Stop Fixing India's Exams
NEET got leaked. CBSE's new marking system melted down. AI is quietly eating the jobs we cram for. We keep debugging a machine we should be replacing.
I did not take NEET. I did sit for CBSE Class 12 boards, and I did sit for JEE — badly. A few thousand ranks behind the cutoff, if I remember right. Then I sat for NID-DAT and got an All India Rank of 33.
That gap still confuses me. The same person, measured twice within a year, came out useless once and promising once. The difference was not me. The difference was what each test was actually optimized for.
Years later, hiring people who came through the same pipeline, I keep noticing the same thing: the signal is getting weaker while the noise gets louder.
So when May 2026 delivered a double failure, I did not see two separate scandals. NEET-UG was cancelled after a “guess paper” matched the real exam, and CBSE’s new On-Screen Marking system produced blurred scans and broken portals. I saw the same machine failing in two different places.
The month that broke the exam machine
NEET-UG went out to over two million aspirants on May 3. By May 12, it was cancelled. Investigators found a 410-question “guess paper” circulating weeks earlier, with around 120 chemistry questions reportedly matching the actual paper. The CBI has made arrests in the alleged leak network. The Supreme Court told the NTA it still had not learnt its lesson, which is a polite way of saying this is a sequel.
The very next day, CBSE declared Class 12 results using a shiny new “On-Screen Marking” system. Students seeking scanned copies found blurred pages, half-evaluated answers, portals crashing under load, and physics marks that made no sense. The contract went to a small Hyderabad vendor that underbid TCS, about 66 days before the nationwide rollout. Teachers who tested it said it needed another year or two. It got zero. Over four lakh students have since applied to see their own scanned answer sheets.
One exam got cheated. The other got “modernized.” Both broke.
And the response, predictably, is a queue of fixes. Tighter security. Biometric verification. Computer-based testing. Better vendors. More audits. A new committee. Every single fix quietly assumes the same thing:
That the exam is worth saving.
I do not think it is.
The machine we keep patching
Strip away the outrage cycle and look at the machine itself.
We keep taping over the same machine and calling it a fix.
India runs one of the most concentrated high-stakes testing systems on earth. A decade of a kid’s life narrows down to a handful of scores on a handful of mornings. Those scores decide medicine or engineering or “general degree, figure it out later.” The stakes for jobs, status, and family honour are all loaded onto a single number.
When you concentrate that much value into one test, the incentives become automatic. A student who actually understands physics can still lose to a student who has memorized the last ten years’ question patterns. A leaked paper is worth a fortune not because students are bad people, but because one morning has become the entire game. A blurry scan or a crashed portal does not cost you a few marks. It can cost you the year.
This is the part nobody says out loud: NEET was not cheated despite being a serious exam. It was cheated because it is the kind of exam where cheating pays for itself. The leak is not the disease. It is the most honest symptom we have.
We are cramming for a future that is already gone
The pipeline is optimized to deliver kids into a small set of “safe” careers. Doctor. Engineer. IT job. Government post. And that destination is being quietly demolished while we argue about the entrance ticket.
Look at the numbers:
- Roughly two in three unemployed young adults are graduates (State of Working India 2026, Azim Premji University). Not dropouts. Graduates.
- The graduate unemployment rate is 23.6%, more than twice the rate for this age group overall. For engineering grads specifically, it is about 25% (NCAER).
- In most countries a degree is a shield against joblessness. In India it has become a liability.
A degree is supposed to lower your odds of being jobless. In India it raises them.
Graduates are jobless at more than twice the overall youth rate. And it compounds: roughly two in three unemployed young Indians hold a degree — not dropouts, graduates.
And now stack AI on top. The IT and back-office economy, the ladder millions of families climbed into the middle class, is exactly the layer generative AI hits first: routine code, support, testing, data processing. IT majors have slowed hiring and shrunk net additions as automation spreads. Early research also suggests entry-level hiring in AI-exposed fields has softened since the LLM wave hit (GlobalSouth.AI analysis).
So here is the actual situation. We put a sixteen-year-old through two years of grinding hell, in a system riddled with leaks and run on software that cannot render a scanned page, for a shot at a “safe” career. And the safe careers are the ones being automated fastest.
The pipeline points at a safe career just as the safe work starts getting automated.
We built a brutal, riggable lottery for tickets to a destination that is being knocked down.
That is the system everyone is so eager to “fix.”
Why we test like this
Be fair to the system for a second. The exam does do one job, and it does it ruthlessly: it sorts scarcity. There are far more aspirants than medical seats, so you need a clean, defensible way to rank a million people and hand out limited tickets.
That is the real function. Not learning. Sorting.
Two million in at the top. A narrow band out the bottom. The rest is exhaust.
Once you see it as a sorting machine, everything makes sense, including why it cannot be patched into honesty. The cheating, the cramming, and the catastrophic fragility all flow from concentrating life-or-death stakes into one rankable number. You can biometric-scan the room and encrypt the question bank all you like. As long as one paper decides everything, somebody will pay to get that paper. You are not fixing a bug. You are fighting the design.
What I would build instead
I am not a policymaker. I am a developer who has built recommendation systems and seen how quickly they can go wrong. But if I were asked to sketch an alternative, I would start with one principle: stop measuring time served and start measuring what someone can actually do.
Everything else falls out of that move.
Stop measuring time served. Start measuring what someone can actually do.
