What an AI-Graded Answer Script Actually Looks Like
Every AI evaluation vendor in India will tell you there is a human in the loop. Almost none of them will show you what that actually means.
So here are two real answers, graded by our system, with the numbers it produced and the decision it made about each one. No screenshots of a dashboard. The actual output.
Both are from an internal test run on past examination scripts. No student or institution is identifiable, and no marks from these runs were ever issued to anyone.
Answer One: A Clear Case
A one-mark question asking the student to name the alternate term for cataclastic metamorphism. The student wrote:
Ans-1) a) Fault zone metamorphism
The system awarded 1 out of 1, against the rubric point "correctly identifying cataclastic metamorphism," with the evidence recorded as: Student correctly identified 'Fault zone metamorphism' as an alternate term for cataclastic metamorphism.
Confidence: 0.97. Flagged for human review: no.
This is the routine majority. Legible handwriting, an unambiguous rubric point, a clear-cut call. An evaluator reading this answer would spend three seconds on it and reach the same conclusion. There is no good reason for a human to spend those three seconds, several thousand times, across a batch.
Answer Two: The Case That Matters
A three-mark question asking for an inverse Laplace transform. The student's handwriting was substantially illegible. Here is what the system was able to read:
Q-4 (a) find Laplace transform.
F(S) = 1 + 4/(S+3) - 5s/(s^2 + 25)
F(S) = [unclear] / [unclear] - [unclear]
F(S) = [unclear] / [unclear]
F(S) = [unclear]
The system awarded 0 out of 3. Its stated reasoning: the student appeared to attempt a Laplace transform rather than the inverse transform that was asked for, and the subsequent working could not be read clearly enough to identify any correct components.
That may well be right. It may also be badly wrong — there could be correct working underneath handwriting the system could not decipher.
This is precisely the answer that generates a revaluation application. A zero on a three-mark question, on a page the student knows they filled with working.
The system flagged it for human review.
The Part Worth Understanding
Here is the detail that matters more than any accuracy statistic.
On that second answer, the AI model reported its own confidence as 0.80 — fairly sure. Our system overruled it and recorded the confidence as 0.60.
The reason is deliberate. We do not trust a model's self-assessment of how confident it should be. We tested it, and found it unreliable: answers the model rated as only partially legible frequently had no illegible content at all, and the reverse happened too.
So the platform measures legibility itself. It counts the unreadable segments in the transcription as a proportion of the answer. Below 5 percent, the answer is treated as clear. Above 10 percent, it is treated as mostly illegible — and when that happens, the confidence score is capped at 0.50 regardless of what the model claims.
The review threshold is 0.75. Anything at or below it goes to a human evaluator. There is no override, no batch setting to raise it when a deadline is tight.
So the sequence on Answer Two was: the model said 0.80, the illegibility cap pulled it to 0.60, 0.60 fell below 0.75, and the answer was routed to a person with the reason recorded as "Low confidence (0.60)."
A separate rule fires on illegibility independently of the number, so an answer that is mostly unreadable reaches a human even if every other signal looks healthy.
Why We Are Showing You a Zero
It would be easy to publish the first answer and stop there. Clean transcription, correct mark, 0.97 confidence — a tidy advertisement.
The second answer is the honest one, because it is the case a Controller of Examinations actually worries about. Not "will the AI grade routine answers correctly" — it will, and so would a tired evaluator at 4pm. The worry is: what happens when the AI is out of its depth, and does anyone find out before the result is published?
On that answer, the system did not produce a confident zero. It produced a zero, marked it as unreliable, explained why in a line an administrator can read, and sent it to a human before it could become a mark.
The Question to Ask Any Vendor
Accuracy percentages are close to meaningless in this category. They depend entirely on the question paper, the subject, the handwriting, and the rubric — and no vendor quoting one will tell you which set produced it.
A more useful question: show me an answer your system got wrong, or was unsure about, and show me what happened next.
If the answer is a confidence number, ask how it is calculated and whether the model grades its own certainty. If the answer is a threshold, ask who can change it. If a vendor cannot produce a single example of their system flagging its own uncertainty, the human in their loop is decorative.
Your evaluators should not be a rubber stamp on AI output, and they should not be reading every answer either. They should be spending their judgment exactly where judgment is required — on the answers a machine has honestly reported it cannot settle.
ParikshAI is an AI-assisted exam evaluation platform built for Indian universities and autonomous colleges. Every answer is scored against your rubric with a confidence level; low-confidence answers are routed to your evaluators; every mark carries an audit trail.
Test it on an exam you have already graded — send us a past batch and compare our scores against your evaluators'. Request a free pilot at pariksha.tech