By — Amritesh Unny
Abstract
A single number can now fail a student, an AI detector returns a percentage, and a committee reads it as a confession. It happened at Lucknow University and at Jindal Global Law School. No court in India has yet said what such a number proves. When does a tool that assists become one that replaces? And who carries the risk when the software is a black box that misreads second-language English as machine output? Fair procedure under Articles 14 and 21, and the rulings surfacing globally help us locate where India must land. The wrong is not the machine; it is the opaque nature of AI detection, and deep reliance on AI by students.
Introduction
At Lucknow University, research scholars say DrillBit, branded properly cited and original works as plagiarised, flagging 116 out of 121 PhD theses with 95% plagiarism, with no real way to contest the result. A similar episode took place at Jindal Global Law School, in 2024, with students claiming Turnitin wrongly flagging their take-home examination as 88% machine written.
This is not an Indian peculiarity. At the University of California, Davis, student Louise Stivers, a political-science major, was put through a two-week integrity inquiry over a brief summarizing a Supreme Court case, flagged by a Turnitin tool her campus had switched on days earlier. Her classmate William Quarterman had already failed after a professor fed his exam answers to GPTZero. At the University of North Georgia, student Marley Stevens drew a zero, a year on probation, a lost scholarship and a $105 bill for a seminar on cheating, all for a two-page paper she had run through nothing more exotic than a free spell-checker. In Britain, the higher-education ombudsman has begun overturning verdicts like these, reminding us that proving misconduct is the job of the university and not the student’s job to disprove.
Trial With No Crime
Indian law has a name for the floor a fair process owes a person, the Principles of Natural Justice. Justice Bhagwati in 1978 in Maneka Gandhi v. UOI, laid down that any procedure encroaching a person’s liberty has to be right, just and fair, and that audi alteram partem is implicit even where a statute is silent. Over the years, this has been read into Article 21. Thus, naturally, a disciplinary process that imposes civil consequences without a fair hearing is constitutionally infirm.
The UGC’s 2018 regulations measure textual overlap with existing sources. Any amount of AI detected measures something different altogether. It measures statistical resemblance to the output of a language model. To punish the latter under a rulebook written for the former, then treat it as strict liability, is not rigor. It is a category error wearing rigor’s robe.
Detector Cannot Tell, And Its Makers Know This
How does a software decide if a human wrote like a machine? Largely through three ideas. The first is perplexity, how surprising each word is to the ones before it. Human writing wanders and surprises. Meanwhile, a model trained to pick the probable next word runs predictably smoothly. So low perplexity gets read as AI generated. The second is burstiness. Humans contrast their sentences, long then short then long, while machine text keeps an even gait. Thus, uniformity in writing is most likely to be predicted as AI generated. Finally, Token-likelihood, the model checks whether each token is one a language model would have been likely to emit. That is more or less the whole science, and it is thinner than it sounds.
In 2023, a Stanford team led by James Zou ran seven popular detectors over essays by non-native English writers. 61.3% of essays got misclassify as AI-generated, while flagging almost none of the native speaker’s essays. The cause is the mechanism itself. A writer with a narrower textbook vocabulary and grammatically safe syntax also produces low-perplexity, low-burstiness prose. So does a law student trained to write in flat institutional English. The tool cannot separate constrained-but-human from the machine. This leaves students writing in a second or third language and drilled in formality, among the most structurally suspect authors on earth.
This is not a fringe worry, OpenAI quietly withdrew its own AI-text classifier in July 2023, admitting a “low rate of accuracy”, it had caught 26 percent of AI text while falsely flagging nearly one in ten human samples. Vanderbilt University switched off Turnitin’s detector after working out that a one percent false-positive rate, across the 75,000 papers it submitted in a year, would wrongly brand some 750 of them. When the firm that built ChatGPT cannot reliably spot ChatGPT, no examination committee should pretend it can.
