Exam Malpractice Detection: Ethics First

We will say the unpopular sentence first. Most campuses should not buy an exam-malpractice model as their second AI project, and many should not buy one at all. A false positive in a hostel ticket wastes a plumber's afternoon. A false positive in an unfair-means file can stall a degree, a visa, and a first job. Vendors talk about computer vision as if it were a second invigilator. It is a second accuser that does not attend the hearing.

This is an opinionated field guide for controllers of examinations, unfair-means committees and syndicates. The opinion is narrow. Detection tools can support a human invigilation system that already exists. They must not become the finding. They must not score faces for 'anxiety'. They must not run on a child's entrance test because a coaching centre shares your campus. And they must live inside the ordinance you already use for unfair means, not beside it in a vendor portal.

UGC and professional councils issue exam-related instructions from time to time. We will not invent a 2026 proctoring regulation. If your council has published one, pin it. If it has not, your ordinance and principles of natural justice are still there. They were enough to restrain a zealous superintendent. They are enough to restrain a model.

Accusation is a legal act

An unfair-means proceeding is an administrative act that affects a student. The student gets notice of the charge, a chance to answer, and a decision by a competent committee. A dashboard that says 'gaze anomaly 0.83' is not a charge. If the superintendent copies that number into the charge sheet without describing a human-observed fact, you have outsourced the accusation.

Write the permitted inputs. A human invigilator's contemporaneous note is an input. A sealed CCTV clip the committee can watch is an input. A vendor's emotion score is not an input we would allow on a system we run. Anxiety looks like cheating to a model trained on other people's idea of honesty.

Remote proctoring of take-home or online tests adds a home. Homes have siblings, shared phones, poor networks, and religious or gender constraints on cameras. A 'room scan' is a search of a private space. Treat it as such: lawful basis, necessity, a less-intrusive alternative, and a real exemption path. If you cannot write those, do not scan rooms.

If the committee cannot watch it and the student cannot answer it, it is not evidence.

Signal	May support a human	Must not be the finding
Invigilator contemporaneous note	Yes, it is the traditional record	A note that only repeats a model score
CCTV or authorised recording the committee can play	Yes, with retention and access limits	A vendor highlight reel the campus cannot export
Browser lock / network log on a campus exam build	Yes, as a technical fact	A claim that a packet equals intent
Gaze, emotion, 'cognitive load'	No, not on our paper	Anywhere near a charge sheet
Voice or face match against a gallery	Only with a written identity purpose and a high error review	Silent exclusion of a student who grew a beard or wore a scarf

Ordinance before SKU

Amend the unfair-means ordinance, or issue a standing order under it, before any tool goes live. Name the tool as an aid to the superintendent. Name the committee as the finder of fact. Give the student the clip or the log you rely on. Give a network-failure protocol that is not 'assume guilt'.

Entrance tests run by the university on behalf of a state or a consortium have another master. You may not have the right to add a vendor camera. Read the conducting-body instructions. Do not improvise.

Malpractice in evaluation — leaked keys, dummy examiners — is a different crime. Do not point a student-facing proctoring tool at that problem and declare victory. Staff the key ceremony. Split duties. An agent that drafts a better seating plan is more useful than an agent that stares at eyes.

Minors, bodies and the camera in the hostel

Seventeen-year-olds sit university-adjacent exams. Children's data, images and behavioural monitoring sit under Section 9 of the DPDP Act: verifiable parental consent, a bar on processing that is detrimental to well-being, and restrictions on tracking and behavioural monitoring, subject to whatever exemptions the live Rules actually prescribe. A gaze model is behavioural monitoring. Do not shrug.

Gendered and disability impacts are not edge cases. A proctoring model that flags looking down will punish students who use a screen reader, a paper draft, or a child in the room. Build exemptions before the first accusation, not after the first protest.

Do not put always-on cameras in hostels in the name of academic integrity. That is a different surveillance programme and it will be read as one.

