Converting text between languages. Effectively solved for widely spoken language pairs in everyday contexts, and far from solved for low-resource languages, specialist terminology, legal precision and anything where tone or cultural register carries meaning. The gap between good enough to understand and good enough to publish remains large.
Technical view
General language models now match or exceed dedicated translation systems on many pairs, with advantages in context handling and register control and disadvantages in latency, cost and consistency. Evaluation has moved from n-gram overlap measures such as BLEU to learned metrics correlating better with human judgement. Document-level context, terminology consistency through glossaries, and handling of named entities are the practical determinants of professional usability.
Related
Part of AI Terminology Index, section: Modalities and Applications.