Can schools use AI to mark exams without sending data abroad? A UAE guide

With AI now a taught subject in UAE schools and national assessment bodies tendering for automated scoring, the question for heads is no longer whether to look at AI marking but how to run it without pupils' scripts leaving infrastructure the school controls. The answer is a criterion-by-criterion pipeline with a teacher as final judge, and it is affordable at school scale.

Yes, and the question is no longer optional

In the week of 2 September 2026, the UAE Cabinet approved a national artificial intelligence curriculum for every public and private school in the country, accompanied by plans to train 22,000 teachers and educators to deploy AI for teaching, assessment, curriculum analysis and lesson planning (Khaleej Times). In Dubai, the Knowledge and Human Development Authority (KHDA), the DP World Foundation and MIT RAISE launched a multi-year AI literacy programme reaching approximately 80,500 private school students and 3,600 teachers across Grades 6 to 8, running through February 2030 (Gulf News). At the federal level, automated scoring is already entering institutional practice: tenders for AI-based automated scoring of speaking and writing in the UAE Standardized Proficiency Assessment confirm that machine-assisted evaluation is national policy rather than a classroom experiment.

For school principals, chief technology officers and heads of IT across the Emirates, the strategic challenge has changed. The operational question is no longer whether automated assessment works, but how an educational institution can adopt AI marking without exporting student papers, identification numbers and diagnostic commentary to external servers in unvetted jurisdictions under opaque commercial terms.

The practical answer is clear: schools can run automated marking pipelines entirely on infrastructure they control. The architecture is straightforward, the computational requirements are modest enough to operate on local hardware or private in-country cloud tenancies, and the workflow preserves the classroom teacher as the ultimate arbiter of every released grade.

Why marking is the right first target

Marking represents the single largest recurring administrative burden on teaching staff. Across OECD school systems, educators spend roughly 9% of their total working hours marking and correcting student work. Furthermore, 40% of teachers identify excessive marking demands as a major source of professional stress, with no surveyed education system reporting a figure below 20% (OECD, TALIS 2024).

Unlike conversational classroom chatbots or open-ended generative lesson planners, automated exam marking possesses structural boundaries that make it an ideal target for engineering deployments:

• Bounded inputs and outputs: A pupil script provides a fixed, immutable input. A grading rubric constitutes a rigid specification. Model performance can be evaluated empirically by checking whether automated scores align with independent human moderation.
• Inherent auditability: Academic institutions already archive physical exam scripts and marks. Retaining criterion-level machine justifications, confidence metrics and teacher edits alongside the script introduces no new regulatory or operational record-keeping mechanisms.
• High-frequency returns: Assessment cycles recur across diagnostic entry tests, term assessments, mock examinations and weekly unit quizzes. Operational efficiencies achieved during one cycle compound across the academic calendar.

Crucially, marking must never be approached as an autonomous scoring exercise. The UK qualifications regulator, Ofqual, published an authoritative working paper titled Principles of AI use in marking. Ofqual determined that deploying artificial intelligence as a sole, unsupervised marker violates regulatory principles. The available empirical evidence supports AI exclusively within defined operational contexts: serving either as an assistive second marker or as an automated quality-assurance filter to catch grading drift (gov.uk). School leadership must treat this human-led principle as an absolute engineering constraint.

How AI marking actually works

In production, an automated marking system is an asynchronous ingestion and inference pipeline, not a generic conversational prompt. The workflow spans six distinct stages:

1. Physical capture: Paper examination papers pass through high-throughput production scanners. In commercial education environments, scanning requires dedicated image pre-processing. Skew correction, boundary cropping, double-feed ultrasonic detection, contrast normalisation and the digital suppression of coloured inks (such as a teacher's red annotation pens) must run before images reach an inference engine.
2. Optical text extraction: Handwriting recognition and optical character recognition (OCR) models convert script imagery into structured machine-readable text. Production OCR engines assign character-level and token-level confidence values; any script displaying poor legibility or irregular layouts gets flagged immediately for manual inspection rather than speculative scoring.
3. Rubric decomposition: Monolithic prompts that instruct an LLM to grade an entire essay consistently fail on edge cases. Reliable pipelines decompose complex multi-part questions into discrete, atomic criteria. A prompt asking a pupil to analyse authorial tone, substantiate arguments with citations, and maintain formal grammar is parsed into three distinct criteria, each evaluated in an isolated inference pass.
4. Criterion-level scoring: A local model evaluates the student text against the specific rubric level under strict schema constraints. Using zero temperature and structured JSON outputs, the engine extracts direct text evidence, matches the evidence against rubric descriptors, assigns an integer mark, and generates concise explanatory notes for the pupil.
5. Confidence routing: The pipeline evaluates model certainty against deterministic thresholds. High-confidence evaluations pass into a teacher review queue with proposed scores pre-populated. Ambiguous prose, unconventional narrative structures, unreadable passages, or rubric boundaries displaying low scoring confidence are routed to a human-first triage queue.
6. Teacher adjudication: The subject teacher inspects the digital script alongside the highlighted evidence quotes and proposed marks. The teacher confirms or amends each criterion score before committing the grade to the school information management system. Discrepancies between human decisions and automated proposals are logged, creating an internal calibration dataset tailored to the school's departmental standards.

