Figure - available from: Applied Sciences
This content is subject to copyright.
Source publication
Bias could be defined as the tendency to be in favor or against a person or a group, thus promoting unfairness. In computer science, bias is called algorithmic or artificial intelligence (i.e., AI) and can be described as the tendency to showcase recurrent errors in a computer system, which result in “unfair” outcomes. Bias in the “outside world” a...
Similar publications
Software defect prediction (SDP) ensures software system quality and reliability. However, class imbalance poses a significant challenge, hindering traditional machine learning (ML) techniques. Biased models often result in impacting defect prediction accuracy. We provide a novel approach that uses reinforcement learning (RL) algorithms to overcome...
The widespread adoption of machine learning systems has raised critical concerns about fairness and bias, making mitigating harmful biases essential for AI development. In this paper, we investigate the relationship between fairness improvement and the removal of harmful biases in neural networks applied to computer vision tasks. First, we introduc...
We address the challenges of Byzantine-robust training in asynchronous distributed machine learning systems, aiming to enhance efficiency amid massive parallelization and heterogeneous computing resources. Asynchronous systems, marked by independently operating workers and intermittent updates, uniquely struggle with maintaining integrity against B...
Deep Learning models tend to learn correlations of patterns on huge datasets. The bigger these systems are, the more complex are the phenomena they can detect, and the more data they need for this. The use of Artificial Intelligence (AI) is becoming increasingly ubiquitous in our society, and its impact is growing everyday. The promises it holds st...
Unsupervised learning has become an essential building block of artifical intelligence systems. The representations it produces, for example, in foundation models, are critical to a wide variety of downstream applications. It is therefore important to carefully examine unsupervised models to ensure not only that they produce accurate predictions on...
Citations
... This study uses 99 fertile and 66 barren adakite analyses, reflecting the availability of high-quality data that met the criteria for inclusion. While the dataset is imbalanced, we did not choose oversampling or class weighting [113], as these techniques can introduce challenges, including the risk of overfitting [114] or biasing the model towards the minority class [115]. Oversampling artificially inflates the minority class, while class weighting can cause the model to overemphasize the minority class, leading to skewed predictions [113][114][115]. ...
... While the dataset is imbalanced, we did not choose oversampling or class weighting [113], as these techniques can introduce challenges, including the risk of overfitting [114] or biasing the model towards the minority class [115]. Oversampling artificially inflates the minority class, while class weighting can cause the model to overemphasize the minority class, leading to skewed predictions [113][114][115]. Instead, the performance of the models was optimized as much as possible and evaluated using metrics such as the area under the curve (AUC) [116] and Matthews correlation coefficient (MCC) [117], which are less sensitive to class imbalance and provide a more balanced measure of model performance. ...
This study applies machine learning (ML) techniques to classify fertile [for porphyry Cu and (or) Au systems] and barren adakites using geochemical data from New Brunswick, Canada. It emphasizes that not all intrusive units, including adakites, are inherently fertile and should not be directly used as the heat source evidence layer in mineral prospectivity mapping without prior analysis. Adakites play a crucial role in mineral exploration by helping distinguish between fertile and barren intrusive units, which significantly influence ore-forming processes. A dataset of 99 fertile and 66 barren adakites was analyzed using seven ML models: support vector machine (SVM), neural network, random forest (RF), decision tree, AdaBoost, gradient boosting, and logistic regression. These models were applied to classify 829 adakite samples from around the world into fertile and barren categories, with performance evaluated using area under the curve (AUC), classification accuracy, F1 score, precision, recall, and Matthews correlation coefficient (MCC). SVM achieved the highest performance (AUC = 0.91), followed by gradient boosting (0.90) and RF (0.89). For model validation, 160 globally recognized fertile adakites were selected from the dataset based on well-documented fertility characteristics. Among the tested models, SVM demonstrated the highest classification accuracy (93.75%), underscoring its effectiveness in distinguishing fertile from barren adakites for mineral prospectivity mapping. Statistical analysis and feature selection identified middle rare earth elements (REEs), including Gd and Dy, with Hf, as key indicators of fertility. A comprehensive analysis of 1596 scatter plots, generated from 57 geochemical variables, was conducted using linear discriminant analysis (LDA) to determine the most effective variable pairs for distinguishing fertile and barren adakites. The most informative scatter plots featured element vs. element combinations (e.g., Ga vs. Dy, Ga vs. Gd, and Pr vs. Gd), followed by element vs. major oxide (e.g., Fe2O3T vs. Gd and Al2O3 vs. Hf) and ratio vs. element (e.g., La/Sm vs. Gd, Rb/Sr vs. Hf) plots, whereas major oxide vs. major oxide, ratio vs. ratio, and major oxide vs. ratio plots had limited discriminatory power.
