Fernando Rodriguez-Merino

Nicolas Rodriguez-Alvarez, Alan Martin Blanch-Marsolini, Samuel Vara-Gutierrez et al.

The ubiquitous presence of digital devices has cemented the 'Digital Native' paradigm, assuming inherent technological proficiency among contemporary youth. This multicenter study ($N=243$ European secondary students) challenges this narrative by investigating the gap between self-perceived digital literacy and actual technical readiness, including Artificial Intelligence (AI) interaction. Our findings reveal a severe Confidence-Competence Divide characterized by a collective Dunning-Kruger effect: students report near-maximum self-efficacy in passive digital consumption but exhibit a sharp decline when evaluating active technological creation and algorithmic logic. Crucially, an intra-pathway analysis demonstrates that the technological gender gap is not universal; rather, it emerges significantly exclusively within Technology-oriented classrooms ($p = 0.046$), indicating the persistence of 'stereotype threat' in formal STEM environments. Additionally, the study uncovers an 'AI Paradox' wherein students significantly overestimate their critical awareness of deepfakes and algorithmic biases compared to their operational AI skills, fostering a false sense of invulnerability against modern misinformation. Ultimately, supported by an overwhelming student demand ($76.5\%$) for pedagogical reform, this research concludes that dismantling this illusion of competence requires abandoning passive theoretical instruction in favor of hands-on, active technological creation.

Nicolas Rodriguez-Alvarez, Fernando Rodriguez-Merino

Deep Neural Networks are highly susceptible to shortcut learning, frequently memorizing low-dimensional spurious correlations instead of underlying causal mechanisms. This phenomenon not only degrades out-of-distribution robustness but also induces severe demographic biases in sensitive applications. In this paper, we propose a geometric \textit{a priori} methodology to mitigate shortcut learning. By deploying a zero-hidden-layer ($N=1$) Topological Auditor, we mathematically isolate features that monopolize the gradient without human intervention. We empirically demonstrate a Capacity Phase Transition: once linear shortcuts are pruned, networks are forced to utilize higher geometric capacity ($N \geq 16$) to curve the decision boundary and learn ethical representations. Our approach outperforms L1 Regularization -- which collapses into demographic bias -- and operates at a fraction of the computational cost of post-hoc methods like Just Train Twice (JTT), successfully reducing counterfactual gender vulnerability from 21.18\% to 7.66\%.