The programmatic decomposition in step 3 guarantees explainability. When a parent or academic inspector questions why a pupil received 2 marks out of 4 on an analytical question, the institution can present the exact text quote, the matched rubric descriptor, and the human teacher's audit signature.

A rubric definition passed to the scoring engine follows a rigid schema:

{
  "question_id": "ENG-Y10-Q3",
  "max_marks": 4,
  "criteria": [
    {
      "criterion_id": "textual_analysis",
      "weight": 2,
      "levels": [
        {
          "mark": 2,
          "descriptor": "Selects highly relevant quotes and analyses metaphorical language accurately."
        },
        {
          "mark": 1,
          "descriptor": "Selects relevant quotes but provides literal or descriptive commentary."
        },
        {
          "mark": 0,
          "descriptor": "Fails to cite textual evidence or analysis is entirely inaccurate."
        }
      ]
    }
  ]
}

The inference server returns a structured payload adhering to a defined format:

{
  "criterion_id": "textual_analysis",
  "assigned_mark": 1,
  "confidence_score": 0.94,
  "evidence_quote": "the shadows swallowed the courtyard",
  "rationale": "Pupil quotes metaphorical language correctly but explains it literally as nighttime darkness.",
  "flag_for_human_review": false
}

Because the output is deterministic and structured, the marking application can visually anchor the evidence quote within the teacher's interface, allowing the educator to verify the mark in seconds.

The teacher must stay the final judge

The rationale for retaining human oversight rests on psychometric validity. As Ofqual's research highlights, statistical correlation between machine scores and human scores does not prove construct validity. Language models are susceptible to proxy biases: rewarding sheer word count, favouring sophisticated vocabulary over substantive logical argument, or penalising non-standard syntactic phrasing that remains factually correct (gov.uk).

Schools should formalise automated marking policies around a three-tier operational framework:

• Low-stakes diagnostic exercises: Weekly homework checks, vocabulary drills, and formative factual quizzes. The pipeline marks the cohort automatically, while the teacher audits a randomised 10% sample to verify scoring consistency. Time savings exceed 80%.
• Medium-stakes assessments: End-of-unit tests, internal mid-term examinations, and practice mock papers. The automated pipeline pre-scores every script, highlights textual evidence, and constructs preliminary feedback. The subject teacher reads the evidence, reviews flagged answers, adjusts marks where necessary, and authorises release. Review time drops by 60% to 75% compared to manual marking.
• High-stakes milestones: GCSE and IB mock predictions, end-of-year promotion examinations, and scholarship assessments. The classroom teacher or an external moderator acts as the mandatory primary marker. The AI operates solely as a blind second marker or an anomaly detector, flagging discrepancies where teacher marks diverge from historical distribution curves.

This multi-tiered governance structure shields the school during parental appeals or regulatory inspections by bodies such as the KHDA or ADEK. The formal grade remains an uncompromised professional human judgement supported by machine evidence.

The data question: what UAE law actually requires

Student examination scripts contain sensitive personal data: pupil names, identification numbers, individual handwriting biometrics, academic performance profiles, and occasionally contextual records such as special educational needs (SEN) accommodations.

In the United Arab Emirates, processing this information is governed by Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data (PDPL) (UAE Legislation portal). Educational institutions operating under federal jurisdiction must comply with Articles 22 and 23 regarding cross-border data flows:

• Article 22 (Cross-Border Transfer to Approved Jurisdictions): Personal data may only be transferred outside the UAE if the recipient state provides an adequate level of data protection officially recognised by the UAE Data Office.
• Article 23 (Exceptions for Cross-Border Transfers): In the absence of an adequacy decision, transfers abroad require specific legal mechanisms, such as standard contractual clauses imposing statutory PDPL protections on foreign processors, or obtaining explicit consent from the data subject.

For primary and secondary schools, relying on parental consent for cross-border cloud processing creates operational vulnerability. Students are minors, parental consent can be withdrawn at will, and coercive consent (conditioning standard school assessments on transferring data abroad) violates standard data protection doctrines. Furthermore, executing contractual due diligence across multiple overseas SaaS sub-processors to audit model retraining policies, retention cycles, and deletion protocols creates severe administrative overhead.

