Education
Designing Assessments AI Cannot Shortcut: Process, Reasoning, Oral Defence

Education has always rested on an old bargain: a finished piece of work is evidence of the mind that produced it. Generative AI has broken that bargain. An algorithm can now write a coherent essay, solve a complex equation or produce clean code in seconds, so a polished submission tells us almost nothing about the student who handed it in. If a learner can use a tool to secure top marks, we have to ask honestly what our grading systems are measuring.
Better AI detection software will not fix this. Those tools remain unreliable. What we need is a deliberate redesign of how and what we assess. The principle is simple. Stop grading artefacts a machine can produce, and start grading the reasoning a person can show in real time.
Why a finished essay no longer proves understanding
For decades, assessment rested on outputs: the essay, the report, the case summary, the program. We treated written eloquence as a fair proxy for understanding. That proxy has collapsed. Advanced AI now writes with more clarity and speed than most literate people, so written communication on its own can no longer be a core human benchmark.
What remains distinctly human is the ability to pull thoughts together, defend decisions and adjust one's reasoning under live questioning, without a digital crutch. Assessment in the AI era should be built around exactly that. Otherwise institutions are protecting old metrics that no longer signal anything.
AI has stripped education to its absolute core. It can synthesize data, but it cannot mentor.
Four assessment designs that resist shortcuts
1. Bring back the viva voce
A much larger share of the grade should come from in-person oral defences. A student should stand before faculty or peers and unpack, dissect and defend an argument when challenged. Doctoral programmes have relied on this rigour for a long time. Why keep it out of undergraduate classrooms?
2. Use live diagnostics that test execution
Give students something that is broken on purpose: faulty code, a flawed mathematical proof, a weak historical argument. Ask them to troubleshoot it on the spot, with no digital help. This tests the process, and it mirrors how problems are actually solved at work.
3. Make students audit the machine
Banning AI tools is fruitless. A better assignment asks students to critique AI output: spot factual hallucinations, verify sources and explain where the machine's reasoning breaks down. The same idea works in professional selection. Hand a candidate a highly polished but logically flawed AI answer and you can measure what I call epistemic vigilance, the habit of not trusting a confident answer simply because it sounds right.
4. Grade to mastery with a Complete/Revise loop
Partial-credit systems tend to reward compliance. An edit-to-mastery contract works differently. Major assignments run on a Complete/Revise loop, and a submission is marked Complete only when it meets strict professional standards published in advance. Work that falls short goes back for rewriting. An A then means the student has shown mastery, and points alone cannot earn it.
Assessing how a student works with the machine
The bigger change is from testing static knowledge to testing how a person works with a machine. I call this system choreography. Two people can use the same AI model and submit results that work equally well. One accepted the first draft without scrutiny and left a hidden flaw in place. The other questioned the tool, tested edge cases and corrected the critical parts by hand. A traditional assessment grades them as equals. A process-focused one sees that only the second student truly understands the work.
In practice, this means looking at how a student questioned the tool, where they caught its errors and how they explain their corrections in an unaided oral review. The record of their reasoning tells us more than the final page does.
What teachers still have to do
Grading used to be hard because teachers had to separate conceptual understanding from ordinary human error. Today the harder task is separating a student's own thought from machine output. That is a philosophical problem more than an administrative one, and it means classrooms must be redesigned to build deep, focused intellect. Processing automated submissions is not the job.
This is where educators matter most. They notice each learner's trajectory, challenge assumptions and hold students to a standard higher than an algorithm can copy. Indian universities, guided by the direction set in the National Education Policy 2020, have a real chance to turn this disruption into lasting reform.
Conclusion
Assessments that AI cannot shortcut have one thing in common: they make thinking visible. Oral defence, live troubleshooting, auditing AI output and mastery-based revision all ask students to show their reasoning along with their results. Institutions that adopt them will give their grades meaning again. More important, they will produce graduates who can direct machines instead of depending on them.


