When the exam hall and the office disagree
A student can pass an Indian engineering examination with a chatbot at their elbow and then walk into a job where that same chatbot is expected to write the first draft of everything. The credential says one thing; the workplace asks for another. This is the finding at the center of a 2026 interview study of 23 engineering and computing students and recent graduates in India, published as From AI-Free Gates to AI-First Work: How Engineering Students in India Navigate University Education and Campus Placements Without AI Guidelines. [1] The researchers, using reflexive thematic analysis, traced a single mismatch through five themes: the gates students pass through were built before generative AI and still assume its absence, while the work beyond them assumes its presence. The human capacity that goes missing in that gap is not coding or recall. It is the ability to stand behind a piece of work as one’s own — to explain it, defend it, and know why it is right. That capacity is exactly what the old assessments were supposed to certify, and exactly what AI-assisted clearing no longer certifies.
What the interviews found before anything else
The study’s first theme is blunt. Placement screens largely excluded AI, while graduates described workplaces that expected it. India’s institutions have an approved intake of nearly 2.9 million seats in engineering and technology for 2026-2027, according to the All India Council for Technical Education, and the route into the sector runs through semester examinations, a cumulative grade point average, and campus placement drives in which each company sets its own minimum CGPA, online aptitude and coding tests, and interviews. Almost none of these gates permitted the tools the destination required. The second theme: participants prepared for placement through self-directed study, with AI as their main coach — all but one participant used AI to prepare for placements or job interviews. The third: they used AI to clear coursework whose grades the screens later filtered on, a practice they distinguished from learning, even as AI also helped them learn where teaching fell short. The fourth: faculty authority and unstated rules shaped how openly they used AI, while peers taught one another how. The fifth: participants worried that dependence on AI, and work they could not fully claim as their own, would leave them unready for the jobs they sought.
The signal that stopped separating anyone
The authors read these gates through signaling theory. An employer cannot observe a candidate’s productivity directly, so it infers it from observable signals such as a degree, a grade, a test score. A signal works only if it is costlier for weaker candidates to acquire. The study’s argument is that AI has made the signal cheap for everyone. Participants used AI to forecast likely questions across fifteen years of past papers, to compress chapters into answers sized to the mark scheme, and to keep their CGPA above company cut-offs. One participant described the result as clearing the exams “but not the learning part.” [1] If a CGPA can be maintained and a screen passed with substantial help from AI, then the score tells an employer correspondingly less about the skill it was meant to indicate. The judgment the grade once stood in for — whether this person can reason through a problem alone — has been quietly replaced by a number that no longer measures it.
The safeguard that guards the wrong thing
Closed-book, invigilated, handwritten examinations are structural safeguards AI cannot enter. The study’s participants still used AI upstream, during preparation, because the examinations rewarded answers that could be predicted and rehearsed. The safeguard secured the conditions of assessment without securing its object. The same structure reappears at the placement gate, in proctored tests whose content can be prepared for or circumvented with AI. By one participant’s estimate, using AI to cheat in placement tests was routine. A signal protected only at the moment of measurement can still be produced elsewhere — and what gets produced elsewhere is an answer without the reasoning that should have accompanied it.
The escalation that may compound it
When a signal weakens, the theory predicts employers will demand stronger ones. Participants reported interviews that had become a lot harder than what the actual job might be, dominated by data structures and algorithms problems they did not expect to use. Harder screens of the same kind raise the cost of the signal without restoring its link to the job, and they intensify the AI-coached preparation that produced the problem in the first place. The judgment being tested drifts further from the judgment being hired for. Meanwhile, access to the best coaching depended on money: free tiers sufficed for coursework, participants said, but not for placement preparation, and subscriptions to the newest models ran to 15,000 to 20,000 rupees, beyond what most students could pay. Where a test-based signal loses force, the study suggests as a hypothesis, household resources can re-enter selection.
The tutor that answers before the student thinks
Participants were candid about what the arrangement left them ready for. AI had become their first tutor, and for many it had also become the first response to any task, however small. One participant typed every task into the chat window without thinking it through, describing an almost loss of struggle. Another applied AI even to problems they knew they could solve: “It’s simple, but I’m not ready to think.” [1] A third said that even when they knew the answer, they would not trust themselves. The study connects this to research warning that skills left unpractised on routine tasks may deteriorate. What deteriorates here is precisely the human judgment the gates were meant to certify — the habit of forming an answer before consulting a machine.
