Artificial Intelligence is increasingly finding its way into satellite operations. Large Language Models (LLMs) promise significant efficiency gains by assisting operators with log analysis, anomaly investigation, technical documentation and operational procedures. At the same time, their adoption raises important questions regarding reliability, trustworthiness and security. To better understand their potential and associated risks, the LINK2AI project investigated the use of a Retrieval-Augmented Generation (RAG) assistant in a realistic satellite operations environment.

A dedicated AI assistant was implemented and connected to a mission-specific knowledge corpus. The corpus included approved technical documentation, system manuals, operational procedures, mission logs, event reports and other relevant mission data. By combining large language models with retrieval mechanisms, the assistant was able to answer operator questions based on authoritative mission knowledge rather than relying solely on its pre-trained capabilities.

The evaluation demonstrated significant potential for supporting operational activities. The AI assistant was capable of navigating large amounts of technical information, answering questions on operational procedures and assisting in the interpretation of mission data.

However, the assessment also showed that a RAG-based approach cannot automatically be considered trustworthy in safety- or mission-critical environments. Despite having access to curated and approved information sources, the system occasionally produced inaccurate responses, followed misleading information contained within the available data, or generated outputs that required expert validation. Furthermore, dedicated security assessments demonstrated that the system could be influenced through adversarial inputs and manipulated knowledge sources.

These findings highlight an important lesson for future AI-enabled space systems: access to trusted information alone does not guarantee trustworthy behaviour. Human oversight, security testing, validation mechanisms and continuous monitoring remain essential when deploying AI assistants in operational satellite environments.