On-Screen Marking (OSM) vs AI Evaluation: What Exam Bodies Should Know

Ask two exam controllers what “digital evaluation” means and you may get two different answers. One is on-screen marking (OSM): answer books are scanned and evaluators mark them on a computer. The other is AI evaluation: software reads the handwriting and proposes marks. They are often mentioned together, but they solve different problems.

This article explains both, what India’s experience with OSM has shown, and how an AI first pass can sit inside an OSM-style process without taking the final mark away from a human evaluator.

What on-screen marking is

In on-screen marking, students write on paper as usual. After the exam, answer books are scanned and the images are sent to evaluators, who log in, read each answer on screen and enter marks question by question. The software adds up the total, records which evaluator marked which script, and lets senior examiners review a sample.

CBSE’s FAQ on OSM puts it plainly: the method changes, not the criteria. The answer book is on a monitor instead of a table, and a mouse is used to award marks. Asked whether computers are checking the answers, CBSE’s answer is no: examiners evaluate the scanned copies.

How OSM is used in India

CBSE Class 12, 2026

In a circular dated 9 February 2026, CBSE told its schools that Class 12 answer books would be evaluated by on-screen marking from the 2026 examinations, while Class 10 would stay in physical mode. Its later FAQ says the board first tried OSM in 2014 but stopped, because answer books could not then be scanned without cutting them, which risked mixing pages. In 2026 the books were scanned with book scanners that did not need the spine cut, marking schemes for 116 subjects were uploaded to the portal, and Head Examiners could randomly check evaluated answer books.

The scale was large. The Ministry of Education said about 98 lakh Class 12 answer sheets were scanned, as reported by PTI. CBSE declared results on 13 May 2026 for about 17.69 lakh candidates. After the results, students and parents raised concerns that a drop in the pass percentage was due to OSM. The Ministry dismissed those concerns and announced a cut in the fee for a scanned copy of an evaluated answer book, from ₹700 to ₹100.

State boards and universities

In February 2026, PTI reported that the Punjab School Education Board would use OSM for one subject of its matriculation exam that year, on answer sheets with a QR code on every page, after evaluating 23,000 answer sheets on screen in its September 2025 supplementary exams.

Universities moved earlier. Deccan Herald reported in April 2012 that Visvesvaraya Technological University’s examination paper evaluation was going digital. The University of Mumbai’s switch in 2017 is the cautionary tale: Newslaundry reported that the rollout was rushed without the agency finalised or teachers trained, results were badly delayed, and over 73,430 results were initially held back because answer scripts were misplaced or not scanned or uploaded properly.

Outside India, OSM is well established in England. Ofqual, the exams regulator there, notes that all four exam boards in its marking study (AQA, OCR, Pearson and WJEC) use on-screen marking and monitoring for some components.

What OSM solves

• Logistics. Answer books stay at the scanning centre. CBSE listed teachers staying in their own schools and savings in transport time and cost among the expected benefits.
• Totalling and posting errors. The system adds the marks. CBSE’s FAQ says OSM eliminates totalling, posting and uploading errors and ensures every answer is evaluated.
• Live supervision. Senior examiners can check samples while marking is under way. In England, boards slip pre-marked “seed” responses into an examiner’s allocation (Ofqual says sampling rates of about 5% are typical) to compare the examiner’s marks with an agreed standard.
• Records for post-result services. Every mark is entered per question against a known evaluator. CBSE gave Class 12 students scanned copies of their evaluated answer books, then a window to ask for verification or re-evaluation.

What OSM does not solve

Evaluators still read every answer. OSM saves the trip to the evaluation centre, not the reading. Ninety-eight lakh answer sheets still need enough evaluators to read every page of every one.

Evaluator variance stays. Two careful evaluators can give the same answer different marks, especially in long answers. Ofqual’s 2018 analysis of on-screen marking data found that the probability of a candidate receiving the “definitive” grade (the one senior examiners would give) ranged from 0.96 for a mathematics qualification to 0.52 for an English language and literature qualification, though the probability of getting the definitive grade or the one next to it was above 0.95 for all of them. OSM measures variance well. It does not remove it.