The skill nobody assessed
The checks participants experienced as reaching understanding were oral, and the costs they tied most directly to work concerned explaining, defending and owning work produced with AI. One participant observed that students presenting AI-assisted projects in placement interviews could not fully account for them. The study’s authors argue that in workplaces where much code and many documents begin with AI, the capacity to evaluate and stand behind that output may be closer to the core of the job than the unaided recall that most screens still measure. The gate tests the wrong human faculty. It asks for recall and rewards rehearsal; the job asks for ownership and rewards judgment. The graduate who has optimized for the first arrives without practice in the second.
The hierarchy that decides without saying
Faculty authority shaped how openly students used AI, and the study found the relationship steep. Professors set the questions, awarded the marks, and decided course by course — often without stating it — whether AI could be used. Many accepted it, but few changed how they taught or assessed in response. Where institutions acted at all, it was through detection. Asked whom they would believe when AI and a professor gave different answers, 15 of 23 participants chose the professor or the textbook, and for some the reason was not expertise but grading power:?” One participant recounted a viva in which a teammate mentioned ChatGPT, after which the professor called in other faculty, mocked the group, and demanded an apology letter from the students’ parents. The student felt unable to object, reasoning that a complaint could cost them the rest of the semester. In this setting, silence did not read as permission; it pushed the cautious toward concealment. The people best placed to define what unaided understanding looks like declined to state a position, and the task of judgment fell to students who had no authority to exercise it.
Peers as the only curriculum
Among peers, AI use was open and routine. Students shared prompts, outputs and entire conversations; in one class, the representative consolidated everything needed for an examination into a single chat an hour or two before it began and shared it with everyone. Knowledge of tools travelled the same way, as classmates passed on news of model releases and a working sense of which model suited which task. Where research on North American students found isolation and shame around AI use, these participants learned from one another what their institutions did not teach. But peer networks spread AI use unevenly, and so did money, language and confidence. The participants who navigated most effectively saw themselves as motivated beyond the syllabus and confident enough with AI to know when to stop. Those who wanted guidance more often spoke of dependence or of struggling to say what they wanted from a model. When institutions and employers leave readiness undefined, the study concludes, the students who arrive best equipped are the ones most likely to acquire it — and the gap does not fall to those who feel it most.
The ownership that goes missing
One of the study’s sharpest observations concerns what a grade earned with AI’s help certifies to its holder. It tells the student as little as it tells an employer. Several participants could no longer tell how much of their own work they understood; some checked answers they had already reached for reassurance. The loss of ownership appeared most sharply at the moment students were asked to account for their work in placement interviews. This is the human role AI renders superfluous in the study’s account: not the person who writes the code, but the person who can vouch for it. The gates never asked for that vouching in a form AI could not supply, and so the capacity atrophied in the one place it should have been trained.
What the authors propose, and what it costs
The study’s recommendations follow the diagnosis. Employers could assess how candidates work with AI as well as what they know without it, test problem solving by asking candidates to explain, critique and extend an AI-produced solution, treat CGPA cut-offs as weaker evidence than they once were, and state plainly what readiness involves. Institutions could require professors to state their AI position per course and task, reward reasoning over reproduction through vivas and explanations of submitted work, and negotiate AI access for students rather than leaving it to company promotions and what families can afford. Tool designers could ask students to commit to their own answer before seeing the model’s, making the provenance of AI-assisted work visible. The authors concede these changes are easier to propose than to make; they require access across very different institutions, literacy for faculty as well as students, and slow curricular change. In the meantime, students bear the cost.
The variable that decides whether any of it works
The study’s sample limits how far its findings travel: students at less-resourced institutions, at rural colleges and in regional-language-medium programmes were not represented, only two participants were women, and all accounts were self-reported. Faculty, recruiters and managers were not interviewed, so their positions appear only through students’ perceptions. What remains open, and what the authors flag as the question that matters, is whether employers and institutions will define readiness for AI-first work and assess it directly — or whether that definition will keep being assembled, unevenly and often with misgivings, by students on their own. The human judgment at stake is not a fixed quantity that AI either preserves or destroys. It is a practice, and practices survive only where something asks for them. The gate that cannot see the machine is still the one place many students are ever asked to show what they can do unaided.