Walk into any faculty lounge at a Nepali college today and you'll hear a version of the same conversation: instructors are increasingly confident that a growing share of student assignments are at least partly AI-written, and increasingly unsure what to do about it. Nepal's universities built their plagiarism systems around a different problem — copying existing text — and that system is now being asked to catch something it was never designed to detect. This piece looks honestly at how widespread AI-written work has become, why detection tools are unreliable, what institutions are actually doing about it, and what assessment models would reduce the incentive to cheat in the first place.
How Widespread AI-Written Assignments Have Become
Free access to capable AI writing tools, spotty enforcement, and assignment formats that reward polished final text over demonstrated understanding have combined to make AI-assisted submissions common across Nepali campuses — from routine essays and lab reports to, in some documented cases, portions of thesis and dissertation work. Faculty members report a noticeable shift in student writing: a sudden jump in polish and vocabulary that doesn't match a student's classroom performance, oddly generic phrasing, or arguments that are technically correct but strangely shallow — all common fingerprints of AI-assisted text, even when no formal detection tool flags anything.
This is not unique to Nepal — universities worldwide are grappling with the same shift — but Nepal's relatively recent and still-developing formal plagiarism infrastructure means the gap between the problem and the institutional response is wider here than in systems that already had mature academic-integrity offices before generative AI became mainstream.
Accuracy Problems With AI-Detection Tools
Even where AI-detection tools are used, they are far less reliable than most people assume. These tools were trained primarily to flag statistical patterns typical of AI-generated text, but they produce meaningful rates of both false positives and false negatives. Non-native English writers — which describes the vast majority of Nepali students — are disproportionately flagged as "likely AI-written" even when they wrote every word themselves, because their sentence structures can statistically resemble AI-smoothed English. At the same time, students can defeat most detectors with simple paraphrasing passes, translation round-trips, or by asking an AI tool to "write like a non-native English speaker," making detection an unreliable foundation for any serious disciplinary decision.
This creates a genuinely difficult position for institutions: acting on unreliable detector output risks unfairly penalizing honest students, while ignoring the issue entirely risks devaluing every degree the institution issues. Neither extreme is defensible, which is exactly why detection tools alone cannot be the primary strategy.
What Nepali Universities Are (Or Aren't) Doing About It
Nepal's largest universities currently rely mainly on conventional similarity-checking software for postgraduate research — comparing submitted text against existing published sources and setting a maximum allowed similarity percentage. This infrastructure was built to catch copy-paste plagiarism between documents, and it does a reasonable job of that specific task. It was not built to detect originally AI-generated text that has no matching source to compare against, which is precisely the kind of content now most likely to slip through undetected.
At the course and assignment level — as opposed to formal thesis submission — enforcement is largely inconsistent and left to individual instructor discretion, rather than governed by a clear, institution-wide policy on acceptable AI use. Some departments have started informal conversations about updating academic-integrity guidelines to explicitly address generative AI, but a comprehensive, consistently enforced national policy specific to AI-assisted academic work has not yet caught up with how common the tools have become in student life.
Better Assessment Models to Reduce AI-Cheating Incentives
The most effective long-term fix is not a better detector — it is redesigning assessments so that AI-only submissions simply cannot succeed. Assessment formats that hold up well include oral defenses or short viva-style follow-up conversations where students must explain and defend their own submitted work in real time; staged submissions that require visible drafts, notes, and revisions rather than only a polished final document; in-class, supervised writing components for at least part of the grade; and applied, scenario-based tasks that require applying course concepts to a specific, unfamiliar situation rather than producing generic, well-covered essay topics that AI tools handle easily.
A parallel, and arguably more important, shift is treating responsible AI use as a skill to be taught rather than only a violation to be punished. Clear institutional guidance on when and how AI tools may be used — with attribution, similar to citing any other source — gives honest students a fair, transparent standard to follow, while making genuinely undisclosed AI submission unambiguously a violation rather than a grey area.
FAQ
Can AI-detection tools reliably prove a student used AI?
No. These tools produce meaningful false positive and false negative rates and should not be used as the sole basis for a disciplinary decision.
Do Nepali universities currently have a specific policy on AI-assisted assignments?
Most rely on general plagiarism policies built around source-matching software; dedicated, consistently enforced generative-AI policies are still developing at most institutions.
What should students do if they're unsure what AI use is allowed?
Ask the instructor directly and in writing before submitting — expectations often vary by course and are rarely written down clearly anywhere else.
Are oral defenses realistic for large Nepali classes with hundreds of students?
Full individual defenses may not scale everywhere, but short, randomly-sampled follow-up conversations on a subset of submissions can meaningfully deter undisclosed AI use without requiring universal one-on-one sessions.
Disclaimer: This article is for general informational purposes only and does not constitute legal, financial, or tax advice. For decisions involving contracts, payments, or tax obligations, please consult an ICAN-registered Chartered Accountant (CA).
Discussion