Mohsen Amoei

Background: AI research increasingly depends on energy-intensive computation, yet energy use, greenhouse gas emissions, hardware life cycle burdens, and water consumption are rarely reported in a standardized way. This lack of reproducible environmental accounting limits comparisons across studies and obscures the trade-offs among model performance, infrastructure choices, carbon intensity, and cooling water demand. Objective: This study aimed to develop and describe an open-access, life cycle–informed AI Environmental Footprint Calculator and propose a minimum reporting dataset for transparent environmental disclosure in AI research. Methods: We developed a browser-based calculator that combines operational energy, regional grid carbon intensity, power usage effectiveness (PUE), water usage effectiveness, hardware embodied emissions, and workload-specific metrics for training and inference. The framework was evaluated in 3 representative scenarios: a single–graphics processing unit laboratory fine-tuning task, a midsized academic cluster workload, and a large-scale industrial training cycle. Outputs were compared with those of established tools to identify how boundary choices and parameter assumptions affect emission estimates. Results: Across the scenarios, the inclusion of PUE and hardware life cycle allocation increased reported emissions compared with operational-only estimates. In the small laboratory scenario, optimization reduced total emissions from approximately 0.07 kg carbon dioxide equivalents (CO ₂ e) to 0.05 kg CO ₂ e and improved the proposed label from C to B. In the midsized cluster scenario, carbon-aware scheduling reduced emissions from approximately 665 kg CO ₂ e/month to 450 kg CO ₂ e/month—a 32% reduction. In the large-scale scenario, shifting to renewable-backed, lower-PUE infrastructure reduced operational emissions by approximately 79% while increasing the importance of water-carbon trade-off reporting. Conclusions: The calculator provides a practical and transparent method for reporting AI environmental footprints using auditable parameters and publication-ready outputs. Routine disclosure of energy use, emissions, water use, hardware assumptions, and regional context can improve reproducibility and support more equitable and sustainable AI research evaluation.

UNSTRUCTURED Health systems are deploying large language models (LLMs) into clinical workflows faster than they are deciding how to integrate them safely. LLMs are often added as a safety net, on the assumption that a second automated check can only reduce error. But a second reasoner reduces error only insofar as it fails differently from the first. One whose failures coincide with the clinician’s adds little safety and can turn agreement into false reassurance. What matters is therefore not benchmark accuracy, but how the model’s failures relate to the clinician’s; we call this property complementary vulnerabilities. Human and LLM clinical reasoning break down through different mechanisms, so their weak points need not line up. These divergent mechanisms are what can make a clinician-LLM pair safer than either alone, but the independence they create is fragile, eroding when clinicians defer to the model or the model conforms to the clinician. A reviewer added without knowing where the model is weak, or a model added without preserving the clinician’s independent judgment, cannot be counted on to capture that safety gain. In this paper, we translate this principle into a predeployment paired-error audit that examines whether divergent failure patterns produce non-overlapping errors locally, and a governance model that preserves that independence over time. We reframe clinical artificial intelligence (AI) deployment from how accurate a model is to how its failures are structured relative to the clinician’s.