... However, as noted in the opening statistic of this article, it is unknown whether the clinicians and expert raters used to train AI models themselves make accurate decisions. Authors in AI ethics commonly discuss the risk of perpetuating human biases in training data (Mavrogiorgos et al., 2024). The extent to which this was addressed in the reviewed articles is unclear. ...
Product headlines have brought artificial intelligence (AI) to the forefront of many industries, including applied behavior analysis (ABA). However, the products behind the headlines may not always be backed by empirical support for claimed benefits. This concise review identifies AI use cases currently with empirical support to aid clinical decision-making in ABA. Six articles suggest weak support for AI-driven systems to supplement data analysis, recommend goals, and recommend comprehensive vs. focused ABA. Replication, increased sample sizes, and increased AI system utility are sorely needed.
... One challenge frequently mentioned in many discussions about the use of AI is that of algorithmic bias (Mavrogiorgos et al., 2024). AI-systems learn from training data. ...
Political decision-making is often riddled with uncertainties, largely due to the complexities and fluid nature of contemporary societies, which make it difficult to predict the consequences of political decisions. Despite these challenges, political leaders cannot shy away from decision-making, even when faced with overwhelming uncertainties. Thankfully, there are tools that can help them manage these uncertainties and support their decisions. Among these tools, Artificial Intelligence (AI) has recently emerged. AI-systems promise to efficiently analyze complex situations, pinpoint critical factors, and thus reduce some of the prevailing uncertainties. Furthermore, some of them have the power to carry out in-depth simulations with varying parameters, predicting the consequences of various political decisions, and thereby providing new certainties. With these capabilities, AI-systems prove to be a valuable tool for supporting political decision-making. However, using such technologies for certainty purposes in political decision-making contexts also presents several challenges—and if these challenges are not addressed, the integration of AI in political decision-making could lead to adverse consequences. This paper seeks to identify these challenges through analyses of existing literature, conceptual considerations, and political-ethical-philosophical reasoning. The aim is to pave the way for proactively addressing these issues, facilitating the responsible use of AI for managing uncertainty and supporting political decision-making. The key challenges identified and discussed in this paper include: (1) potential algorithmic biases, (2) false illusions of certainty, (3) presumptions that there is no alternative to AI proposals, which can quickly lead to technocratic scenarios, and (4) concerns regarding human control.
... The rapid growth of natural language processing (NLP) has led to signi cant advancements in text classi cation tasks. However, machine learning models often exhibit biases due to imbalanced datasets, biased annotations or algorithmic limitations and a comprehensive analysis are discussed in [1]. Such biases can lead to unfair outcomes, where speci c groups or contexts are systematically disadvantaged. ...
The rapid growth of natural language processing (NLP) applications has highlighted concerns about fairness and bias in text classi cation models. Despite signi cant advancements, the evaluation of bias and fairness in Bangla text classi cation remains underexplored. This study investigates model bias in Bangla text classi cation models, focusing on key fairness metrics such as Demographic Parity, Equalized Odds, and Accuracy Parity. We analyze the performance of widely used models, including Naive Bayes (NB), Support Vector Machine (SVM), Random Forest (RF), LSTM and Bangla-BERT, on a comprehensive dataset. The results reveal disparities in fairness across models, with Bangla-BERT achieving the highest fairness scores but still exhibiting measurable bias. To address this, we conduct an error analysis, highlighting the prevalence of bias-induced misclassi cations across sensitive attributes. Additionally, we propose actionable recommendations to enhance fairness in Bangla NLP models, bridging gaps in ethical AI for low-resource languages. Our ndings provide valuable insights for developing more equitable Bangla text classi cation systems and emphasize the need for fairness-aware methodologies in future NLP research.
