Stellar classification is a fundamental challenge in astrophysics, often complicated by the well-known age-metallicity degeneracy. While traditional methods rely on fitting specific indices or full spectra, we propose a novel approach using an unsupervised dense autoencoder to analyze spectra from the Medium-resolution INT Library of Empirical Spectra (MILES). This model is trained to learn the intrinsic features of stellar spectra, compressing them into a low-dimensional latent space. A K-means algorithm is then applied to this space to group the spectra into six distinct clusters. The core strength of our method lies in the concept of a pseudospectrum, which is generated by reconstructing the centroid of each cluster's latent representation. Unlike a simple average, the pseudospectrum is a noise-free archetype that encapsulates the most significant features of a stellar spectral type as perceived by the AI, providing a clear visual representation of each cluster. This work stands apart from prior AI applications involving the MILES database, such as the supervised deep learning approach by Wang et al. (2019) for predicting stellar parameters, or the semi-empirical library created by Knowles et al. (2021) to model variable abundances. Our unsupervised method represents a paradigm shift, focusing on data-driven discovery and classification rather than on predictive or modeling tasks. The resulting pseudospectra not only validate the model's efficacy but also open a new era for understanding stellar populations through the lens of AI-driven, unsupervised pattern recognition.