Abstract
The rapid advancement of artificial intelligence (AI) technologies presents both transformative opportunities and critical challenges across technical and societal domains. This thesis combines a technical investigation into data security for machine learning systems with a socio-technical analysis of how emerging technologies should be integrated into education. On the technical side, we develop CleanSight, an automated pipeline for detecting label-flipping data poisoning attacks in image classification datasets. Using the MNIST benchmark, we evaluate four detection strategies—k-means clustering, Gaussian mixture models, Mahalanobis distance, and Wasserstein distance—across a comprehensive experimental design. Results show that k-means clustering achieves the strongest overall performance (AUC = 0.9761, recall = 0.9312), while Gaussian mixture models provide the highest precision, demonstrating the feasibility of scalable, pre-training data validation methods.
Complementing this technical work, we introduce the Task Reallocation Framework, a theoretical model for integrating new technologies into education. Drawing on historical case studies of calculators and spell-check systems, the framework argues that technology should automate computational tasks that learners have already mastered while reallocating cognitive effort toward higher-order critical thinking. Applied to generative AI tools, the framework highlights both the risks of cognitive offloading and the opportunity to enhance reflective problem-solving through carefully designed instructional use.
Together, these contributions demonstrate a unified perspective on AI as both an engineering system requiring robust safeguards and a societal force reshaping human cognition and learning. By linking secure AI development with responsible educational integration, this work emphasizes the importance of designing technologies that not only perform reliably but also augment, rather than diminish, human critical thinking.
