This paper presents an autoencoder-based Variable Active Antenna Spatial Modulation (AE-VASM) approach to improve multiple-input multiple-output systems in Rician fading channels with moderate antenna correlation. Leveraging autoencoders (AEs) for joint transmitter and receiver optimization, the model integrates modulation and channel encoding into a unified structure. This design mitigates noise and antenna correlation, enhancing system robustness and power efficiency. Simulation results show that the AE-VASM model achieves approximately a 2 dB bit error rate (BER) improvement over the benchmark at a Rician factor of 3 dB, demonstrating its advantage in moderate scattering environments. The findings highlight the potential of autoencoder-based methods to advance spatial modulation through neural network-driven dynamic signal adaptation. This study supports the integration of deep learning into spatial modulation for next-generation wireless networks and suggests future research to enhance the AE-VASM model’s scalability and performance in diverse communication scenarios.