Evidence · strength 1 of 5 · searched to 2 Nov 2026
Does AI help professional translators?
Not enough evidence yet to say whether AI helps professional translators. Across 2 studies, time per task changed by −31.1% to +3.5%.
What a large trial with professional translators and smaller workplace studies found about translating with large language models: time, quality and earnings.
VerdictNot enough evidence yet
No good evidence
ImpactNo clear effect
Across 2 studies, time per task changed by −31.1% to +3.5%.
Evidence strength1 of 5
We have very low confidence in this estimate.
Cost per user a year$
General AI models or features inside translation software; costs are small and not reported.
Pilot rating under our v0.3 rules. The rules are published on our method page but not yet independently reviewed, so ratings may change as the method is checked and new studies arrive. Our method.
What the studies found
Time (headline)
No clear effect−31.1% to +3.5%- Merali (2024): −31.1% Time per translation task, any LLM vs no AI (600.7 vs 413.8 seconds) (p = 0.000).
With access to any AI model this was reduced to 413.8 seconds (p=0.000), a statistically significant reduction in time of 31.1%.
- Castaldo et al. (2025): −44.4% Total editing time, post-editing GPT-4 output vs translating from scratch (64.33 vs 115.68 minutes, all translators) (no statistical test reported). Not statistically clear.
GPT-4 64.33 ... HT 115.68 ... Table 1: Editing Times (in Minutes)
- Castaldo et al. (2025): +3.5% Total editing time, post-editing GPT-3.5 output vs translating from scratch (119.74 vs 115.68 minutes) (no statistical test reported). Not statistically clear.
HT 115.68 ... GPT-3.5 119.74
Errors
No clear effect- Tellingen et al. (2026): +21.9 points Share of deliberately inserted critical errors fixed: LLM highlights plus correction suggestions (S-APE) vs plain post-editing (no test of this comparison reported). Not statistically clear.
Perc. % 62.5 90.62 90.62 84.37
Output
Small gain- Martikainen et al. (2025): +14% Translation speed (characters per second), post-editing GPT-4 output vs translating from scratch, within the same documents (P = .03).
postediting machine translation is 14% faster than translating texts from scratch (2.75 vs 2.40 characters per second, P =.03), and up to 37% faster
Quality
No clear effect- Merali (2024): +0.14 SD Grade by three professional graders (7-point scale), any LLM vs no AI (p = 0.148). Not statistically clear.
This equates to an increase of 0.14 standard deviations in the average grade.
- Castaldo et al. (2025): +1 creativity score points Creativity score (creative shifts per unit of creative potential x COMET), GPT-4 post-editing vs human translation (26.5% vs 25.5%) (authors report no significant difference). Not statistically clear.
GPT-4 0.32 0.83 26.5% ... HT 0.30 0.85 25.5% ... the creativity score does not differ significantly between the four models
A minus sign on time means the task took less time. Bands are provisional; see how impact is banded. Effects are not pooled yet: each band is set from one result per study, and where studies use different units we show only the direction.
How sure are we?
- Starting level2 of 5, from 2 studies graded A, B or C.
- Down one: old models0 of 4 studies show they used a model released since May 2025 (the 18-month window); the rest used older models or do not say which.
- Publication biasnot testable with fewer than 10 studies.
- Result1 of 5. We have very low confidence in this estimate.
Behind the average
- Translators slower than the median on the baseline task ('low-skill') (Merali (2024)): Time fell 21.1% per tenfold compute increase vs 4.9% for faster translators (difference p = 0.017)
- Individual translators (Martikainen et al. (2025)): Speed change ranged from -2% (the translator with most segments) to +102%
How, where and why it works
Why it works
2 of 4. Translators post-edit an AI draft instead of translating from scratch; newer models saved more time.
For whom and where
2 of 4. Slower translators gained most in the largest trial.
What it takes
1 of 4. Most studies are small; one ran inside a public-sector translation team.