Educational institutions located within financial free zones face distinct legislative frameworks. The Dubai International Financial Centre (DIFC) enforces the DIFC Data Protection Law No. 5 of 2020, while the Abu Dhabi Global Market (ADGM) applies the ADGM Data Protection Regulations 2021. Both free-zone regimes enforce strict cross-border export requirements and stringent accountability standards for processing children's data.

The cleanest operational strategy is data localisation: maintaining student papers, inference models, and score registries entirely within the borders of the UAE, or within the school's physical premises.

Running it on infrastructure the school controls

Modern open-weight models have rendered proprietary external cloud APIs unnecessary for structured assessment tasks. Extracting evidence and evaluating text against a bounded rubric does not require trillion-parameter frontier networks. It requires compact, instruction-tuned architectures executed with constrained decoding.

Schools can run production inference on modest hardware via serving frameworks like vLLM. Because exam grading is an asynchronous batch workflow rather than an ultra-low-latency chat application, a single dedicated GPU server can process thousands of papers overnight.

For the majority of large school groups and independent private schools, private in-country cloud hosting or on-premises servers offer the ideal balance between operational resilience and regulatory compliance. The inference engine sits behind the school's local firewalls, connected directly to production document scanners over an isolated VLAN.

Azrty designed NovaGrade to address this exact architectural need. NovaGrade delivers automated, criterion-level grading of open-ended examination responses integrated directly into Kodak high-speed scanning hardware. Examination papers are digitised, pre-processed, OCR-converted, and evaluated locally on school-managed infrastructure, ensuring that sensitive student records never depart internal boundaries while teachers transition from manual markers to expert adjudicators. When schools require custom integration into legacy student information systems or specific departmental workflows, Azrty's AI solutions team configures and deploys the underlying platform to meet institutional data residency specifications.

A worked example: one school, one exam cycle

To quantify the operational and computational mechanics, consider a representative Dubai secondary school educating 800 pupils across Years 7 to 11. During an end-of-term assessment cycle, students sit written examinations across six core academic subjects containing open-ended written sections: English Language, English Literature, Science, History, Geography, and Computer Science.

The operational volume scales as follows:

• 800 pupils across 6 examination papers yields 4,800 physical scripts.
• Each examination paper contains 3 substantial open-ended prose questions.
• Each question is assessed against 4 distinct rubric criteria, resulting in 12 criterion-level judgements per script.
• The entire assessment window generates 57,600 individual criterion evaluations alongside cohort-wide feedback summaries.

Manual marking baseline

Under traditional conditions, a secondary school teacher requires 6 to 10 minutes to read, annotate, cross-reference, and grade a multi-page open-ended exam paper.

For 4,800 scripts, the total marking demand consumes between 480 and 800 teacher-hours. Distributed across a departmental faculty of 25 subject teachers, each educator faces 19 to 32 hours of intensive marking over an assessment window. This administrative load routinely displaces lesson preparation and generates severe professional fatigue. Over three academic terms and an additional mock examination series, the institution expends between 1,500 and 2,500 hours annually solely on marking open-ended responses.

Local automated pre-marking with human adjudication

Under an on-premises automated pipeline, examination papers are scanned in bulk at the end of each examination sitting. The processing engine executes optical handwriting extraction, decomposes the rubrics, scores each criterion against embedded evidence, and assigns preliminary confidence metrics.

When teachers open their marking consoles, the student text is already transcribed, rubric criteria are aligned with specific text selections, and suggested marks are pre-populated. Educators spend their time reviewing evidence quotes for unambiguous responses and dedicating clinical attention to flagged borderline answers.

Average review time falls to between 1.5 and 3 minutes per paper:

• Cohort review time: 120 to 240 teacher-hours across the entire faculty.
• Net reduction in manual workload: 60% to 75%.
• Administrative hours reclaimed per cycle: 360 to 560 professional hours.

Computational requirements and economics

Automated assessment tasks are computationally light compared to continuous conversational generation. In an evaluation pipeline, token consumption per criterion breaks down into predictable components:

• System instructions and rubric context: roughly 300 tokens.
• Student written answer extract: roughly 400 tokens.
• Generated JSON evaluation (mark, quote, rationale): roughly 100 tokens.
• Total volume per evaluation: approximately 800 tokens.

Across 57,600 criterion judgements, the assessment cycle processes approximately 46 million tokens of inference. On a standard enterprise workstation equipped with a single modern local accelerator running vLLM, throughput ranges between 60 and 120 tokens per second for batched generation.

