The rapid proliferation of Large Language Models (LLMs) and generative artificial intelligence has fundamentally disrupted traditional higher education paradigms, shifting the focus from knowledge retention to critical evaluation. This comprehensive mixed-methods study explores the pedagogical efficacy of integrating generative AI as a collaborative partner rather than treating it merely as a vehicle for academic dishonesty. By implementing controlled, AI-assisted research modules across 15 undergraduate humanities and social science courses involving 1,200 students, this research evaluated changes in academic writing quality and analytical depth. The empirical findings demonstrated that students who were explicitly trained to use AI for brainstorming, outlining, and counter-argument generation—while strictly authoring the final prose themselves—produced essays that scored 28% higher on rubrics measuring critical synthesis and argumentative complexity. Conversely, control groups lacking structured AI guidance frequently exhibited superficial engagement with source materials. Qualitative interviews with faculty underscored a significant reduction in time spent correcting basic syntactic errors, allowing educators to focus feedback on higher-order theoretical concepts. However, the study also highlights profound ethical and cognitive risks, notably the phenomenon of "automation bias," wherein students blindly accept AI-generated hallucinations without secondary verification. To combat this, the authors propose a robust instructional framework centered on "AI literacy," which mandates that students critically audit and document their interactions with LLMs. The paper ultimately argues that academic institutions must urgently pivot from punitive plagiarism-detection models to proactive integration strategies, ensuring graduates are equipped to navigate an AI-augmented professional landscape ethically and effectively.