[visarga]

Beautiful idea, an autoencoder must represent everything without hiding if is to recover the original data closely. So it trains a model to verbalize embeddings well. This reveals what we want to know about the model (such as when it thinks it is being tested, or other hidden thoughts).

[sobellian]

It could just invent its own secret language embedded into English akin to steganography. The explanation would not lose information but would remain uninterpretable by humans