A ₹40 Lakh Budget, a Mobile-Only Cohort and an AI Proctor
A state open university with 12,000 active learners in a two-year PG programme plans to make terminal examinations online and to adopt a single learning management system. A learner survey found that 64% access the internet only through a 4G mobile handset, the median monthly data spend is ₹250, 22% share a device with others, and three districts report frequent outages - 1,800 learners live in one of them. Two vendor quotes are on the table: AI-based remote proctoring with face detection at ₹180 per learner per examination, and an LMS at ₹95 per learner per year that requires 2 Mbps per concurrent video session. The ICT budget is ₹40 lakh for the year and must also cover content development, translation and the helpdesk. A 400-learner pilot last year saw 18% of candidates lose connection mid-examination, and an internal audit found that flagged 'suspicious behaviour' was 2.6 times more frequent for women candidates, mostly because they sat in low light or with a veil.
1. Cost and feasibility: what does the proposal consume from the ₹40 lakh, what is left for everything else, and which components would you keep, modify or drop?
Read the worked answer
The arithmetic. Proctoring: 12,000 × ₹180 = ₹21.6 lakh. LMS: 12,000 × ₹95 = ₹11.4 lakh. Together ₹33 lakh, or 82.5% of the budget, leaving ₹7 lakh - about ₹58 per learner - for content development, Hindi translation, helpdesk staffing, accessibility support and connectivity mitigation. That is not a plan; it is a licence purchase with a residual.
Feasibility. A 2 Mbps-per-concurrent-video requirement is incompatible with a mobile-only, shared-device, data-capped cohort - and the last pilot already produced an 18% mid-examination disconnection rate, which on a 12,000-learner examination means roughly 2,160 disrupted candidates per cycle.
Keep. A learning management system, but re-specified: low-bandwidth mode, text-first content, downloadable packs under 50 MB per unit, audio summaries, asynchronous discussion with email digests, and no mandatory live streaming. Re-quote or use an open-source platform on university or national cloud infrastructure; the ₹11.4 lakh licence is the first line to renegotiate.
Modify. Examination mode: centre-based online testing on a local network at study centres for the 1,800 learners in the outage district and for anyone who opts in, with the online-from-home route available only where connectivity is demonstrably adequate.
Drop. Blanket AI proctoring for all 12,000. At 54% of the budget it buys the least pedagogical value and the most risk.
2. Is AI-based remote proctoring defensible here? Assess the fairness and data-protection issues and propose an assessment design that meets the same integrity goal without it.
Read the worked answer
Fairness. The audit is decisive: flags were 2.6 times more frequent for women candidates, driven by lighting and clothing rather than by misconduct. A system that generates adverse findings at different rates for a protected group, on grounds unrelated to the behaviour being detected, produces disparate impact. Every flagged candidate needs a human appeal, and the base rate matters - if the true misconduct rate is 1–2%, most flags will be false, and the cost falls on the honest.
Data protection. Face images and facial templates are sensitive personal data. Under India's data protection law the requirements are: a clear consent notice stating purpose, a lawful basis with explicit consent for biometric processing, strict data minimisation (verification in memory rather than storage of templates), a defined retention period with erasure, a grievance officer, and a breach process. Recording a learner's bedroom for three hours is not the minimum necessary for the purpose.
A design that meets the integrity goal differently.
- Open-book papers built on application, analysis and case reasoning, where copying is worth little because the answer is not in the text.
- Question randomisation from a large bank plus per-candidate parameter variation in numerical items.
- Centre-based supervised online examinations on a local network where connectivity is poor.
- Post-hoc authenticity checks: plagiarism detection, a short viva for a random 5% and for any anomalous script, and a statistical screen for improbable answer patterns.
Integrity comes from assessment design and sampled verification, not from universal surveillance. Reserve supervised conditions for cases where a grade is challenged.
3. Set out a phased rollout and the equity safeguards, with two measurable indicators you would report to the Vice-Chancellor after the first cycle.
Read the worked answer
Phase 1 (months 1–3): build small and offline-first. One course on the LMS in low-bandwidth mode; offline packs distributed through study centres; a helpdesk in Hindi and English with a named coordinator per centre; training for the 40 centre coordinators. No examinations yet.
Phase 2 (months 4–6): test at scale. A full mock examination for all 12,000, load-tested for concurrency; publish the technical failure protocol in advance; a 48-hour window in which a candidate may switch from home to centre mode without penalty.
Phase 3 (months 7–12): live examinations with a dual mode. Publish the appeal route with a named officer and a stated turnaround (say, 72 hours); do not charge a learner for a technical failure; hold a reserve sitting for affected candidates.
Equity safeguards. Device and connectivity declaration at registration, with automatic eligibility for centre mode; content available as downloadable packs so data spend does not decide grades; no mandatory synchronous attendance; translated interface and notices; shared-device users given a private submission slot rather than being flagged for a second login; and an explicit assurance that no automated flag results in adverse action without human review.
Two indicators to report.
- Assessment completion rate, disaggregated by district, gender and device type - the gap between the outage district and the rest should be under five percentage points.
- Technical failure and grievance rate: proportion of candidates unable to complete an examination for technical reasons (target under 2%, against the 18% pilot), with average resolution time.
A third, if permitted: the flag rate by gender, which must not differ materially - if it does, the integrity tool is the problem.