An autoencoder learns to compress input data into a lower-dimensional latent (bottleneck) representation via its encoder, then reconstruct the original input from that representation via its decoder, trained by minimizing reconstruction error on normal data. Because the model is optimized specifically to reconstruct patterns it has seen frequently during training, it becomes proficient at compressing and reconstructing 'normal' instances but performs poorly --- producing high reconstruction error --- on inputs that deviate structurally from the training distribution, i.e., anomalies. Thresholding reconstruction error thus provides a natural, unsupervised anomaly score without requiring labeled anomalous examples, which are often scarce or unavailable in real-world settings.

This mechanism is the operative characteristic tested here, not classification accuracy (B, which describes a supervised discriminative task the autoencoder is not directly trained for), image enhancement (C, a description closer to denoising autoencoders' side effect rather than the core anomaly-detection mechanism), or forecasting (D, which describes sequence models like RNNs/LSTMs applied to time series, a different architecture family and objective).

Variants such as variational autoencoders (VAEs) extend this idea probabilistically, and in multimodal settings, cross-modal autoencoders can flag anomalies where reconstruction fails to reconcile one modality given another.