Brightcon 2026, hackathon and courses

Highlights
1. AI integration in LCA introduces both new uncertainties and undetectable errors that differ fundamentally from traditional LCA uncertainty
2. A clear distinction is drawn between AI-induced uncertainty (stochastic, plausible outputs) and AI-induced error (hallucinated, factually wrong outputs)
3. A stage-based framework is proposed mapping AI-induced uncertainty and errors onto all four LCA phases
4. Unlike traditional LCA errors, AI-induced errors accumulate across stages, making fully automated pipelines inherently prone to error propagation
5. The framework provides a foundation for developing verification and quality control standards in AI-assisted LCA

Context
Life Cycle Assessment (LCA) involves uncertainty across four phases: goal and scope definition, life cycle inventory (LCI), life cycle impact assessment (LCIA), and interpretation. Uncertainty in LCA is commonly classified into model uncertainty, parameter uncertainty, and scenario uncertainty. Various methods have been developed to characterize and propagate these uncertainties, including pedigree matrix for data quality assessment and Monte Carlo simulation for uncertainty propagation.

While integrating artificial intelligence (AI) into the LCA process can reduce certain human-related uncertainties and enhance transparency, it also introduces new challenges in uncertainty characterization. The effect of AI integration depends on how AI-supported LCA systems are designed and applied. In this context, two distinct issues must be differentiated: uncertainty, where AI fills knowledge gaps and produces different but equally plausible models due to its stochastic nature; and error, where AI generates outputs that are simply wrong, such as hallucinated emission factors or fabricated characterization factors, yet remain undetectable without independent verification. Instead of introducing new types of uncertainty, AI can amplify the existing uncertainty in LCA, or introduce errors, such as generating fabricated emission factors that inflate data uncertainty. Unlike traditional uncertainties in LCA, which are often traceable and manageable, AI-induced uncertainty and errors are more difficult to detect due to the nature of AI algorithms. In addition, the traditional errors in LCA are detectable through expert review, AI-induced errors accumulate across four LCA stages, making fully automated LCA pipelines inherently prone to error propagation. Therefore, there is a need to systematically identify and define the AI-induced uncertainty and errors within the LCA context.

Approach
This study proposes a stage-based framework characterizing both AI-induced uncertainty and errors in AI-assisted LCA, mapped onto the four stages of LCA. In goal and scope definition, AI introduces uncertainty through misinterpretation of research intent, ambiguous functional unit definition, and opaque system boundary decisions. In life cycle inventory analysis, primary uncertainty sources include hallucinated inventory data, erroneous foreground system construction, and incorrect background database linkage. In life cycle impact assessment, uncertainty arises from the misassignment of impact categories and the selection of unverifiable characterization factors. In the interpretation stage, AI-generated conclusions lack traceable reasoning, compromising the reliability of hotspot identification and the overall credibility of the assessment.

Brightcon Spirit
This session invites the Brightway and broader LCA community to critically examine the reliability of AI-assisted LCA workflows. By presenting an open framework for uncertainty and error characterization, we aim to stimulate discussion on community standards for verification, quality control, and responsible AI integration in LCA practice.