1. A public skill graph instead of a single syllabus. Replace “the Class 12 syllabus” with an open, national map of competencies: literacy, numeracy, computational thinking, trades, arts, communication, and how they connect. Not one ladder everyone climbs in lockstep. A graph everyone navigates differently. Think of it as open infrastructure, like UPI for skills: a public standard anyone can build on.
One route, same rungs, same order for everyone. Miss a rung and you fall the whole way down.
2. An adaptive layer that figures out where you are and what is next. This is where AI actually earns its place, not as an exam invigilator, but as a tutor that never sleeps. It reads what you have done, infers what you know, and recommends the next problem, project, or person to learn from. We already build recommendation engines that are scarily good at suggesting the next video. Point that same machinery at learning instead of attention span.
Flip any result and the engine re-routes: more practice where you slipped, a stretch where you are ready. The same recommendation machinery that picks your next video, pointed at learning.
3. A portable, signed “learning ledger” instead of a marksheet. Every real thing you do gets logged as verifiable evidence: a project, a solved problem, a peer review, a mentor sign-off. It is cryptographically signed, owned by you. A git history of what you can do, that an employer or university can actually inspect. Not a rank on a curve. A profile with receipts.
One number, and your place on a curve. It says nothing about what you can actually do.
4. Assessment as a stream, not a one-shot. Instead of one high-stakes morning, you get hundreds of tiny, continuous signals embedded in actual work. And this is the quiet superpower: when no single test decides anything, leaking a single test buys you nothing. You cannot cram a portfolio of two years of real output the night before. The incentive to cheat does not get policed away. It evaporates.
Leak that one paper and you compromise 100% of the result. It is worth a fortune.
Leak one of two hundred signals and you compromise 0.5%. Nobody bothers.
Spread the signal across hundreds of small, continuous tasks and the incentive to cheat does not get policed away — it evaporates.
Here is the whole thing side by side:
Notice what the right column does to this month’s headlines. A NEET-style leak stops being an apocalypse, because there is no single paper to leak. A CBSE-style marking meltdown stops being possible, because there is no one-shot script to mis-scan. You do not make the exam more secure. You make the exam unnecessary.
And it maps better onto the AI-shaped job market too. A continuous, evidence-rich profile of what you can do across domains is exactly what a volatile labour market can read, far better than “94.2%, Science stream.”
And it travels. A signed, machine-readable profile is something a foreign university or an employer abroad can actually read, without you re-proving yourself through one more standardized gauntlet (SAT, GRE, IELTS, and the rest) for every border you want to cross. The one thing Indians grind hardest for, a way out and up, gets easier, not harder.
Where this absolutely could go wrong
I have shipped enough products to know the prettier the diagram, the more I should distrust it. So, honestly, here is how this breaks if you build it badly.
The digital divide eats it alive. A kid in a metro with fibre internet and three tutors will lap a kid on a shared phone with patchy data. Build this as a premium app and you do not dissolve inequality. You encode it, permanently, in the infrastructure. This has to be publicly funded, open-source, and offline-first from day one, or do not bother.
Same promise, two children, and the network quietly decides which one it reaches.
Surveillance. A system that watches every keystroke to “personalize learning” is also a data goldmine and a privacy disaster waiting to happen. Continuous assessment means continuous data collection on minors. That demands hard, legislated limits: data minimization, no secondary use, kids owning their own ledger. Not a privacy policy nobody reads.
Algorithmic bias. Auto-scoring trained on Indian data inherits Indian biases: caste, language, region, gender. An AI that quietly marks down a dialect or a name is worse than a human examiner, because it does it at national scale and calls it objective. This needs independent audits, not blind trust.
And if anyone is tempted to wave away these risks because “it is AI, it is neutral”, the CBSE OSM rollout is your cautionary tale. Bolting half-baked tech onto a high-stakes process did not remove human error. It industrialized it.
Teachers do not get cheaper. They get more important. The failure mode here is a policymaker reading “AI tutor” as “fire the teachers.” Wrong. The machine handles drilling and tracking. Humans handle the part that matters: mentoring, judgment, ethics, the messy social stuff a model cannot hold. Staffing is the investment, not the cost to cut.
I am not pretending these are small. I am saying they are design problems, hard, solvable, and worth solving, versus the exam system’s problems, which are structural and, I would argue, not solvable at all without changing the structure.
”But you cannot flip the whole country overnight”
Correct. You do not.
You run both systems in parallel for a decade. Competency profiles ship alongside the old marksheet. You pilot the new system in real districts and measure it honestly. You give employers and colleges time to start trusting the profile. The moment they trust it more than the rank, the rank quietly dies on its own. No big bang. Boring, phased, multi-year. That is a feature, not a bug; that is how you actually migrate a system a billion people depend on.
The hard part was never the technology. We have adaptive learning, signed credentials, and AI tutors today. The hard part is political will and an entire test-prep industry whose business model is the disease.
Cancel the format
We spent this month treating a leak and a software meltdown as two separate scandals to be patched. They are not. They are the same machine showing us its wiring under stress: a machine that turns learning into a riggable, fragile, all-or-nothing sort, and aims it at a future that is disappearing.
We cancelled NEET this month.
We should have cancelled the format.
Sources and further reading: NEET-UG 2026 timeline (SCC Online) · CBSE On-Screen Marking controversy (The Week) · State of Working India 2026 · Graduate unemployment breakdown (NCAER)