... To address this issue, the authors employed an association algorithm to examine keyword relationships and engaged in team discussions during the final stage to reduce subjectivity in interpretation. Although the abstract and conclusion aim to summarize the key content and findings of each artifact, there is a possibility of interpretation bias (Mavrogiorgos et al., 2024). For future research, consider expanding the analysis to include other sections such as the introduction or methodology. ...
Purpose
The study aims to explore how knowledge creation helps accounting scholars in developing accounting-entry systems (AeS) in their research domain.
Design/methodology/approach
The authors conducted a systematic literature review to gather 474 artifacts from 1923 to 2023. Data mining techniques created a knowledge dashboard to achieve the study’s objectives.
Findings
This study confirms that accounting-entry systems (AeS) are integrating with technological structures, requiring the transfer of accounting knowledge. Schools play a crucial role in enhancing AeS knowledge. This involves transferring knowledge through research and method development, which is applied in professional accounting practices. Additionally, the study reveals that artifacts often highlight the environment, art, and mathematics aspects essential for developing AeS knowledge within the triad construct. Afterward, technological aspects became increasingly important in the digitalization era, expanding the triad construct.
Research limitations/implications
Theoretical implications indicate that the use of knowledge dashboards in foresight studies can lead to innovation in accounting. Practical implications suggest that accounting professionals and knowledge engineers can collaborate to develop technology-driven accounting solutions that are compatible with contemporary information systems.
Practical implications
The study developed a data mining protocol to analyze the shift and evolution of knowledge. This protocol can be used and assessed in future research. It introduces a knowledge dashboard to aid in forecasting accounting and information systems research topics. The AeS journey outlines key events to elucidate the knowledge transition. The AeS evolution demonstrates how the triad construct emphasizes the need for a substantive shift, leading to knowledge creation.
Originality/value
This study employs knowledge dashboard visualizations to explore the progress of AeS through social reproduction events and evolution values. The AeS journey illustrates the shift in knowledge, focusing on social reproduction events. The evolution of AeS implicitly demonstrates the change in knowledge substance by creating a dyad and triad to explain knowledge creation.
... First, the preparation of datasets for AI applications introduces the risk of embedding biases during the data curation and pre-processing stages and increases the number of "customized" input data products adapted to AI applications whose quality might not be easily verified. These biases, whether intentional or unintentional, can influence the outcomes of ML models by shaping the patterns that the algorithms are designed to detect [Masuda et al., 2024;Mavrogiorgos et al., 2024]. The key issue here is the quality assessment of input data impacting the reliability of AI results. ...
Open Science is the paradigm driving the sharing of research data worldwide. It includes the ambition to make FAIR (Findable, Accessible, Interoperable and Re-usable) data sharing the default. FAIR guiding principles for research data have been recently proposed to scientific communities as the new horizon for sharing data. The FAIR principles create the conditions to foster data sharing and improve data stewardship, provided that several legislative, organizational, and ethical issues are addressed. In this paper, we aim to discuss the ethical dimension of sharing solid Earth science data. Earth scientists have a long-lasting tradition in data acquisition, quality control, and standardization, being the key actors in feeding and implementing metadata and services for qualification, storage, and accessibility. Pan-European Research infrastructures like EPOS (European Plate Observing System) involve scientific communities and research organizations federating facilities and resources to ensure data management and interoperability through e-science innovation. After introducing the ethical issues associated with the protection of personal data, intellectual property rights, and data misuse, we will focus on the impartiality for public good. This opens a 1 new horizon to the ethical dimension of open access to research data, going well beyond research integrity. This assumes an outstanding relevance when referring to solid Earth science data since they also concern natural and anthropogenic hazards and risk communication relying on sharing scientific information with different stakeholders. Although we present a specific perspective for solid Earth science, we believe that the addressed ethical dimension is relevant for environmental science in general.
... Furthermore, software relying on stochastic processes for equitable distribution of outcomes can introduce bias if the random number generation mechanism deviates from true randomness, potentially skewing selections towards items at the extremities of a list (Greene, 2022). (Mavrogiorgos et al., 2024). ...
