A network trained to reproduce its own input after squeezing it through a narrow middle. Because the middle is too small to hold everything, the network must learn what matters. That compressed middle is the useful part; the reconstruction is just the training signal.
Technical view
An encoder maps input to a lower-dimensional latent code and a decoder reconstructs, trained on reconstruction error. Used for dimensionality reduction, denoising and anomaly detection, where high reconstruction error flags unusual inputs. Without constraints it can learn an identity mapping, hence the bottleneck, or sparsity and denoising variants. The sparse autoencoders used in interpretability are a deliberate inversion, expanding rather than compressing.
Related
Part of AI Terminology Index, section: How the Machinery Works.