Machine translation quality has improved faster in the last three years than in the previous twenty. Here is what is genuinely different now.

Context is the big unlock

Modern systems translate paragraphs and documents as coherent units rather than sentence by sentence, which fixes pronoun agreement, consistent terminology and register drift.

Tone and formality became controllable

Explicit formality settings are now standard, which matters enormously for languages with grammaticalised politeness.

Low-resource languages caught up — partly

Transfer learning has dramatically improved coverage for languages with limited training data, though the quality gap with major European pairs has not closed entirely.

What is still hard

Humour, poetry, deliberate ambiguity and highly specialised legal terminology remain firmly in human territory.