Evaluating Trustworthiness of LLMs for Contrail Detection via Ground-Based Cameras

Authors: Aizierjiang Aiersilan

Abstract

This project investigates the trustworthiness of frontier large language models (LLMs), specifically Gemini 3.1 Pro and Claude Opus 4.6, when deployed as research assistants within a real scientific workflow. Using contrail identification from ground-based visible cameras as an entry point, a domain relevant to aviation's climate impact where condensation trails are estimated to account for roughly 35% of aviation's total climate forcing, the study structures the workflow into three stages: a brief literature review generated by Gemini 3.1 Pro with mandatory verbatim source quotations for every factual claim; an experimental design proposed by Gemini 3.1 Pro for differentiating real contrails from optical glare in whole-sky imagery, including neural architecture selection and ADS-B data fusion; and an adversarial cross-model audit in which Claude Opus 4.6 independently reviews that design to surface technical pitfalls, physical impossibilities, and architectural biases. The work probes a central question in trustworthy AI, namely whether LLMs can serve as reliable collaborators in scientific research, and what safeguards are necessary to prevent confidently stated but technically flawed outputs from entering the research pipeline.

Bibliographic Reference

@misc{aiersilan2026contrail,
  title={Evaluating Trustworthiness of LLMs for Contrail Detection via Ground-Based Cameras},
  author={Aiersilan, Aizierjiang},
  year={2026},
  note={Research Project, The George Washington University}
}