This paper explores the experience and trust of voice-enabled artificial intelligence (AI) technology with Jamaican users, paying close attention to the intersection of dialect recognition and trust. Drawing on data from a national survey conducted by the Sir Arthur Lewis Institute of Social and Economic Studies (SALISES) at the University of the West Indies, Mona, the study analyzes the responses of individuals who reported both awareness of AI and experience with voice technologies such as Siri, Alexa, and ChatGPT. While many respondents expressed concerns that AI tools fail to accurately recognize their dialect, this experience did not strongly correlate with reduced trust or lower adoption rates. That disconnect raises important questions about user expectations, adaptation, and the normalization of linguistic bias within digital systems. Descriptive statistics and recoding techniques were used to identify patterns across variables like education, trust, and AI interaction.The findings suggest a complex interplay between experience and trust, one that warrants closer investigation in contexts where language is a key barrier to inclusion. This paper puts forward the argument that persistence of voice AI use, despite ongoing issues with dialect misrecognition, highlights not just a technical shortfall, but also a social dynamic shaped by possible necessity, familiarity, and broader digital optimism.
The rise of generative AI in fabricating or altering audio poses a significant risk to public confidence in what has traditionally been a reliable medium, amplifying the spread of synthetic content to an unprecedented degree. Safeguarding the integrity of media requires research dedicated to developing dependable methods for differentiating between authentic and AI-generated content. When combined with legislative measures that drive adoption on social platforms, these detection systems could help mitigate this concerning threat. This project focuses on strengthening such detectors against a specific challenge they may encounter in real-world deployment: data poisoning attacks. Our defense mechanism called Feature Inspection and Learning for Threat Examination and Removal (FILTER), is robust against two common types of data poisoning attacks: targeted label flipping attacks. and backdoor attacks.
Extensive prior research highlights the benefits AI possesses in education, and how it can enhance students’ overall self-efficacy and confidence in a wide range of capabilities. However, the study of AI use leading to improvements in the verbal and written self-efficacy of higher education students has been a topic largely unexplored. Given the widespread adoption of AI tools by college students, and AI’s multifaceted features that can both enhance and harm students’ communication abilities, it is important to determine which AI usage patterns yield the best results for students’ communication skills and minimize the risks of using AI. As such, this study, investigates the different usage patterns of AI by college students, focusing on AI uses that lead to students’ verbal and written self-efficacy. The data was gathered from 60 Penn State Abington students via an online survey with constructs aimed at measuring students’ communication self-efficacy and determining AI usage patterns. Through in-depth analyses of students’ responses, the results indicate that different AI usage patterns yield distinct benefits and that generally more involved approaches that do not create an overreliance on AI lead to students’ perceptions of quality enhancement in their writing and speaking, which can indirectly influence their self-efficacy in their verbal and written communication skills. Moreover, no direct correlation was found between AI use frequency and students’ verbal and written self-efficacy, indicating that constant use alone does not improve student confidence in their communication skills, and suggesting that the way students use AI is more significant than how often they use it. This study also discusses the implications these findings can have in education and suggests directions for future research regarding the creation of AI guides of use and the possible implementation of AI into education.
A prolonged employee absence following an occupational injury can disrupt business operations and incur significant costs related to lost productivity and wages. The ability to anticipate which injury cases will result in extended time away from work allows for effective resource allocation. Although machine learning models can estimate absence duration, their predictions typically lack the precision or interpretability required for effective decision-making. To overcome this, we propose two novel frameworks, rank-ordering and order-classification, which prioritize workers’ compensation claims with long-term absences, enabling businesses to concentrate efforts on the most severe cases. The rank-ordering framework adapts an established ranking technique for absence duration prediction. We first apply a CatBoost regression model to predict the number of days an injured employee remains absent from work. After ordering the predictions, we evaluate how well the ranking captures the longest actual absences. Then, we use the order-classification framework to assign binary labels to the regression outputs based on different absence duration thresholds and cutoff points. Lastly, we compare its performance against a baseline CatBoost classifier in three scenarios. Our findings show that while the CatBoost classifier performed best under balanced conditions, the order-classification model produced equal or stronger results in imbalanced scenarios. To the best of our knowledge, our study presents the first comparison between our order-classification framework and a traditional classifier for identifying and prioritizing long-duration injury absences.
As crime rates rise and citizens become increasingly reluctant to report incidents due to fear of retaliation, there is growing demand for technological interventions. This need is especially urgent in developing countries like Jamaica, where public safety directly impacts tourism and economic stability. In response, this study proposes a human-in-the-loop framework that leverages the strengths of both human and machine capabilities in gunshot detection and reporting. Using a sample of Jamaican citizens, the study compares the accuracy of human versus machine classification of gunshot sounds and examines citizens’ willingness to report such incidents. The proposed framework addresses technological limitations such as miscalculation, as well as social barriers like fear and distrust, offering a balanced solution for environments where a single strategy may be ineffective or unavailable.