The entire cohort workload finishes in approximately 100 to 200 GPU-hours. Scheduled across an overnight batch run over three days of testing, a single on-premises GPU workstation handles the entire school's examination load without pipeline congestion. The capital expenditure of local hardware amortises over several academic terms, delivering clear financial returns when measured against recovered faculty time and reduced staff attrition.

Establishing the calibration baseline

Prior to production deployment, academic departments must validate scoring alignment. The school runs a calibration exercise using 200 historical student scripts marked independently by two senior teachers:

1. The automated pipeline grades the 200 scripts blindly against the official departmental mark scheme.
2. Inter-rater reliability is calculated across each criterion using quadratic weighted kappa metrics.
3. If human-to-AI agreement on a specific criterion drops below 0.80, the rubric is flagged for refinement. Ambiguous descriptor phrasing is clarified, or the criterion is temporarily designated for mandatory manual scoring until the model prompt is tuned.

Executing this calibration protocol requires an afternoon of departmental moderation, providing empirical defensibility before parents, executive boards, and educational regulators.

Questions to ask any marking tool before a pilot

School leaders should require vendors to answer these critical questions in writing before authorising any software evaluation:

1. Where does student data physically reside during processing and rest? Demand precise geographical locations, cloud provider names, and the complete registry of international sub-processors. If any data leaves the UAE, request the statutory PDPL Article 22 adequacy classification or Article 23 transfer instrument.
2. Are student exam responses ingested into training datasets? Commercial agreements must state explicitly that student prose, handwriting images, and teacher feedback will never be utilised for public or proprietary model training.
3. What is the statutory role allocation under the PDPL? The school must remain the sole Data Controller. Any software provider must act strictly as a Data Processor bound by explicit institutional mandates.
4. Does the engine deliver granular criterion-level evidence citations? Reject platforms that provide only a monolithic grade and a superficial overall comment. The tool must return the exact quoted evidence supporting every point awarded or deducted.
5. How does the human override workflow function? Can a teacher amend a mark or rationale with a single click, and does the system log overrides to inform ongoing institutional calibration?
6. What is the platform confidence thresholding architecture? What percentage of student work is automatically escalated to human-first triage? Any vendor claiming near-total autonomous accuracy lacks psychometric credibility.
7. How does the system handle Arabic and mixed-language scripts? In UAE educational settings, models must handle regional handwriting variations, bidirectional text layouts, and code-switched responses without transcription failure.
8. What technical controls govern data retention and automated purging? Can the school programmatically erase scanned images, OCR transcriptions, and scoring logs upon conclusion of an academic appeal cycle without logging support tickets?
9. What empirical validation data supports scoring accuracy? Request independent validation studies demonstrating inter-rater reliability against experienced human examiners on comparable student demographics and qualification boards.
10. Can the software execute entirely on an on-premises server or in-country private cloud? If the solution cannot function without constant outbound connectivity to external shared services, evaluate the associated operational and regulatory risks accordingly.

What to do next

Adopting machine-assisted assessment successfully requires disciplined, step-by-step institutional governance:

1. Scope an isolated pilot: Select a single secondary year group and two subjects characterised by well-structured mark schemes, such as Year 9 Science and History. Avoid rolling out cross-school deployments simultaneously.
2. Assemble the calibration dataset: Collate 150 to 200 anonymised student papers from the previous academic year. Have two teachers mark them independently to establish a reliable baseline of human agreement before benchmarking the automated tool.
3. Establish your deployment model: Decide early whether the engine will run on a dedicated local hardware unit in the school's server room or within a UAE-hosted virtual private cloud. This decision eliminates compliance uncertainty before vendor demonstrations begin.
4. Publish your assessment charter: Draft a clear two-paragraph internal governance policy. Clarify that automated tools perform preliminary rubric extraction and pre-scoring, that teachers review and sign off on every mark, and that summative exit evaluations remain strictly human-led.
5. Track operational metrics: Monitor three quantitative indicators across the pilot cycle: total teacher-hours saved, human-to-AI scoring agreement across rubric criteria, and the frequency of post-results mark adjustments.

When educational institutions combine criterion-level prompt architecture, local hardware execution, and mandatory teacher adjudication, automated exam marking ceases to be a speculative compliance hazard. It transforms into an efficient, defensible internal utility that protects pupil data sovereignty while returning hundreds of weekend hours to teaching faculties.

https://www.azrty.com/blog/can-schools-use-ai-to-mark-exams-without-sending-data-abroad-a-uae-guide

<a href="https://www.azrty.com/blog/can-schools-use-ai-to-mark-exams-without-sending-data-abroad-a-uae-guide">Can schools use AI to mark exams without sending data abroad? A UAE guide</a> (Azrty)