The full cost
1 of 4. One trial reports earnings per minute; tool costs are not counted.
Four separate scores from 0 to 4, never added up: how much the studies tell us about why it works, for whom, what it takes and what it really costs (the EMMIE method).
What to do next
Measure your pilot →If you are trying it yourself
- Post-edit with a current model, and check meaning, not just fluency.
If you are rolling it out
- Measure time and quality on your own texts before changing rates or workflows.
What it takes
- Results depend on the model: post-editing an older model's output was no faster in one study.
What can go wrong
- Quality gains were not statistically clear in the largest trial.
- Apart from one large trial, the studies so far are very small: four to eight translators each.
The studies
| Study | Grade | Design | People | AI model (released) | Company role |
|---|---|---|---|---|---|
| Scaling Laws for Economic Productivity: Experimental Evidence in LLM-Assisted Translation Ali Merali, 2024. arXiv 2409.02391v2 (working paper) | B | RCT, realistic task | 300 professional translators (Spanish, Hindi, Arabic; at least one year of paid experience) recruited on Freelancer and Fiverr, 1,800 tasks; about 1,500 task observations in the AI vs control models and 1,392 in the compute models | 13 LLMs of differing training compute (not individually named; largest is Claude 3.5 Sonnet) | none |
| Evaluation of Generative Artificial Intelligence Implementation Impacts in Social and Health Care Language Translation: Mixed Methods Case Study Miia Martikainen, Kari Smolander, Johan Sanmark, Enni Sanmark, 2025. JMIR Formative Research 2025 (journal) | C | other quasi-experiment, deployed work | 4 in-house translators at a Finnish wellbeing services county; 908 segments in the productivity comparison; 1,373 segment pairs for quality metrics | GPT-4 via Microsoft Azure OpenAI, through the RWS 'AI Professional' plugin for Trados Studio | none |
| Extending CREAMT: Leveraging Large Language Models for Literary Translation Post-Editing Antonio Castaldo, Sheila Castilho, Joss Moorkens, Johanna Monti, 2025. arXiv 2504.03045v1 (preprint) | D | other quasi-experiment, realistic task | 4 professional literary translators (English to Italian), about 2,200 words each of a novel, rotated across three LLMs and translation from scratch | GPT-4; GPT-3.5; Mistral-7B fine-tuned for literary translation | none |
| Smarter edits? Post-editing with error highlights and translation suggestions Fleur V. J. van Tellingen, Gautam Ranka, Dora Žugčić, Joyce van der Wal, Andrea Camasta, Livio Guerra, Alina Karakanta, 2026. EAMT 2026 conference proceedings (journal) | D | other quasi-experiment, realistic task | 8 professional English-Dutch translators, 8 texts (news and biomedical, 200 words each), four conditions in a Latin-square design | TowerInstruct-13B-v0.1 (xTower) for translation, error explanations and corrections; xCOMET-XXL for quality-estimation highlights | none |
Grades: A, a well-run randomised trial in real work; B, a strong natural experiment in real work or a well-run trial on realistic tasks; C, a trial with clear limits or a matched comparison; D, before-and-after, surveys or case studies. How studies are graded.
Does the rating hold up?
- Run without company-run studiesNo change: no such studies.
- Run without working papers and preprints2 studies: strength 0, verdict no good evidence.
How up to date is this evidence?
Method →Evidence searched to 2 Nov 2026. The studies do not all state model release dates. Rating computed 5 Nov 2026 under Module 15 v0.3 (pilot).
How this page was made: AI research agents searched for studies and read each paper; a second pass checked every number on this page against the paper itself, with the quote and where it appears. The rating is computed by fixed rules, not chosen. No person outside AI Adoption has reviewed it yet. Editorial owner: AI Adoption editors. Spot a mistake? Corrections.
Disclosure: AI Adoption uses AI tools, including Claude from Anthropic, in its research and editorial work. No company pays for or has any say in a rating.