THE REST OF THE WORLD IS REIMAGINING EDU-TECH POLICY
Slowly, the rest of the world is reaching an obvious conclusion. Matter of Newby v. Adelphi University annulled a misconduct finding built on a 100 percent Turnitin score, calling it “without valid basis and devoid of reason,” and ordered the record expunged. In Haishan Yang v. University of Minnesota, where the University expelled a doctoral student for AI use, the Court affirmed the expulsion. But the panel there leaned on faculty judgment, internal inconsistencies, and a full hearing, not just AI detector reading.
The European Union’s AI Act, Annex III(3), now treats systems used to evaluate learning outcomes and to monitor students during exams as “high-risk,” demanding human oversight and proven accuracy. Its data-protection law already bars decisions with “legal” or “similarly significant” effects taken solely by machine. Article 14 and 15, of the EU’s AI Act classifies such machine reporting as high-risk status which triggers obligations for risk management, data governance, transparency, human oversight, accuracy, robustness and cybersecurity. Furthermore, as per Article 22 of General Data Protection Regulation, citizens have a right “not to be subject to a decision based solely on automated processing”.
The Real Concern Is Not Ai, It Is Reliance
None of this makes artificial intelligence the villain. That framing is lazy, and the better economics says so. Nobel Laureate, Daron Acemoglu and Economist, Pascual Restrepo have argued for years that technology helps when it augments human work and invents new tasks, and harms when it merely replaces people and lets their skills wither. Carried into a classroom, the lesson is to use AI to build higher-order skills, not to automate the old exercises and then prosecute the students who reach for it.
The real danger is reliance, not use. A 2025 MIT Media Lab study wired up essay-writers and found that the group leaning on a chatbot showed the weakest neural engagement and the faintest sense of owning their own words, what the authors called “cognitive debt.” Researchers at Jindal Institute of Behavioral Sciences describe the same trap as an “illusion of learning”, fluent output laid over a hollow understanding. Professor Adithya Variath, frames the institutional stakes more sharply still, that without reform, classrooms risk becoming sites of credentialing rather than intellectual formation. That is the harm worth naming. Not that students write with machines, but that they might stop thinking while they do.
Build The Classroom, Then Change The Exam
Is imperative to note that this is not a public trial against any university in particular. It is simply a call for action. To reevaluate traditional practices and analyze the internalized fears of our current education system. Rendering student’s incapable of defending themselves against a number developed by software with opaque functioning, must be abhorred. Universities across the nation must realize that software infamous for its false positives and false negatives being utilized in a routine manner is a foreseeable miscarriage of justice.
Professor Dabiru Sridhar Patnaik, in “Rethinking higher education in the age of AI” writes:
“Artificial intelligence is reshaping the very grammar of education… With digital transformation and artificial intelligence, traditional modes of assessment are clearly under strain… Faculty reorientation will therefore be essential to this pedagogical shift.”
So, what should universities do? Firstly, they must begin with an evidentiary firewall, i.e., no student penalized on an AI detector score alone. Add real guidelines with procedure, an offence the student can find in the rulebook, with provisions safeguarding written notice of the charge and the evidence, disclosure of how the software reached its number, a hearing with the right to reply, and a reasoned, appealable decision. Universities should not develop quasi-criminal tribunals themselves and rely on something that cannot be scrutinized.
If this is brought to fruition, consequently the next step is then to redesign the assessment itself. In-class examinations, vivas, supervised components, collaborative drafts and more, so that integrity no longer rests on a guess. Students ought to be taught tools such as AI rather than hunt for it. Have students declare and annotate their AI to use and show them what outsourcing the first draft quietly costs them. Even more, the UGC should finally write the framework it has not, separating AI assistance from AI authorship, and forbidding sole reliance on detectors, before every institution improvises its own private criminal code.
The student who writes well should not have to prove their innocence to a machine that cannot read, let alone appreciate human vicissitude. Until we fix this process, the only thing these tools reliably detect is how little thought we have given to fairness.
About the Author:
Amritesh Unny is a three-year LL.B. (Hons.) student at Jindal Global Law School, O.P. Jindal Global University. He has a deep interest in human rights.
Image Source: https://lettercrafted.com/ai-vs-human-writing/