Leak of an entrance examination paper for admission to India’s medical schools (NEET-UG 2026), a subsequent massive protests by students and resignation of Union Education Minister Dharmendra Pradhan has put National Testing Agency (NTA) under intense scrutiny.

NTA conducts six major entrance examinations of India, NEET-UG, NEET-PG, JEE Main, CUET-UG, CUET-PG, UGC-NET- in 13 languages across 5000 centers for millions of applicants every year.

But recent examination paper leak and criticism over the accuracy of use of AI in question papers, has pushed the agency into a wider overhaul with tighter controls around confidential operations, new security infrastructure and changes to how sensitive examination content is handled.

In that regard, on October 3, Higher Education Secretary Deepti Gaur Mukherjee had visited newly relocated NTA’s office in New Delhi to review the infrastructure and administrative setup ahead of upcoming examinations.

Now the agency is looking to empanel ‘Translation Reviewers’ for AI-based question-paper translation verification across its examinations, in conjunction to AI-based translation platform.

The Expression of Interest (EOI) says the translation workflow is being restructured into two parts: bulk translation by an air-gapped AI-based platform deployed on NTA premises, followed by verification, certification and quality assurance by empaneled human professionals.

The agency’s application invite said that the role is being introduced alongside the increasing use of AI-produced translation, with human reviewers responsible for checking translations for accuracy, clarity, consistency and contextual relevance across 13 Indian languages.

The application window is open until October 9, 2026.

Under this role, the reviewer will not translate question papers from scratch. “The Reviewer shall review AI-produced translations for accuracy, terminological appropriateness, linguistic fidelity and clarity, and shall certify each translated item as either ‘accepted’, ‘accepted with corrections’, or ‘rejected with reasons’,” said NTA in EOI.

The role also limits reviewers’ access to sensitive examination content. They will see individual questions one at a time in randomized order and will not see the assembled paper. The work must be carried out at an NTA secure facility using NTA-provided workstations, with no personal devices or external transfer of content permitted in the work area under continuous CCTV monitoring with AI-based behaviour analytics.

The reviewer role also comes with strict confidentiality requirements. Reviewers must sign an NDA and are prohibited from retaining, copying, photographing, transcribing or otherwise preserving examination content. They must also follow security protocols including device deposit, biometric authentication, two-person rules and activity logging.

The agency has used sovereign AI models installed on its own GPUs to translate examination material for the NEET-UG retest.

In an exclusive interview with AI FrontPage, NTA Director General Abhishek Singh explains what has changed in the agency’s AI-generated translation workflow, why human review remains crucial and mandatory in the exam process, what checks are performed before a translation reaches a reviewer, and how NTA is restricting access to confidential examination content.

Question: What is an “air-gapped AI-based translation platform”?

NTA DG Abhishek Singh: An air-gapped platform is one that is physically and logically isolated from all external networks. In the case of the National Testing Agency’s translation platform, this means that:

• The AI translation model is deployed on dedicated servers at NTA premises, and not on any public cloud or internet-connected environment.
• The premises have no internet connectivity, no external network access, no cloud connectivity, and no permitted use of personal devices, personal e-mail, USB drives or any other removable storage media.
• Only authorized NTA-provided devices are used, within a controlled workspace with continuous CCTV coverage and access-control protocols.

The purpose is straightforward: confidential examination material remains within the boundary of a physically secured NTA facility during the translation stage, with no external network or removable-media pathway for transferring the content.

Question: What are the key challenges in using AI to translate examination questions into Indian languages, and why is human linguistic expertise important?

NTA DG Abhishek Singh: AI translation of examination questions into Indian languages is materially different from routine machine translation and presents specific challenges:

• Domain-specific terminology in Physics, Chemistry, Biology, Medicine and Botany, including taxonomic names, formulae, units and technical vocabulary, requires standardized, subject-consistent rendering across all languages.
• Semantic precision matters absolutely: in a multiple-choice question, a subtle shift in a word or an option can inadvertently reveal or alter the correct answer.
• Grammatical structure differs materially across Indian languages, including word order, gender, case markers and agglutination, and this affects the framing of the question and the options.
• Regional and dialectical variation requires calibration to the standard form of each notified language.
• Equivalence of difficulty across all language versions of the same question has to be preserved so that no candidate is advantaged or disadvantaged by the language chosen.

For these reasons, the AI output is only a first-cut draft. The role of human linguistic and subject-matter expertise is central and non-negotiable: the AI accelerates and standardizes the first pass; human reviewers ensure fidelity, precision, difficulty equivalence and cultural appropriateness.

The HLCE  (Humanity’s Last Code Exam), comprising 235 of the most challenging problems from the International Collegiate Programming Contest (ICPC World Finals) and the International Olympiad in Informatics (IOI) spanning 2010-2024, itself observed that reliance solely on AI or machine learning for translation of examination questions is ‘not completely feasible’ because of these variations.

NTA’s workflow therefore retains mandatory human review of the AI-generated output.

Question: How does NTA measure the accuracy of AI-generated translations before human review?

NTA DG Abhishek Singh: Before any AI-generated translation is placed before human reviewers, it is subjected to a set of automated quality signals on the platform itself, including:

• Model confidence scores on each sentence, flagging low-confidence segments for priority review.
• Terminology consistency checks against a curated, subject-wise glossary of standard terms in each language.
• Structural integrity checks for the preservation of numerics, units, symbols, formulae and named entities between the source and translated text.
• Round-trip/back-translation validation. 
• Anomaly detection for unusually short, unusually long or structurally incomplete translations.

These signals provide a preliminary quality picture and help determine the sequence of human review. They do not, in themselves, certify a translation.

Every question, without exception, is placed before human reviewers, and no translation enters the examination without full human verification.

Question: How will translation reviewers verify the questions? What are the steps in the review process?

NTA DG Abhishek Singh: Every AI-generated translation, without exception, is subjected to human review on the air-gapped platform before it enters the question paper. Each question is reviewed by a bilingual subject-matter expert who verifies fidelity to the source, correctness of technical terminology and equivalence of difficulty across languages.

The reviewer also checks linguistic correctness, natural readability and the absence of regional bias.

A back-translation validation is also carried out, in which the translated text is retranslated into the source language and compared with the original to detect any semantic drift. Every action of every reviewer on every question is captured in an auditable log.

The combined effect is that no translation enters the examination without independent, multi-layered human verification, with AI used for the first-cut draft.

Also Read: Important for Teachers to Learn AI: India’s Education Secretary to AI FrontPage