Evaluating Offline AI Assistants in Analog Missions: Implications for Astronaut Trust, Training, and Operational Readiness

Pilot study presented at the AnalogAstro Conference 2025, Analog Astronaut Training Center (AATC), Poland. The study was conducted during the ASTRAEA-103 analog mission, August 2025. This work addresses the safe operational use of AI in isolated environments, where crews depend on AI assistance without ground support. Analog missions provide a safe yet realistic environment to test how astronauts will collaborate with artificial intelligence during future Moon and Mars expeditions. This pilot study evaluated crew interaction with three locally hosted language models of different sizes (Gemma-3 with 4B parameters, Llama-3.1 with 8B, and DeepSeek-R1 with 14B) across operational tasks in an isolated analog habitat. Using NASA-TLX, System Usability Scale, and Trust-in-Automation metrics (N = 3 participants, 20 task datapoints), results revealed distinct user archetypes: one prioritized speed and clarity (Llama-8B), another favored friendliness and low latency (Gemma-4B), and a third preferred slow but verifiable reasoning (DeepSeek-14B). Although based on a small analog crew, these early patterns highlight consistent human-AI interaction tendencies that can guide future astronaut training. Checklist-style AI outputs reduced workload and frustration, while domain-drift episodes briefly eroded trust without destroying it. Beyond usability, findings underscore a critical operational insight: AI literacy must become an integral part of astronaut training. Crews must understand how a local, offline AI differs from cloud-connected systems, including its computational limits, response behavior, and reliability in isolation, and learn prompt engineering to prevent or recover from hallucinations. Even small, locally run models can support mission tasks effectively when precisely prompted, but their apparent confidence and compliance can mask subtle inaccuracies with potentially hazardous consequences. Therefore, critical evaluation and cross-checking of AI responses must become standard astronaut practice. As AI capabilities advance through onboard hardware acceleration and increased computational power, future missions will benefit not only from stronger models but, more importantly, from better-trained and more critically aware humans. Embedding AI literacy and verification protocols in analog missions and astronaut education is a key step toward human readiness for autonomous deep-space operations. This pilot's observations directly motivate a follow-on study, currently in design, that will examine crew verification behavior, trust calibration over repeated interactions, and safeguards against overreliance with a larger participant group. This record preserves the conference presentation as delivered.