Generative AI: A degree is a public promise: this person knows, understands and can do what the certificate claims. Generative artificial intelligence has not broken that promise, but it has made universities less certain that they can still keep it.
The problem is larger than cheating. Calculators altered arithmetic, search engines altered recall and online databases altered research. Universities adapted because the link between a student and the submitted work remained broadly credible. Generative AI weakens that link. A polished essay, computer programme or policy brief may now demonstrate the sophistication of a prompt more clearly than the depth of the student’s understanding.
READ | Generative AI is forcing universities to rethink assessment
This changes the central question of assessment. Universities can no longer ask only whether an answer is good. They must also ask what the answer proves about the learner. UNESCO has noted that generative-AI tools are evolving faster than regulatory frameworks and that educational institutions remain largely unprepared to validate them. At the same time, AI-detection scores cannot establish misconduct on their own. Australia’s higher-education regulator, TEQSA, explicitly says that additional evidence is required before alleging misuse of generative AI.
The consequences extend beyond employability. A degree certifies not only occupational skill but the capacity to weigh evidence, recognise uncertainty and defend a conclusion. Democracies need precisely these habits when synthetic text, images and statistics can manufacture plausibility at scale. If universities reward fluent output without examining the reasoning behind it, they may graduate people who can produce answers but cannot judge them.
The result is epistemic dependence: professionals and citizens become increasingly reliant on systems whose errors they are unequipped to detect. In medicine, law, engineering or public administration, confident simulation is not a harmless shortcut; it can become a material risk. The educational objective, therefore, is not to preserve human authorship in every task. It is to preserve human responsibility for consequential judgement. That distinction should anchor India’s response.
Assessment has to establish what a student can actually do
Some universities have begun redesigning assessment rather than merely policing it. The University of Sydney has adopted a two-lane approach: secure, in-person assessments that establish unaided attainment, alongside open assessments in which students can use contemporary tools, including AI, under clearly defined conditions. The university says the two approaches are intended to operate together rather than replace one another.
READ | Generative AI and journalism: Hidden risks reshaping information
For India, the stakes are formidable. The latest All India Survey on Higher Education puts total enrolment at 4.50 crore students in 2023–24, while the number of universities has risen to 1,289. Expansion is a democratic achievement. Yet mass scale creates pressure for assessments that are cheap to administer: standardised examinations, recycled assignments and high-volume take-home submissions. These are also among the formats most vulnerable to rote reproduction or undisclosed use of AI.
This is where AI exposes a gap between policy and practice. The National Education Policy 2020 calls for assessment based on programme learning goals and says higher education institutions should move away from high-stakes examinations towards more continuous and comprehensive evaluation. UGC has also issued guidelines on innovative pedagogical approaches and evaluation reforms. But continuous assessment is not automatically credible assessment. Ten AI-assisted assignments do not establish learning merely because they replace one final examination.
India therefore needs a verification spine within every degree. Students may use AI for exploration, feedback and productivity, but programmes should identify a limited set of outcomes that every learner must demonstrate independently. Short oral defences, supervised problem-solving, laboratories, field tasks, staged drafts with revision histories, studio performances and capstones before mixed panels can test reasoning rather than textual polish. Randomly selected vivas can authenticate learning in large classes without turning every course into an administrative marathon.
Assessment should inspect the path to an answer: the assumptions made, the evidence considered, the corrections introduced and the judgement exercised. The final product alone is increasingly inadequate evidence of learning.
READ | Generative AI trust crisis grows amid booming adoption
Verification cannot become another source of inequality
Oral assessment, however, is no magic shield. Poorly designed vivas can reward confidence, accent and social familiarity while disadvantaging anxious students, first-generation learners and those answering outside their strongest language. Verification therefore needs published rubrics, multiple modes of assessment, brief examiner training, reviewable decisions, disability accommodations and, wherever feasible, Indian-language options.
This matters because a reform intended to restore trust in assessment could otherwise reproduce existing inequalities under the name of rigour. A student who is less fluent in English should not be presumed to understand less; nor should confidence in an oral examination be mistaken for mastery of a subject.
The deeper danger is institutional. If employers begin treating degrees as weak signals, students with greater financial and social resources will be better placed to acquire private credentials, interviews and networks to prove competence. Others will bear the cost of declining trust. An assessment problem could then become another mechanism of inequality.
Universities do not need to prove that every sentence submitted by a student was written without technological assistance. They do need to establish that the graduate who receives the degree possesses the knowledge, skills and judgement that the qualification represents.
AI should be welcomed as a tool, disclosed as an influence and bounded as a substitute. The test of a university degree in the AI era is ultimately whether the institution can still establish who has learned what the certificate says they have learned.
READ | Generative AI-driven SaaS challenges legacy software giants
Debdulal Thakur is Professor, Department of Economics, Alliance University, Bengaluru.