By: Dr. Tamara Patzer
Ask a chatbot a question in English and the answer usually sounds fluent. Ask the same question in Tamil, Finnish or Haitian Creole and the quality can slip. The grammar may wobble, the tone may feel off, or the answer may quietly miss what a local reader would take for granted. Companies building these systems know it, and many are now hiring bilingual people to find the gaps.
The reason traces back to a simple imbalance in the material AI systems learn from.
Key Takeaways
- A survey from Beijing Foreign Studies University, reported July 23, 2025, found English makes up 70 to 80 percent of the training text in more than 50 multilingual models.
- Researchers who hired native speakers to rate AI answers in Hindi, Tamil, Bengali and seven other languages collected about 90,000 evaluations across 30 models.
- Automated AI judges tend to score answers higher than people do, which is one reason companies still pay humans to evaluate.
- Advertised hourly ranges for bilingual roles on one company’s listings ran from $30 to $65 when last checked.
- Requirements and pay change by project, so every posting should be read in full.
The Imbalance Behind the Demand
The EurekAlert summary of the Beijing Foreign Studies University survey, dated July 23, 2025, reached a blunt conclusion about the raw material. Across more than 50 multilingual models, English accounted for between 70 and 80 percent of the text used in training. Everything else, from Spanish and Mandarin to Swahili and Icelandic, shares the remainder.
That does not mean the models cannot speak other languages. It means they learned them from a thinner, more uneven diet. Idioms, politeness levels, local references and regional spelling are exactly the details that a thin diet misses.
What Native Speakers Actually Do
The clearest public example is a research project called Pariksha, described in a June 2024 paper posted to arXiv. The team recruited native speakers through Karya, a data work platform, and asked them to score AI-generated answers in Hindi, Tamil, Bengali and seven other languages. Across 30 models, the project gathered roughly 90,000 evaluations. The task was not translation. It was judging the answers the models produced.
Day to day, bilingual contract work follows the same logic. A reviewer may read two chatbot answers and pick the better one, correct a mistranslated phrase, flag an offensive or confusing reply, or rewrite a response so it sounds the way a person in a specific region would say it. Voice projects may ask for recorded speech, and audio roles may involve transcription. Details vary by company and by listing, which is why the posting matters more than the headline.
Why Machines Cannot Simply Grade Themselves
A natural question is whether one AI could check another and spare everyone the expense. Researchers led by Rishav Hada tested this in 2023 and found that automated AI judges tended to score answers higher than human raters did. Native-speaker evaluation, they wrote, remains the gold standard, though it is costly. A grader that flatters the work is a poor grader.
A second study points the same way. Pinzhen Chen and colleagues at the University of Edinburgh and Tsinghua University, writing in June 2024, compared models tuned on translated text with models tuned on text written natively in the target language. The comparison asks whether text written natively teaches a model something that translated text does not. Native writing is harder to obtain, which is where human contributors come in.
What It Looks Like at Company Scale
Vendors sell this as a service. Welo Data published a case study on May 12, 2026, describing work across 20 locales and more than 36,000 annotations. It is the vendor’s own account, so readers should treat it as a description of what the work involves rather than an independent audit. Even so, the scale is useful context: a single project can need dozens of reviewers across many regions.
On the Expert Opportunity Report’s reading of one company’s listings, checked on November 2, 2026, advertised hourly ranges for bilingual roles ran from about $30 to $65. The site has since gathered nine listings in a single overview of bilingual and voice roles, so readers can compare language pairs and requirements side by side.
Who Tends to Qualify
The research above points to a clear pattern. The valued skill is not a certificate in translation. It is the ability to judge language as a native or near-native speaker would. That includes people who grew up speaking a language at home, teachers, interpreters, journalists, nurses and others who use two languages in daily work. A listing may ask for a particular level of fluency, a country of residence or a subject specialty, and some projects pair language skills with professional knowledge, such as medicine or law.
Applicants should not assume that a rare language guarantees work. Projects open and close with customers’ needs. A language that is in demand this quarter may be quiet next quarter, and a common language may have many applicants for few seats.
Questions Worth Asking Before You Sign Up
- Does the posting name the exact language and dialect, such as European or Latin American Spanish?
- What test or interview is used, and how long does it take?
- Is the work contract work, and who handles taxes?
- How is quality scored, and what happens if an answer is rejected?
- What are the expected hours, and can the project end without notice?
A Note on Caution
Demand for languages makes a tempting hook for scammers, who pitch “translation jobs” with fake checks or fees. A real employer does not charge to hire you and writes from its own official domain. When in doubt, go to the company’s website directly and look for the same listing there.
Frequently Asked Questions
Do I Need to Be a Professional Translator?
Not necessarily. The research examples relied on native speakers judging whether answers were natural and correct. Individual listings set their own requirements.
Why Not Just Translate English Data Into Other Languages?
Researchers compare it with text written natively to see whether the two teach a model different things.
Is the Pay the Same in Every Language?
No. Rates vary by company, project, region and skill level, and they change.
Where Can I Follow Open Language Roles?
The free Expert Opportunity Report publishes a regular list of open expert and contributor roles, with the hourly range shown on each posting.
About the Author
Dr. Tamara Patzer is a publisher and former newspaper editor who writes about AI and business data.
You may reach her through: info@dailysuccessinstitute.com
Sources
EurekAlert, summary of the Beijing Foreign Studies University survey of multilingual models, July 23, 2025.
Pariksha study, arXiv, June 2024.
Rishav Hada and colleagues, arXiv, 2023.
Pinzhen Chen and colleagues, University of Edinburgh and Tsinghua University, arXiv, June 2024.
Welo Data case study, May 12, 2026.
Expert Opportunity Report listing notes, checked November 2, 2026.
Disclaimer: The author publishes the Expert Opportunity Report and is a referral partner of micro1. This article is information, not a promise of work or income, and not legal or financial advice.