... For example, optimization functions that prioritize accuracy over fairness can result in models that perform well on average but poorly for minority groups. Additionally, the choice of regularization methods and hyperparameters can influence the model's behavior, potentially introducing unintended biases (Mavrogiorgos et al., 2024). ...
This research investigates the integration of Large Language Models (LLMs) into aerospace defense systems engineering to automate two critical processes: eliciting requirements through System Theoretic Process Analysis (STPA) and assigning Means of Compliance (MoCs) to aerospace defense systems’ requirements. The motivation lies in addressing the labor-intensive and error-prone nature of traditional methods, which heavily rely on human expertise. The study specifically evaluates the feasibility and performance of LLMs, such as GPT-3.5 and GPT-4, when guided by advanced Prompt Engineering techniques and fine-tuning methodologies. These approaches aim to maintain or surpass the accuracy and quality typically achieved by experts in the field. The problem under investigation is the inefficiency and variability of manual requirements engineering and compliance processes, which are critical in defense aerospace systems due to stringent safety and reliability demands. Using a hypothetical Unmanned Combat Air Vehicle (UCAV) as a case study, the study situates the research in the context of the Brazilian Air Force (FAB), where these challenges are particularly acute. The methodology involves automating Phase 1 of STPA through tailored prompts to generate system requirements and training a fine-tuned model to assign MoCs accurately. Performance was benchmarked against real-world system data and domain experts’ outputs. The findings highlight that LLMs guided by Prompt Engineering can generate requirements that meet or exceed eight of nine evaluated quality attributes, including testability, completeness, clarity, and modifiability. The fine-tuned ‘gpt-3.5-turbo’ model achieved an 80.18% accuracy in MoC assignments. Finally, with appropriate techniques, it was possible to generate safety assessment reports such as PHAs (Preliminary Hazard Analysis) from the technical documentation of real products. The implications of this research are profound. By streamlining requirements elicitation, MoC assignment, and the generation of engineering reports, LLMs reduce the time, effort, and cost associated with engineering processes while maintaining high standards of rigor and reliability. This work advances academic understanding of LLM applications in safety-critical systems, introduces a scalable and replicable framework for integrating LLMs into engineering workflows, and offers practical tools to the aerospace defense industry.
... While the integration of AI/ML techniques in evidence synthesis has been shown to significantly reduce the time and effort required to complete these projects [34], [35], several challenges impede broader adoption. These include the lack of established guidelines for implementing AI/ML techniques (particularly regarding defining stopping criteria for screening tasks [36]), ensuring reproducibility of ML practices [37], and addressing biases inherent in algorithmic systems [38]. Inconsistencies in documenting the application of these technologies further complicate evaluations of their effectiveness and applicability [39]. ...
will perform data collection. Kristen will conduct the analysis and draft the manuscript, while all authors will provide critical feedback to refine the work, including contributions to the analysis discussion and manuscript revisions. 2. INTRODUCTION AND BACKGROUND Evidence synthesis involves systematically aggregating and, in some cases, evaluating research findings to expand their applicability and generate new knowledge [1], [2], [3], [4]. It plays a critical role in advancing scientific knowledge by building consensus [5], [6], informing evidence-based practices [7], [8], [9], identifying knowledge gaps [10], [11], and guiding policy [12], [13]. Across disciplines, evidence synthesis methods (e.g., systematic reviews [14], meta-analyses [6], and evidence gap maps [10]) are used to address complex questions and support decision-making. For example, in medicine, evidence synthesis is integral to the development and evaluation of clinical guidelines [7]; in education [15] and environmental policy [13], it also helps to inform best practices. The reproducibility crisis, marked by challenges in replicating research findings across disciplines, has underscored the need for rigorous evidence synthesis methods as a mechanism for identifying sources of variability and promoting the reliability of research [6], [16]. The evidence synthesis process typically involves a series of structured steps, including formulating research questions, conducting systematic searches, screening results for relevance, extracting data from included studies, assessing the quality and potential biases of studies, and analyzing and synthesizing the findings. These projects typically demand significant human effort, often requiring teams of experts months or even years to complete [17], [18]. These efforts are compounded by the exponential growth of research outputs [19], [20], which makes comprehensive synthesis tasks increasingly challenging. Reducing the time and effort required for evidence synthesis, while maintaining methodological rigor, is essential for delivering timely insights to stakeholders and decision-makers. The automation of evidence synthesis tasks has been explored extensively [21], [22], [23], [24], [25]. Initial approaches relied on rule-based methods to improve efficiency, such as optimizing database searches [26] and using text mining to facilitate study selection [27]. The growing
... A bias toward positive findings can distort the perceived effectiveness of AI chatbots. Encouraging the publication of null or negative results is vital for maintaining a balanced understanding [75,138,139]. ...