New failure points appear. Scanning adds steps that can go wrong. CBSE’s June 2026 window for post-result verification listed the issues students could report: missing pages, missing supplementary sheets, missing maps or graphs, blurred pages, an incorrect answer book, and evaluation against a different set. The Ministry said more than 13,000 answer sheets written in very light ink were identified and checked manually. CBSE’s February circular had expected that post-result verification of marks would no longer be required. In practice a verification route was still needed, for a different kind of error.

Security becomes a bigger topic. Scanned answer books put a whole cohort’s personal data in one system. On 1 June 2026, The Week reported CBSE’s statement that it had been monitoring vulnerabilities flagged in public in its service provider’s portal, had brought in cybersecurity experts from government and the IITs, and that the identified vulnerabilities had been contained.

What AI first evaluation adds

AI evaluation of answer sheets does what OSM does not: it reads first. A typical system reads the handwriting, finds each answer, compares it with a marking scheme and proposes a mark with reasons. The evaluator’s job changes from reading every answer from scratch to checking a proposed mark and changing it where they disagree.

• One rubric for every script. The same criteria are applied to the first script and the ten-thousandth, without end-of-day fatigue.
• Reasons you can audit. Per question: what was read, which criteria were met, what mark was proposed and why.
• Triage. Low-confidence reads and unreadable pages are flagged for a human instead of being marked by guesswork.
• An annotated copy. Some systems write ticks, crosses and notes onto the script itself, so the student or a re-evaluator can see where marks were won and lost.

The limits matter just as much:

• It should not make the final decision. Marks stay a draft until a human releases them.
• It reads text. Diagrams, maps and graphs need a human eye; a text-based system can judge only their labels and written explanation.
• It must be tested on your subjects, languages and handwriting. Ask for results on your own past scripts, not general claims.
• It can be wrong in ways that look confident, so human review and sampling must continue.

How the two combine

Think of OSM as the desk and AI as a first reader sitting at it. Three patterns work:

1. 1

   AI first pass, human reviews every script

   The AI proposes marks and reasons per question. The evaluator sees the scan beside them, confirms or changes each mark, and the human mark is final. Suits university semester exams, internal exams and mock tests.

2. 2

   Human marks first, AI as a check

   Evaluators mark as they do in OSM today. An AI mark is computed in the background, and large disagreements go to a senior examiner, like a seed check that covers every script. Suits bodies whose rules require a human to enter every mark.

3. 3

   Split by question type

   AI first pass on text answers with clear marking schemes; human-only evaluation for diagram-heavy questions and for languages the system has not been tested on.

Whichever pattern you choose, protect the human edit: once an evaluator changes a mark, no later AI run should overwrite it.

Risks and governance

• Write down who owns the final mark. Your rules should say that a named human role decides every mark and that nothing is published without human release.
• Keep re-evaluation unchanged. Students keep the same rights to see their copy and ask for re-evaluation. The re-evaluator should see the AI’s proposed mark and reasons, and every human change.
• Keep an audit trail. Per question: the scan, what the system read, the proposed mark and reason, the evaluator’s mark, and who changed what, and when.
• Gate scan quality. Borrow CBSE’s layered checks: at scanning, by a quality control team, and by the evaluator, who can reject an unreadable copy for rescanning.
• Pilot before go-live. Run the new process on past scripts with known marks, in parallel with the old one, and compare. Mumbai’s 2017 experience shows the cost of skipping this step.
• Protect the data. Answer books carry students’ names, roll numbers and handwriting. India’s Digital Personal Data Protection Rules, 2025 were notified on 14 November 2025 with an eighteen-month phased compliance period, and include duties on security safeguards and telling affected people about a breach without delay. Check where scans are stored, who can open them, how long they are kept and how vendors are audited, and take legal advice on your obligations.
• Be open about AI use. Tell evaluators and students where AI is used and how a human checks it. UNESCO’s 2023 guidance on generative AI in education calls for protecting data privacy and a human-centred approach to validating such tools.

A decision table

For a side-by-side view of the two approaches, see Evalezy vs on-screen marking.