Translation was one of the first professional fields disrupted by AI, and in 2026 the disruption is mature enough to see clearly. Machine translation has not eliminated translators — but it has fundamentally changed what translators are paid to do. The volume play (translate 2,000 words from scratch for a flat rate) is mostly gone. The expertise play (ensure a medical device manual will not cause harm in German; capture the emotional register of a literary novel in Japanese) is worth more than ever. Here is where the profession stands.
What changed in 2026
- Neural MT quality crossed a threshold for general business content in major language pairs (EN-DE, EN-FR, EN-ES, EN-ZH, EN-JA). For commodity content, post-editing is now often faster than translating from scratch.
- Specialized domain models. DeepL, ModernMT, and SYSTRAN now offer domain-adapted models for legal, medical, and technical content — still imperfect, but meaningfully better than general-domain MT.
- Context-aware MT. Document-level context (not just sentence-by-sentence) is now standard; paragraph-level coherence improved significantly.
- AI in CAT tools. memoQ, SDL Trados Studio (now Language Weaver integrated), and Phrase all have AI suggestions that blend MT, translation memory, and glossary in real time.
- AI translation quality metrics. Tools like COMET and BLEURT are now widely used for automated quality scoring; buyers use them to set post-editing thresholds.
The honest picture: MT quality by content type
| Content type |
MT quality (2026) |
Human value add |
| Software UI strings |
High |
Low — mostly light editing |
| Marketing copy |
Medium |
High — brand voice, cultural adaptation |
| Technical documentation |
High |
Medium — terminology, edge cases |
| Legal contracts |
Medium-Low |
High — legal precision, liability |
| Medical device / pharma |
Medium-Low |
Very High — regulatory requirement |
| Literary fiction |
Low-Medium |
Very High — voice, style, culture |
| Financial reports |
High |
Medium — numbers check, regulatory terms |
| Subtitles / captions |
Medium-High |
Medium — timing, spoken register |
Post-editing machine translation (PEMT): the new baseline
PEMT is now the standard workflow for most translation agencies. The process: MT generates a draft, translator reviews and edits to target quality, with time and rate adjusted based on MT quality and expected edit density.
Full PEMT vs. light PEMT:
- Light post-editing: Fix only errors that impede meaning or cause compliance risk. Stylistic improvements not required. Fast — ~60–80% time saving vs. from-scratch translation.
- Full post-editing: Bring the translation to the same quality as human-from-scratch. Typically 40–55% time saving vs. from-scratch.
Translators who resist PEMT entirely are pricing themselves out of volume work. Translators who do only PEMT without deep expertise in specialized domains are competing on a commodity that MT is winning.
AI-assisted CAT tools: the daily workflow
The most productive 2026 translator workflow uses a CAT tool with AI suggestions, a domain-specific glossary, and a translation memory. In practice:
- Import source document into CAT (memoQ, Trados, Phrase)
- AI + TM suggests a translation for each segment
- Translator accepts, rejects, or edits each suggestion
- Glossary terms auto-highlighted and auto-inserted where consistent
- Final QA pass (automated consistency check + human scan)
This workflow outperforms manual translation by 30–50% even in literary domains when glossary discipline is maintained.
Where human translators are irreplaceable
Literary translation. The quality gap between human literary translation and MT remains large. Literary translators make thousands of micro-decisions about register, rhythm, cultural allusion, and implied meaning that MT systems get wrong in ways readers notice.
Legal translation. Mistranslated contract terms cause real legal liability. Certified legal translators carry professional insurance; MT does not. For anything entering legal proceedings, human certification is non-negotiable.
Medical and pharmaceutical translation. Regulatory bodies (FDA, EMA) require qualified human review of translated label copy, package inserts, and clinical documentation. MT errors in dosing instructions can cause patient harm.
High-stakes marketing localization. Brand campaigns require cultural adaptation (not just translation) that MT cannot perform — idiomatic phrases, humor, and cultural references need native human judgment.
How to pick AI translation tools
- For volume post-editing: DeepL Pro ($30–60/month) or ModernMT (usage-based, ~$0.01–0.03/word) with your CAT tool of choice.
- For specialized domains: Request domain-adapted models from your MT provider; test quality on real content before committing to a workflow.
- For terminology management: memoQ has the most mature terminology management; integrates with AI suggestions naturally.
- For quality scoring: COMET-based scorers are available as standalone tools and are integrated into enterprise platforms. Use them to set QA thresholds, not to replace human judgment.
Common mistakes
Using general-domain MT for specialized content without adjustment. A general MT model translating medical device instructions will produce plausible-sounding but potentially dangerous errors.
Skipping terminology management. Inconsistent technical terms across a document are the most common, most damaging MT quality failure. Build your glossary before the MT pass, not after.
Under-pricing PEMT. PEMT is cognitively demanding — reading critically, catching errors, rewriting for style. It is not "cheap translation." Rates for full PEMT should reflect that; many agencies and clients need education here.
Treating MT quality metrics as ground truth. COMET scores predict human quality judgment well on average; they fail on domain-specific terminology and cultural nuance. Human QA is still required for high-stakes content.
What to skip
- Raw MT output for any regulated content — legal, medical, pharmaceutical — without qualified human review.
- AI translation tools marketed specifically to "replace translators" — they work for low-stakes content only and create legal risk if misused.
- Monolingual post-editing (editing target text without access to source) — it produces fluent-sounding content that may be semantically wrong.
FAQ
Is the translation profession dying?
Volume is shifting to PEMT and MT with human oversight. The profession is restructuring, not dying. Specialized, literary, and certified translation commands higher rates than ever. The mid-range generalist market has compressed.
Which language pairs are best handled by MT in 2026?
European language pairs with English (EN-DE, EN-FR, EN-ES, EN-IT, EN-PT) have the highest MT quality. Asian language pairs (EN-JA, EN-ZH, EN-KO) improved significantly but have higher post-editing effort for nuanced content.
How should I price post-editing work?
Typical PEMT rates run at 40–70% of full translation rates depending on MT quality and domain. Request a sample segment before agreeing to a rate — MT quality varies significantly by content type.
Can AI tools translate audio or video content?
Yes — combined ASR (speech recognition) + MT pipelines handle subtitling and dubbing prep. Quality is good for clear audio in supported languages; still requires significant human editing for broadcast quality.
Where to go next