Beyond Automation: Human-centered Approaches To Neural Network Integration In Architectural Design

This research explores the contemporary relationship between AI-driven tools and human perception and examines both the opportunities and challenges emerging from the intersection between digital technologies and human-centered design approaches. The integration of neural networks in architectural workflows --- spanning from simple image generation to more complex algorithmic optimizations --- arises critical questions concerning the emotional, cultural, and phenomenological quality of the spaces that such AI-based models can produce. Tools such as Midjourney, Stable Diffusion, Finch3D, etc. have demonstrated a remarkable efficiency in terms of form-finding and spatial optimization, even though prioritizing computational perfection over perceptual nuance, with results of leading to the generations of bidimensional spaces that lack cultural specificity, place attachment or emotional resonance. From these premises, this paper aims to identify and analyze the limitations of existing generative design tools, propose and prototype alternative frameworks for human-centered AI in architecture. Grounded on a mixed-methods approach, it combines case study analysis and user perception studies through eye-tracking and heatmap visualization, for the development of custom AI model called TheChair.net. The prototype is designed to tackle the identified limitations of existing generative tools by incorporating user-uploaded architectural vocabularies, enabling personalized, memory-informed design generation that maintains authorial identity while exploring the potentialities of computational power. Indeed, the research proposes a human-oriented generative framework where artificial intelligence is intended as a collaborative partner rather than a mere autonomous designer. Through the integration of perceptual feedback, cultural sensitivity, and ethical concerns into neural network training and deployment, some bias patterns inside AI-based tools are identified and analyzed to prevent architectural homogeneity towards diversity. The finding challenges deterministic fears of AI replacing architects, instead positioning technology as a medium through which human spatial intelligence could be improved while becoming computationally scalable without losing its essential human touch.