameter robustness, counterfactual validity, dynamic adaptability, and analytical rigor. I then synthesized the nine evaluations as the “Highest Judge,” comparing where the models converged, disagreed, and flagged omitted assumptions. The main finding is that frontier LLMs were strongest as adversarial reviewers, not final judges: they exposed fragile assumptions and missing counterfactuals, but could not independently resolve causal proof. Across the panel, causal attribution was the recurring weakness: the papers described post-merger outcomes, but struggled to prove whether those outcomes were caused by the merger itself or by broader market shifts.
Recommended Citation
Portillo, Pablo, "The Digital Supreme Court: A Meta-Analysis of Antitrust Retrospectives Using Frontier LLMs" (2026). IPHS 300: AI for Humanity. Paper 62.
https://digital.kenyon.edu/dh_iphs_ai/62
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.