Language load is the share of an assessment score driven by reading the question rather than by the ability the question is meant to measure. It is construct-irrelevant variance — real, measurable, and invisible in a total score. In India, where the 2011 Census recorded English for 10.6% of the population, an English-only aptitude test quietly measures reading alongside reasoning.
That is not an argument against testing in English. It is an argument against pretending you know which of the two a candidate failed.
The construct you meant to measure, and the one you did
The Standards for Educational and Psychological Testing — the joint AERA, APA and NCME document that most defensible assessment practice in the world is built on — treats construct-irrelevant variance as a central fairness problem, not a rounding error. Its fairness chapter is explicit that linguistic demands unrelated to the construct are a threat to valid score interpretation. The open-access front matter and chapter list are published at testingstandards.net.
Here is what that looks like on a real screening form. A numerical reasoning item for a warehouse supervisor role opens with a 45-word scenario in English about pallet throughput across three shifts. The arithmetic is two steps. The reading is harder than the arithmetic. A candidate who can do the arithmetic and cannot parse the sentence gets the same score as a candidate who can do neither — and the hiring manager sees one number.
Across a 30-item test with a 25-minute limit, that reading load stops being an item-level quirk and becomes a systematic penalty on a specific, identifiable group of candidates — one your score distribution absorbs without comment.
How large is the affected pool in India
The best public evidence on the entry-level end of India's talent pool is the ASER Centre's 2023 "Beyond Basics" survey of 34,745 young people aged 14 to 18 across 28 rural districts in 26 states. Its reading and arithmetic findings are specific:
- About 25% could not fluently read a Standard II level text in their own regional language.
- 57.3% could read sentences in English.
- Of those who could read the English sentences, 73.5% could say what they meant.
- 43.3% could correctly do a three-digit by one-digit division.
Take the second and third figures together and roughly 42% of this cohort can both read an English sentence and state its meaning. The remainder are not low-ability candidates. They are candidates whose ability your English-language instrument cannot see.
Two honest caveats. ASER's sample is rural and adolescent, so it describes the pool feeding frontline retail, logistics, field sales and entry-level BPO roles — not an urban engineering campus, where the English picture is very different. And the Census figure is self-reported proficiency from the 2011 bilingualism and trilingualism tables, now fifteen years old, with fewer than 260,000 people reporting English as a mother tongue and about 129 million reporting it as a second or third language.
Note also who has already acted on this. India's large public recruiters do not test in English only: Indian Railways widened its recruitment examinations beyond Hindi and English in 2018, and Railway Recruitment Board computer-based tests are now offered in 15 languages. Private screening at comparable volumes frequently still runs one English form for everybody.
Translating the test is not the fix most people assume
This is where the critique turns on the obvious remedy. The experimental literature on language accommodations is larger than most hiring teams realise, and it does not support blanket translation.
Li and Suen's meta-analysis of accommodation studies, published in Applied Measurement in Education in 2012 (ERIC EJ981547), found that test accommodations improved English-language-learner performance by about 0.157 standard deviations on average. After controlling for other variables, the only significant moderator was the examinee's own English proficiency — low-proficiency candidates gained substantially more than high-proficiency ones. The authors found no support for the common claim that linguistic simplification outperforms other approaches.
Pennock-Roman and Rivera's meta-analysis of 14 randomised and counterbalanced studies, in Educational Measurement: Issues and Practice in 2011 (ERIC EJ940995), reached compatible conclusions across 50 groups of English learners and 32 comparison groups:
- Plain English rewrites produced very small average effects overall, with signs of more benefit at intermediate proficiency.
- A full native-language version produced the single largest individual effect in the set, +1.45, for low-English-proficiency Spanish-speaking students.
- Computer-administered glossaries worked even when time limits were tight.
- Printed-material accommodations only worked when time limits were generous for both groups — which the authors read as evidence that the accommodation reduced construct-irrelevant variance rather than inflating scores.
Both are US school studies, not Indian hiring studies, and should be read as evidence about the mechanism rather than as effect sizes you can transplant. But the mechanism is the point: the benefit is concentrated in the candidates whose language proficiency is actually binding, and simplifying the English is the weakest of the available moves. Translation is a targeted intervention masquerading as a universal one.
Three ways a translated assessment goes wrong
You create two forms and report one scale. A Tamil version and an English version of the same test are two instruments until you have evidence they behave the same way. That means differential item functioning analysis and measurement invariance testing before scores from the two are ranked in a single shortlist. The ITC Guidelines for Translating and Adapting Tests, second edition, set out what the statistical equivalence evidence has to cover; most internal translation projects produce none of it.
You translate the items and keep the clock. Reading in a second language is slower before it is less accurate. A time limit set on English-medium pilot data turns a comprehension difference into a speed penalty on every translated form.
You translate a test whose construct genuinely includes English. If the role is English-language voice support, reading and composing English is the job, and stripping the language load out of the assessment destroys the measure instead of improving it.
When an English-only test is the right call
Sometimes it is. The test is correct when the job analysis establishes that English reading and writing are part of task performance, not an access route to the test — international voice processes, English-language written customer correspondence, documentation-heavy technical roles. The discipline is that the requirement has to come from the job analysis and be documented there. In the US, where this has been litigated hardest, the EEOC's enforcement guidance on national origin discrimination treats fluency requirements as lawful only to the extent they are genuinely required by the position. The underlying logic — measure what the job needs, and be able to show why — is portable to any jurisdiction.
What to do instead
- Audit the reading load of your stems. Count words and clauses per item on every non-language test. A reasoning item whose stem is longer than its solution path has a problem.
- Measure English separately, and report it separately. A short dedicated English assessment plus a low-language ability measure gives you two interpretable numbers instead of one ambiguous one. AssessAll's AI-graded scenarios make this manageable at volume, provided the rubric scores decision quality separately from fluency of expression.
- Set the language policy per role family, not per company. The warehouse supervisor and the international voice agent should not be screened on the same linguistic assumptions.
- Loosen the clock on any form with heavy text. Where speed is not part of the construct, generous limits remove a confound at no cost to validity.
- Collect invariance evidence before you merge score distributions. If you cannot run differential item functioning analysis, keep the language cohorts in separate shortlists and set separate cut scores.
- Pilot the translated form on incumbents. Incumbents give you a performance criterion to check the translated version against.
Steps 1, 3 and 4 cost nothing but attention. Steps 2, 5 and 6 are where a custom assessment build earns its keep, and where pay-as-you-go credits at ₹30 per assessment make piloting a second language version affordable enough to actually test rather than assume.
The takeaway: an English-only aptitude test in India does not fail to measure — it measures two things and reports one. Decide deliberately which of them belongs in the score, document why, and you have an instrument you can defend; leave it to chance and you have a reading test with a reasoning label on it.