Western Uusimaa Wellbeing Services County
Western Uusimaa county tests GPT-4 post-editing in its in-house health and social care translation team
Finland's Western Uusimaa Wellbeing Services County let its four in-house translators use GPT-4 inside their translation software for four months. Post-edited GPT-4 output was 14% faster than translating from scratch, but gains ranged from -2% to 102% across translators.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
14%
faster (2.75 vs 2.40 characters per second, P =.03)
Speed of post-editing GPT-4 output compared with translating from scratch, 4 translators, 908 timed segments1
up to 37% faster
Speed gain when segments were equalized across translators (a supplementary analysis the authors call non-standard)1
-2% to 102%
Range of individual productivity change across the 4 translators1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Finland's public social and health care sector is under pressure to do more with the same budget. Western Uusimaa is a bilingual region, so documents must exist in Finnish and Swedish, and increasingly in other languages. Its in-house translation team has four people. The county wanted to know whether generative AI could raise their output without hurting quality.
What they built
The county's translators used GPT-4, a large language model from OpenAI, through a plugin inside Trados, the translation software they already use. The plugin calls GPT-4 through Microsoft Azure. It suggests a machine translation for each sentence-sized segment, and the translator then edits it. This is called post-editing. The department had no standardized practice of using machine translation aids before this.
How they did it
Four translators worked on their normal assignments from March to June 2024. They were asked to translate half of each text's segments from scratch and half with GPT-4, picking the starting method at random. Material with sensitive or personal data was left out. The study recorded timestamps in Trados to estimate speed, scored output quality with four automatic metrics, and held five interviews, including the team supervisor.
The setup was basic. The translators used only one system prompt they could not see. No term bases and no task-specific prompts were used.
Results
Post-editing GPT-4 output ran at 2.75 characters per second against 2.40 when translating from scratch, which the authors report as 14% faster. When the authors evened out how many segments each translator contributed, they report up to 37% faster. Individual results ranged from -2% to 102%.
Quality was mixed. Of the Finnish-Swedish outputs, 11% were accepted without any edit, and 16% of the Finnish-English outputs. Translators reported problems with specialist terminology and the tool's lack of context across a whole document. Three of the four wanted to keep using a similar tool.
What to copy
Run a short, instrumented trial on real work before buying. Track speed per person, give the team training in post-editing, and collect their view of terminology problems before you scale.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Speed of post-editing GPT-4 output compared with translating from scratch, 4 translators, 908 timed segments | 14% faster (2.75 vs 2.40 characters per second, P =.03)1 | |
| Speed gain when segments were equalized across translators (a supplementary analysis the authors call non-standard) | up to 37% faster1 | |
| Range of individual productivity change across the 4 translators | -2% to 102%1 | |
| Finnish-Swedish GPT-4 outputs accepted without edits | 11% (141/1261)1 | |
| Finnish-English GPT-4 outputs accepted without edits | 16% (18/112)1 | |
| Translators who wanted to keep using a similar AI tool, from 5 interviews (4 translators and the team supervisor) | 3 out of 41 |
Lessons for your project
Measure per person, not only on average. Gains ranged from a small loss to more than double, so the team average hides who benefits.
Budget for training in post-editing. Reviewing and fixing machine output is a different skill from translating, and the authors tie results to it.
Fix the tool's weak points early. Translators named inconsistent terminology and missing document context as the main blockers.
Expect staff to rate the gain lower than the stopwatch does. Translators were more cautious than the measured data, so plan change management.
Keep sensitive text out of the trial. The county excluded material with sensitive or personal data.
Caveats
- The source is a peer-reviewed paper by researchers from LUT University, the county's own research and development department and HUS. The county approved the study. No separate outlet confirms the figures, so the case is company-reported.
- Only 4 translators took part, and one of them produced 61% of the timed segments. The authors say results cannot be generalized.
- Translation time was estimated from software timestamps, with filters applied by the authors. The 37% figure comes from a supplementary analysis the authors call non-standard; the abstract says up to 37% while the results text says on average 37%.
- The journal site could not be read directly. The figures were checked against the full text on Europe PMC (PMC12443352), which prints the publication date of 17 September 2025.
- The deployed date is the start of the data collection period (March 2024). The paper does not say whether the tool stayed in use afterwards.
- Quality was scored with automatic metrics against the translators' final versions, not by independent human reviewers.
Sources
- Evaluation of Generative Artificial Intelligence Implementation Impacts in Social and Health Care Language Translation: Mixed Methods Case Study
JMIR Formative Research · 2025-09-17 · Research paper · formative.jmir.org
Summarized in our own words from the sources above. Spotted an error? See corrections.