Machine Translation Quality in 2026: Good Enough or Not?
If you run a shop or a business site, you have probably asked yourself whether AI translation is finally good enough to publish without a translator looking over every line. The honest answer in 2026 is “often yes, sometimes absolutely not.” Machine translation quality has improved so much that the best systems now rival human translators on research benchmarks, yet the same research shows exactly where things still break. Here is what the evidence says, so you can decide which pages to automate and which deserve a human.
What the latest benchmarks actually show
The most rigorous yardstick for machine translation is the annual General Machine Translation task at the Conference on Machine Translation (WMT). Professional annotators score real system outputs, which makes it far more trustworthy than vendor marketing.
WMT24: the LLM era arrives
The 2024 findings paper carries a telling title: The LLM Era is Here but MT is Not Solved Yet. Organizers evaluated 11 language pairs and collected outputs from 8 large language models and 4 online translation providers on top of the submitted systems. A few results stand out for website owners:
- The best systems for English to Spanish produced translations the authors described as close to flawless, making it the easiest pair in the test (WMT24 findings).
- Claude 3.5 Sonnet was the best system overall, winning 9 of 11 language pairs (WMT24 findings).
- Human reference translations landed in the winning cluster in only 7 of 11 pairs, although the authors note the English to Hindi reference was of suboptimal quality (WMT24 findings).
- Speech was the hardest domain, likely because errors from automatic speech recognition carried into the translation (WMT24 findings).
WMT25: harder tests, same pattern
For 2025 the organizers deliberately picked more difficult source documents and titled their paper Time to Stop Evaluating on Easy Test Sets. They evaluated 60 systems in total, 24 of them LLMs or popular online translation services. Even on harder material, Gemini 2.5 Pro sat in the top cluster for 14 language pairs, while human references made the winning cluster in only 6 of 15 pairs (WMT25 findings).
The detail that matters most for you is the spread. The best system almost always scored 90 or higher on the 100 point scale, yet mid-tier models often scored 60 or below and the weakest could drop to 0 depending on language pair and domain (WMT25 findings). “AI translation” is not one quality level; the engine you pick matters.
Where the sources disagree
The research is not perfectly consistent:
- Automatic scores vs. human judges. In WMT24, Unbabel Tower70B ranked first on automatic metrics across all languages but did not win across the board with human annotators (WMT24 findings). In WMT25, Shy hunyuan MT topped the automatic rankings for all but one pair, while human evaluation placed it considerably lower (WMT25 findings). Treat any vendor quality score based only on metrics with caution.
- Literary text. WMT24 found news, literary and social content similarly difficult, while WMT25 rated literary texts the easiest. The WMT25 authors themselves caution that error-based scoring can reward translations that are technically correct but miss the author’s voice (WMT25 findings).
- Vendor studies. Intento, a company that sells translation automation, reports that its 2025 evaluation of 46 engines across 11 language pairs found that customizing terminology, tone, formatting and consistency produced five to ten times fewer errors than the baseline (Intento). Useful, but not an independent benchmark.
Where AI translation is good enough
AI translation is reliable when three things are true: the language pair is well resourced, the text is reasonably standard and an occasional slightly awkward phrase would not hurt anyone.
Even in medicine, high-resource languages perform well. A 2024 study in Pediatrics compared ChatGPT and Google Translate with professional translators on 20 discharge instruction sets and found the machine output comparable to professional translation for Spanish and Brazilian Portuguese (American Academy of Pediatrics). A 2019 study of emergency department instructions found Google Translate 92% accurate for Spanish (PSNet).
For a typical business site that translates into German, French, Spanish, Italian or Portuguese, this means product descriptions, blog posts, FAQs and category pages are realistic candidates for AI translation with light spot checks.
Where AI translation is not good enough
Low-resource languages and dialects
Quality drops sharply once you leave the big languages. A large evaluation of GPT models across 204 languages found them underperforming traditional machine translation for 84.1% of the languages covered, with resource level the strongest predictor of quality (Robinson et al.). A 2025 study of 200 languages on FLORES-200 still found low-resource languages clearly behind high-resource ones (Song et al.).
In WMT25, translation into Maasai was described as largely unusable, with systems often producing Swahili instead. For Egyptian Arabic, systems tended to output Modern Standard Arabic rather than the dialect requested (WMT25 findings). If your audience speaks a regional variety, check that the output is actually in that variety.
Medical content
The Pediatrics study that looked fine for Spanish told a different story for Haitian Creole: potentially clinically significant errors appeared in 33.3% of ChatGPT translations and 23.3% of Google Translate translations, compared with 8.3% for professional translations (HealthDay). A 2021 study of Google translated discharge instructions, summarized by the National Center for State Courts, found accuracy ranging from 94% for Spanish down to 55% for Armenian (NCSC). Some of this research tested older tools, but none of the studies we found shows the gap for smaller languages closing.
Legal content
The NCSC’s 2025 guide for US courts states that machine translation cannot yet be relied on to convey complex information or legal concepts and that courts should never use it to convey legal or procedural information, while allowing drafts for review by professional translators (NCSC). The same guide notes US Department of Justice guidance that discourages machine translation alone even with a disclaimer. For a business site, that applies to your terms and conditions, privacy policy, warranty terms and anything contractual.
