Abstract Large language models (LLMs) frequently generate plausible but factually inaccurate outputs, commonly known as hallucinations. Moving beyond treating hallucination as a simple binary factuality problem, this study evaluates it as a multidimensional risk encompassing generation frequency, confidence accuracy mismatch, detectability, and downstream trust impact. Analyzing a comprehensive dataset of 10,000 interactions across six frontier AI models and platforms, we assessed hallucination prevalence, epistemic calibration, and... Large language models (LLMs) frequently generate plausible but factually inaccurate outputs, commonly known as hallucinations. Moving beyond treating hallucination as a simple binary factuality problem, this study evaluates it as a multidimensional risk encompassing generation frequency, confidence accuracy mismatch, detectability, and downstream trust impact. Analyzing a comprehensive dataset of 10,000 interactions across six frontier AI models and platforms, we assessed hallucination prevalence, epistemic calibration, and human consequences. Results reveal distinct variations in hallucination rates, with Llama and Mistral exhibiting the highest frequencies. Crucially, we identified a pronounced confidence accuracy mismatch across systems; notably, Mistral presented the most concerning risk configuration by combining high hallucination rates, maximum AI confidence, and the lowest detection rates. Conversely, GPT-5 demonstrated the most favorable factual accuracy profile. Cross-tool analysis demonstrated that hallucination occurrence, rather than specific tool identity, serves as the primary driver of severe trust erosion. Specifically, the mere presence of an error caused a massive trust penalty that subsequent detection alone could not resolve. These findings proved that AI hallucination is a systemic, cross-platform reliability challenge. Mitigating its impact requires integrated strategies addressing confidence calibration, proactive detection mechanisms, and user interface designs that foster appropriate skepticism, rather than relying solely on reducing raw factual error rates. Read More Read Less

Abstract Despite the rapid advancement of Large Language Models (LLMs), hallucinations remain a critical barrier to their trustworthy deployment in high stakes domains. This study examines how prompt clarity influences LLM response quality and trustworthiness, focusing on the mediating role of hallucination risk. Utilizing a synthetic benchmark of 15,000 prompt response pairs spanning five languages, multiple domains, four difficulty levels, and five model families, we... Despite the rapid advancement of Large Language Models (LLMs), hallucinations remain a critical barrier to their trustworthy deployment in high stakes domains. This study examines how prompt clarity influences LLM response quality and trustworthiness, focusing on the mediating role of hallucination risk. Utilizing a synthetic benchmark of 15,000 prompt response pairs spanning five languages, multiple domains, four difficulty levels, and five model families, we employed causal style mediation analysis, fairness audits, and routing policy simulations. Results indicate that hallucination risk is the dominant mediator, accounting for approximately 66% to 71% of the total effect of prompt clarity on response quality. While LLM performance demonstrated strong cross-lingual and cross-domain robustness, trustworthiness measured via a novel Trustworthy Response Index (TRI) declined significantly as task difficulty increased. Furthermore, model family significantly moderated the mediation pathway; Reasoning-LLMs showed greater resilience to prompt ambiguity than other families. System level simulations revealed that intelligent model routing, particularly strategies prioritizing Reasoning LLMs, yields substantial gains in overall quality and TRI. These findings underscore that prompt engineering and model aware routing are interconnected determinants of reliable AI, necessitating integrated strategies for trustworthy LLM deployment. Read More Read Less

Abstract Automated construction monitoring increasingly relies on geospatial metadata and machine learning to track urban development, yet the true generalizability and reliability of metadata driven models remain underexplored. This study evaluates the predictive capabilities of geospatial metadata including polygon geometry, temporal duration, and annotation attributes using Random Forest, Extra Trees, and Logistic Regression models on a large scale GeoJSON dataset of construction sites. We conducted... Automated construction monitoring increasingly relies on geospatial metadata and machine learning to track urban development, yet the true generalizability and reliability of metadata driven models remain underexplored. This study evaluates the predictive capabilities of geospatial metadata including polygon geometry, temporal duration, and annotation attributes using Random Forest, Extra Trees, and Logistic Regression models on a large scale GeoJSON dataset of construction sites. We conducted binary (site presence) and multiclass (construction type) classification tasks, revealing a stark contrast: while Random Forest achieved near perfect binary accuracy (99.9%), multiclass construction type recognition remained highly ambiguous (macro-F1 = 0.610). Crucially, permutation feature importance analysis exposed that the binary predictive signal was almost entirely concentrated in a single annotation derived variable (n_features), with all physical and geometric features registering zero importance. This indicates that the near perfect performance is an artifact of the annotation workflow and potential target leakage rather than genuine physical site detection. We conclude that metadata only models are insufficient for fine grained semantic typing and issue a critical methodological warning for the GeoAI community: rigorous spatial cross validation and annotation leakage audits are mandatory to prevent spatially and procedurally inflated performance metrics. Read More Read Less