A Systems Approach to Redesigning Human-AI Interactions in Gaslift Operations for Upstream Oil and Gas

Abstract

Artificial intelligence tools are increasingly used to support engineering decisions, yet their value depends on whether humans trust and act on their recommendations. This thesis investigates how LLM response style and user expertise influence trust in a technical context. Gas‑lift troubleshooting was used as a domain case. One hundred thirty‑one participants across five expertise levels were randomly assigned to evaluate a diagnostic scenario answered by one of four curated LLM personas that varied in tone, confidence, and accuracy. After each of five steps, participants rated trust and explained their rationale. We analyzed trust ratings using one‑ and two‑way ANOVA with post‑hoc tests and complemented the analysis with text mining and exploratory factor analysis of perceived trust drivers. Trust varied by LLM persona and by participant expertise, with no evidence of interaction effects. Beyond accuracy, three attributes consistently associated with higher trust were clarity, source citation, and presentation of a confidence score. Text analysis showed that experts emphasized technical validity, while less‑experienced users prioritized clarity and information sufficiency. Based on these results, the thesis proposes a conceptual redesign for human-AI interaction that adapts explanations to expertise, discloses confidence, and links recommendations to supporting sources. The framework provides measures of operational effectiveness to evaluate adoption and impact in production settings. Findings suggest that tailoring communication rather than using a one‑size‑fits‑all style can improve trust and decision quality in gas‑lift operations and similar engineering workflows.