University of the Potomac

The environmental impacts of generative artificial intelligence are increasingly measurable, but ethical guidance remains difficult to operationalize at the point of model selection. This article argues that developers and deployers have a prima facie duty to avoid materially greater ecological burdens when a lower-impact system can satisfy the legitimate purpose of a task, unless the additional capability provides a proportionate and context-specific benefit. We operationalize this claim as an Ecological Capability Proportionality Test and use a public infrastructure-aware dataset of large-language-model inference to illustrate its application. Among 42 model configurations with usable capability and energy estimates, capability was positively associated with estimated energy (Spearman ρ = 0.588, p < 0.001; bootstrap 95% CI 0.281–0.812). Eight configurations formed a capability-energy Pareto frontier. At the upper frontier, increasing the capability index from 43 to 45 increased estimated energy per long prompt from 9.79 to 21.31 Wh, while the Environmental Cost of Marginal Capability rose to 5.76 Wh per index point. The ethical significance is not that a universal capability threshold can be inferred from these data, but that marginal ecological premiums should be disclosed and justified. The article proposes a governance procedure combining sufficiency thresholds, least-burden model selection, escalation justification, distributive scrutiny, and rebound-aware monitoring.

Diabetes mellitus is a chronic metabolic disease that is a significant global public health concern. Complications may potentially be avoided or postponed with early diabetes diagnosis and treatment. The development of ML and DL has created new possibilities in the analysis of clinical data, allowing to identify the pattern concealed in them and increase the accuracy of diagnosis. This research proposes a diabetes prediction model based on the PIMA Indian Diabetes Dataset through the application of the Machine Learning (ML) and Deep Learning (DL) methods. Random Forest (RF) and Long Short-term Memory (LSTM) are two high-performance models that were implemented to extract nonlinear and temporal relationships in the data. Experimental testing showed that RF had 97.54% accuracy and 96.32 F1-score, whereas LSTM had a steady 98.85% accuracy and 98.20 F1-score. The proposed RF and LSTM models showed a clear superiority when compared to more traditional models, like Naive Bayes, Decision Trees, AdaBoost, and SVM (72-79% accuracy), to the more advanced models, like ANN, CNN-LSTM, and DNN (89-93% accuracy). The proposed framework demonstrates that combining ensemble and sequential learning offers a scalable and accurate solution for early diabetes detection, representing a key contribution to clinical decision support.

Test-time adaptation (TTA) can improve predictions after deployment, but the update itself can become a new failure mode when unlabeled streams are transient, contaminated, imbalanced, or recurrent. We introduce TRUST-ADAPT, a risk-gated TTA algorithm that treats adaptation as a reversible intervention rather than a default response. A frozen source model converts confidence nonconformity into split-conformal p-values; a batch-level Hoeffding gate triggers learning only when the fraction of anomalous samples exceeds a finite-sample threshold. Triggered batches are further filtered by predictive uncertainty, adapted with pseudo-label and entropy objectives regularized by an immutable source anchor and selective class-balanced memory, and accepted only if an anchor-risk validation check passes; otherwise the parameters are rolled back. Under super-uniform conformal p-values and conditional within-batch independence, the gate controls the probability of a false adaptation trigger at a user-specified level. We implement the complete procedure in PyTorch and execute a locked five-seed controlled benchmark on streaming digit classification with abrupt, recurring, and 15% contaminated shifts. Across 15 seed-scenario pairs, TRUST-ADAPT improves mean streaming accuracy over the frozen source by 3.20 percentage points (Holm-adjusted Wilcoxon p=0.0039) while using 49.8 update batches on average versus 90.0 for TENT-style adaptation. The method is particularly competitive under contamination, but recurring shifts remain a limitation. All reported values are generated by the accompanying executable code; no unexecuted numerical result is claimed.

The digital revolution in procurement has turned out to be an essential facilitator towards operational effectiveness, cost minimization and strategic decision-making in contemporary businesses. The conventional procurement systems are usually associated with manual operations, disintegrated systems of data, and weak analytical tools, which create inefficiencies and slow down the decision-making process. In order to overcome these issues, this paper suggests a smart automation-based system of digital procurement change incorporating such innovative technologies as artificial intelligence (AI), machine learning (ML), robotic process automation (RPA), and data analytics. The suggested solution will allow complete automation of the procurement processes, such as the selection of suppliers, demand forecasting, contract management, and real-time decision support. The framework improves procurement visibility, mitigation of risks, and the evaluation of supplier performance with the help of predictive analytics and intelligent agents. Moreover, the system has adaptive learning systems so that it keeps enhancing the accuracy of decisions and operational robustness in dynamic market conditions. The results of experimental assessment indicate that there are huge differences in the time of procurement cycle, saved costs, and efficiency of the decisionmaking process, which are better than the traditional approaches. The suggested model offers an upscaled and customizable approach to companies that intend to shift to datadriven, intelligent procurement ecosystems, which would allow building sustainable and competitive businesses.

Although many organizations publish high-level artificial intelligence principles, most lack a coherent framework to evaluate whether their communication practices are operationally ready for responsible use. Grounded in the NIST Generative AI Profile (NIST AI 600-1), algorithmic work scholarship, and communication-centered accountability, this study develops and validates a multidimensional model of organizational readiness for responsible AI communication.Using a multi-site empirical meta-synthesis and cross-sectional survey data (N = 1,420), we reveal a severe operationalization gap: while 73.9% of organizations exhibit symbolic policy presence, only 37.4% demonstrate functional operational readiness. Major vulnerabilities emerge in plain language and usability (28.4%), multilingual inclusion (22.1%), and monitoring routines (25.3%).Structural Equation Modeling shows that procedural verification and Human-in-the-Loop review routines completely mediate the link between enterprise AI capability and recipient trust (beta = .62, p < .001). We propose an actionable 4-tier human oversight architecture and an agenda for institutional AI governance.