What we will help with — and what we will not

We will help a controller retrieve the ordinance, draft a seating plan, calendar evaluators, and assemble a pack the committee already asked for. We will not sell a gaze score. We will not run a room scan. If that loses a deal, it loses a deal.

If you already bought a proctoring suite, write the ordinance around it now. Turn off emotion features. Export clips to your store. Delete vendor copies on a schedule. Put a human between the dashboard and the charge sheet. That is mitigation, not enthusiasm.

Seating is the better exam agent

Before you buy a gaze model, buy a seating agent that respects published constraints: distance, accessibility, course clash, invigilator load, and rooms the fire officer has actually cleared. That agent does not accuse. It produces a chart a superintendent can change. False seats are embarrassing. False accusations are careers.

Evaluator logistics are the same class of problem. An agent that calendars packets, tracks dummy numbers, and refuses to show a name next to a dummy is doing integrity work. An agent that 'predicts who will leak a key' is doing superstition. Keep dummy isolation in the ordinance, not in a model.

If a professional council or a conducting body later publishes a live proctoring instruction, pin it and amend. Until then, do not let a vendor PDF become the missing instruction. The 2018 integrity regulations and your unfair-means ordinance are still the texts you can show a court.

Objections you will hear — and what to do with them

These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.

Cheating exploded after COVID. We have no choice.

You have choices: better in-person invigilation, better question design, oral components, honour codes with real hearings, seating agents, evaluator split. A camera in a bedroom is not the only remaining instrument.

The vendor says accuracy is 99 percent.

Ask accuracy on your students, by gender, disability and network quartile. Ask who labelled the training set. Ask what happens to the one percent in a batch of five thousand. Then ask counsel whether a 99 percent accuser is still an accuser.

Students already accept the terms at login.

A mandatory exam is a poor place to claim free consent. Necessity and ordinance, plus DPDP analysis, have to do the work. A clickwrap is not a hearing.

Without AI we will be seen as soft.

A syndicate that cannot explain a model to a court will not look hard. It will look careless. Hard is a competent committee and a reconstructable file.

A pause-and-write month before any proctoring buy

If a tool is already on, do the same month and turn features off until the ordinance exists.

1. Week 1: read the live unfair-means ordinance and any council instruction. List gaps: evidence, hearing, network failure, disability.
2. Week 2: decide whether you need a new tool at all. Price better invigilation against the SKU.
3. Week 3: if you proceed, ban emotion and gaze-as-guilt. Require exportable clips and a human charge sheet.
4. Week 4: student notice, exemption path, deletion schedule. Then, maybe, a limited pilot on a low-stakes internal test — never first on a degree exam.

How this shows up in the file

Controller's note: the model is not a member of the unfair-means committee. No emotion scores. Clips live in our store. Students see what we rely on. Minors and room scans have a separate counsel paragraph.

If that note cannot be signed, the purchase cannot be signed.

What the next file must contain

“Exam Malpractice Detection: Ethics First” earns a line in the noting only if a P3 University can attach proof of “exam proctoring AI ethics.” A heading is not proof. A vendor slide is not proof. A workshop photograph is not proof.

Write three dated sentences: what was decided, who owns it after the next posting order, and when it will be re-checked. If you cannot write the three sentences, you are not ready to buy, to sell, or to go live.

Leave unsourced percentages out of the note. DPDP is not a blanket localisation statute. The November 2025 AI governance text is guidance, not an Act. CERT-In’s 28 April 2022 directions still set specified incident and log clocks. A PAC, when lawful, lives in GFR Rule 166.

• Name the designation that owns “exam proctoring AI ethics.”
• Attach one artefact a stranger can open next year.
• Record the instrument you are actually using.
• Revisit when the model, the SI, the notice or the posting changes.

This article is informational field guidance for Indian public universities and colleges, not legal, regulatory, academic, labour or procurement advice. Confirm against the live AICTE / UGC / NAAC notice, the DPDP Act and Rules as commenced, your Act and statutes, reservation orders, labour counsel and the syndicate before you file it.