Anxiety disorders are among the most prevalent mental health conditions globally, causing significant personal and societal burdens. Traditional therapies, while effective, often face barriers such as limited accessibility, high costs, and the stigma associated with seeking mental health care. The emergence of artificial intelligence (AI) chatbots offers a novel solution by providing accessible, cost-effective, and immediate support for individuals experiencing anxiety. This comprehensive review examines the evolution, efficacy, advantages, limitations, challenges, and future perspectives of AI chatbots in the treatment of anxiety disorders. A methodologically rigorous literature search was conducted across multiple databases, focusing on publications from 2010 to 2024 that evaluated AI chatbot interventions targeting anxiety symptoms. Empirical studies demonstrate that AI chatbots can effectively reduce anxiety symptoms by delivering therapeutic interventions like cognitive-behavioral therapy through interactive and personalized dialogues. The advantages include increased accessibility without geographical or temporal limitations, reduced costs, and an anonymity that encourages openness and reduces stigma. However, limitations persist, such as the lack of human empathy, ethical and privacy concerns related to data security, and technical challenges in understanding complex human emotions. The key challenges identified involve enhancing the emotional intelligence of chatbots, integrating them with traditional therapy, and establishing robust ethical frameworks to ensure user safety and data protection. Future research should focus on improving AI capabilities, personalization, cultural adaptation, and user engagement. In conclusion, AI chatbots represent a promising adjunct in treating anxiety disorders, offering scalable interventions that can complement traditional mental health services. Balancing technological innovation with ethical responsibility is crucial to maximize their potential benefits.
... The biases inherent to each modality are well known [Alelyani, 2021, Mavrogiorgos et al., 2024, Mehrabi et al., 2021, Cadene et al., 2019, Liang et al., 2021, Navigli, Roberto and Conia, Simone and Ross, Björn, 2023, Chen et al., 2024. For example, language models are often found to be capable of reinforcing stereotypes [Miller, 2024, Kotek et al., 2023, Manzini et al., 2019, and image classifiers have the potential to enhance misrepresentation [Zhang et al., 2024]. ...
Multimodal machine learning models, such as those that combine text and image modalities, are increasingly used in critical domains including public safety, security, and healthcare. However, these systems inherit biases from their single modalities. This study proposes a systemic framework for analyzing dynamic multimodal bias interactions. Using the MMBias dataset, which encompasses categories prone to bias such as religion, nationality, and sexual orientation, this study adopts a simulation-based heuristic approach to compute bias scores for text-only, image-only, and multimodal embeddings. A framework is developed to classify bias interactions as amplification (multimodal bias exceeds both unimodal biases), mitigation (multimodal bias is lower than both), and neutrality (multimodal bias lies between unimodal biases), with proportional analyzes conducted to identify the dominant mode and dynamics in these interactions. The findings highlight that amplification (22\%) occurs when text and image biases are comparable, while mitigation (11\%) arises under the dominance of text bias, highlighting the stabilizing role of image bias. Neutral interactions (67\%) are related to a higher text bias without divergence. Conditional probabilities highlight the text's dominance in mitigation and mixed contributions in neutral and amplification cases, underscoring complex modality interplay. In doing so, the study encourages the use of this heuristic, systemic, and interpretable framework to analyze multimodal bias interactions, providing insight into how intermodal biases dynamically interact, with practical applications for multimodal modeling and transferability to context-based datasets, all essential for developing fair and equitable AI models.
