H₃ is Combustible: The HHH Trilemma in AI-Assisted Evaluation
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Commercial large language models are aligned using the HHH framework: Helpful, Honest, and Harmless. This paper demonstrates, through controlled experimental data, that these three optimization targets form an unstable trilemma when applied simultaneously to human evaluation tasks. Using a single candidate's resume evaluated across five iterative rounds by three frontier models (Claude, GPT-4o, and Perplexity), we show that each model prioritizes a different H, producing divergent hiring decisions on identical data. We further demonstrate that multi-model consensus does not resolve the instability—it compounds it. The resulting instability mirrors molecular physics: H₂ forms stable bonds; H₃ does not exist in nature because the geometry cannot support it. We propose that the missing stabilizing element—the 'oxygen'—is a human governor that contextually prioritizes which optimization target applies when. We introduce H₆ as the equilibrium configuration: three resolved H₂ pairs, contextually governed. Companion paper to AI Governance Failure Analysis: When Models Evaluate People (DOI: 10.5281/zenodo.19350383).
Fields of Study & Subjects
Ethics and Social Impacts of AI
AI and HR Technologies
Explainable Artificial Intelligence (XAI)
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 1 April 2026
- Language
- EN