Messy input
WMT25 found that state-of-the-art systems still struggle with non-standard input, linguistic complexity, domain-specific terminology and gender agreement in some language pairs (WMT25 findings). Think slang, puns and product jargon. For brand names and product terms, a glossary helps a lot; see our guide on translating brand terminology.
What Google says about machine-translated pages
Google’s spam policies list “automated transformations like synonymizing, translating” as a form of scaled content abuse when content is scraped to generate many pages that provide little value. That policy was introduced in March 2024 and targets content made mainly to manipulate rankings, however it was produced.
In June 2025, Google told SEO consultant Glenn Gabe that its policies “do not strictly define content that has been translated by AI as spam” (GSQI). Gabe also noted that Google removed older advice about blocking auto-translated pages from its multilingual documentation. The practical reading: translating useful pages is fine, while mass-translating thin pages to grab traffic is the risk.
Good enough, review needed or human only?
The table below sorts typical website content by how much human involvement the evidence above suggests. It assumes a well-resourced language pair; for smaller languages, move everything one column to the right.
| Content type | Good enough with AI | Review needed | Human only |
|---|---|---|---|
| Product descriptions and category pages | ✓ | ||
| Blog posts and help articles | ✓ | ||
| Navigation, buttons and UI text | ✓ | ||
| Headlines, slogans and ad copy | ✓ | ||
| Checkout, shipping and returns info | ✓ | ||
| Health, safety or dosage information | ✓ | ||
| Terms, privacy policy and contracts | ✓ |
The pattern is simple: the higher the cost of a misunderstanding, the further right a page belongs. UI text sits in the middle because short strings carry little context, so a quick check on the live page is worth it. Marketing copy sits there because, as WMT25 hinted, a translation can be error-free and still lose your voice.
How to make AI translation work on your site
You do not need a big localization budget for this:
- Pick the engine by evidence. The WMT spread between top and mid-tier systems is huge, so test a few pages in your target languages before committing.
- Keep a glossary and translation memory. A glossary targets the terminology problems WMT25 flagged.
- Review where it counts. Spend human time on checkout, legal and safety pages, not on your 400th product description. Our post-translation editing guide covers light vs. full review.
- Fix errors in context. Mistakes are easiest to spot on the live page, not in a spreadsheet.
- Send legal text to a professional. Use AI drafts at most as a starting point, as the NCSC suggests for courts.
This is the workflow heylingo is built around: you add one script tag, pages are translated by AI with a translation memory and you can correct any string inline on the live site, in 30+ languages, for EUR 20 per 1,000,000 translated characters per month. It does not replace a lawyer reviewing your terms and it is not meant to.
FAQ
Is AI translation better than human translation in 2026?
For some well-resourced language pairs, top systems now match human references. In WMT25, human translations made the winning cluster in only 6 of 15 pairs (WMT25 findings). That does not hold for every engine or language.
Will Google penalize my machine-translated pages?
Not for being translated. Google’s spam policy targets scaled, low-value content. Google has also said AI-translated content is not strictly defined as spam (GSQI). Translate pages that are useful in the first place.
Which languages are risky for AI translation?
Languages with little digital training data, such as Haitian Creole, Armenian or Maasai in the studies above, plus dialects like Egyptian Arabic. Have a native speaker check those before publishing.
Should I use ChatGPT or a dedicated translation engine?
The evidence does not crown one approach. LLMs won WMT overall, but the 2023 FLORES study found traditional machine translation stronger for most low-resource languages (Robinson et al.). Test on your own content.
Conclusion
Machine translation quality in 2026 is good enough for most everyday website content in major languages. Top systems now rival human translators on research benchmarks. It is not good enough for legal text, health information or smaller languages without a human check. Sort your pages by the cost of a mistake, automate the low-risk majority and put your review time where errors would actually hurt. If you want to compare approaches in more depth, read our take on Google Translate vs. professional translation services.
See heylingo pricing and start translating your site
Sources
- Kocmi et al., Findings of the WMT24 General Machine Translation Shared Task, ACL Anthology, 2024
- Kocmi et al., Findings of the WMT25 General Machine Translation Shared Task, ACL Anthology, 2025
- Robinson et al., ChatGPT MT: Competitive for High- (but not Low-) Resource Languages, arXiv, 2023
- Song et al., Is LLM the Silver Bullet to Low-Resource Languages Machine Translation?, arXiv, 2025
- American Academy of Pediatrics, Research evaluates use of machine language translation platforms for pediatric discharge instructions, 2024
- HealthDay, Report on the Pediatrics machine translation study, 2024
- AHRQ PSNet, Assessing the use of Google Translate for Spanish and Chinese translations of emergency department discharge instructions
- National Center for State Courts, Machine Translation: Considerations and Cautions for Courts, 2025
- Google Search Central, Spam policies for Google web search
- Google, New ways we’re tackling spammy, low-quality content on Search, 2024
- Glenn Gabe, Blog post on auto-translating content and scaled content abuse, GSQI, 2025
- Intento, The State of Translation Automation 2025