Harrison C. Margolin
Model evaluations, SpaceXAI
About
I evaluate frontier models at SpaceXAI to advance their capabilities. My research asks how safely models use those capabilities, and how they should be governed.
My technical research focuses on how internal states can predict misaligned behavior. In my safety evaluations of long-horizon agents, I study agent behavior through the lens of behavioral economics. A Push Down the Mountain tests these questions on collusion in a market simulation.
My governance research draws on Posner’s economic theory of tort law, and studies how evaluations can inform regulatory design. In The Ratio Method, I assign values to AI risks using law-and-economics variables, then draw from the regimes that govern similar industrial harms.
Research interests: agent safety evaluations, technical AI governance, experimental economics with AI agents and interpretability and steering.
Research
AI Safety and Evaluations
A Push Down the Mountain: Framing Effects and Internal State Interventions on Collusive Behavior in a Long-Horizon LLM Agent · Working paper, October 2026 (submitted to arXiv) · PDF · Code
Studies the influence of behavioral-economic framing (gain, loss, blame, credit) on agent misconduct in a market simulation. Phase I tests Gemma 3 12B across 260 episodes built from the transcripts of the London silver-fixing case. Among other results, blame framing raised participation in collusion from 5% to 33%. Phase II analyzes the agent’s internal states and reruns the experiments with activation steering.
AI Governance and Liability
Precedents: Matching Frontier AI Risks to Liability Regimes · Report, 2026 · PDF · Code
Profiles the eight frontier AI risks of the International AI Safety Report 2026 on law-and-economics variables, and matches each to the closest of sixteen liability regimes.
The Ratio Method: Addressing Complex Tort Liability in the Fourth Industrial Revolution · St. Mary’s Law Journal, vol. 52, p. 679, 2021 (with Grant Frazier) · PDF
A law-and-economics framework for designing liability around emerging technological risk. Categorizes each risk on a set of variables (e.g., frequency, average damages, causation mechanism, litigation costs), and matches it to precedent regimes. Applied to AI-driven cybersecurity, IoT, gene editing, and biological weapons.
An Analytical Comparison of the Ratio Method and Amodei’s Policy on the AI Exponential · Essay, June 2026 · PDF
Traces convergences between 2026 frontier AI governance proposals and the Ratio Method; separates the two methodological approaches, ex ante controls and liability; assesses the challenges of regulating automated R&D and loss of control.
Economics of AI
AI Investment and Economic Outcomes in the U.S. · Working paper, SSRN, October 2026
Documents the AI build-out, in which investment in AI-related sectors and infrastructure rose to over 5% of GDP after 2022. Analyzes the cycle in comparison to earlier technology booms and reviews productivity gains across 32 industries.
Economic & Policy Review, 2025 · Semiannual review, Zenodo, 2025
Reads macro and cross-asset signals through the Treasury term premium, using the post-WWII period as its historical comparison. Examines the national debt and the bet on AI-driven growth that appears to be priced into markets.
Education

- University of California, BerkeleyB.A., Economics, with Highest DistinctionProbability & Statistics, Econometrics, Advanced Micro, Advanced Macro, Game Theory & Behavioral Economics, World Economic History, Financial Economics, Demography. Favorite elective: Art History
- University of California, Los AngelesJ.D., Dean’s ScholarGovernance Responses to Transformative Societal Trends, Legislation & Regulation