Definition
Natural Language Autoencoder is a system that compresses a model’s internal activations into a plain-English description and then reconstructs (or approximates) those activations from that description, using natural language as the bottleneck instead of a learned latent vector. It was proposed as an interpretability monitor on the theory that a human-readable bottleneck would surface whatever concepts the network is actually representing, but in practice it failed to flag concepts a model had learned to suppress or encode indirectly, showing that fluent verbal summaries can miss what is functionally driving behavior.