From AI-Free Gates to AI-First Work: How Engineering Students in India Navigate University Education and Campus Placements Without AI Guidelines

Synthesis: In a Global South setting that research on AI-mediated technical work has rarely sampled, Goutham and Gu (2026) interviewed 23 current students and recent graduates of Indian engineering and computing programs, 13 still enrolled, between September and October 2026 about how they used Generative AI across coursework, semester examinations and campus placements. Reflexive thematic analysis produced five themes tracing a single route from study into work, and that route runs between two incompatible regimes: placement screens that mostly excluded AI, and workplaces graduates described as AI-first, where deliverables that once took a month were expected in one or two weeks. All but one participant used AI to prepare for placement, largely as a self-directed coach, while the CGPA those screens filtered on was itself partly produced with AI. Drawing on signaling theory, they argue that grades and screen scores certify less — to employers and to students — once AI can help produce them, even as 15 of 23 participants said they would believe a professor or textbook over the model when the two disagreed.

Key Findings

1. Two regimes inside one degree. All but one of the 23 participants used AI to prepare for placement screens that excluded it, while graduates described workplaces that required it from the first day.
2. Clearing over learning. Participants generated exam-ready answers trimmed to the mark scheme and named the trade-off themselves: "we clear the exams but not the learning part."
3. Deference follows the grader. 15 of 23 participants said they would believe a professor or textbook over AI when the two disagreed, and several explained this by who awards the marks.
4. Policing without policy. No participant could point to an institutional policy; enforcement ran through checker limits (one university set about 5%), detection software and network blocks on supervised tests only.
5. Coaching you had to buy. Access came from company student promotions and telecom bundles, not institutions; the newest model subscriptions were put at 15,000 to 20,000 rupees.
6. Ownership as the cost. Participants described dependence and work they could not fully claim: "you don't go with the same ownership confidence that you used to do."
7. A split over guidance. Eight participants wanted AI guidance from professors or their university, five saw little value in it, and the split tracked confidence with the tools.

Two regimes inside one degree

India holds an approved intake of nearly 2.9 million engineering and technology seats for 2026–2027, and the route into the technology sector runs through gates that predate generative AI: entrance examinations or 12th-grade board marks, semester examinations marked against a professor's notes, and campus placements that screen on CGPA cut-offs, aptitude and coding tests, and interviews. GenAI arrived without institutional frameworks for its use. Because most graduates had entered college around 2022, the norms moved inside a single degree: P9 remembered language models as "gimmicky toys" and AI-written code as "a kind of a cheaty way to go about things," while P4 said that "by the third or fourth year, AI basically took over everything." At P9's employer, "we're meant to do it AI first," with deadlines that once ran to a month now at one or two weeks. Many participants concluded that placement screens tested something other than the job they were applying for.

Clearing, not learning

Semester examinations were closed-book, handwritten and marked against the professor's own notes, which made the expected answer predictable and, in participants' accounts, rehearsable. AI made the rehearsal cheap: participants summarized chapters, forecast likely questions (P17 reported that "most of the time, uh, the questions given by chat GPT will appear in my examination"), sized answers to the two-mark and five-mark schemes, and stripped out explanation so that material could be mugged up. Participants did the work themselves in two cases — numerical subjects, where AI's answers were often wrong, and assessments they expected to defend aloud. AI also tutored where teaching assumed knowledge students did not have, so scoring and learning ran as separate tracks, the second mostly outside the assessed curriculum. The Academic Integrity problem is not smuggling but a grading system whose top grade can be produced without the understanding it certifies.

Rules that were never stated

Upward, the relationship was steep. Professors set the questions, awarded the marks and decided course by course whether AI could be used, often without saying so. No participant could point to an institutional policy, and where institutions acted it was through numerical limits on AI-checker scores, detection software and network blocks, measures that covered only supervised examinations. Participants read the same silence in opposite ways, as tacit permission or as risk, and the cautious ones defaulted to concealment. Detection was met with evasion, most often asking the model to "humanize" its output so that it passed the checker, or to "make it sound more Indian." Sideways, among peers, norms were flat and open: prompts, whole chats and news of model releases circulated freely, and in P8's cohort a class representative consolidated an examination's material into one shared ChatGPT conversation an hour before it began.

What readiness became

The burden of defining readiness fell on students. AI became the first response to almost any task, and many participants felt their own thinking receding. P2 said that "in using AI, I'm shrinking my capacity to learn," and P14 typed in even problems they could already solve: "It's simple, but I'm not ready to think." The cost they tied most directly to work was ownership: P7 could "sleep peacefully that the work is completed" but doubted they could answer questions about it. Signaling theory gives this a shape — a CGPA above a cut-off separates candidates only if it is costlier for the weaker ones, and AI-assisted clearing makes it cheaper for everyone — which is why participants welcomed AI-permitted rounds as closer to the work itself, and why accounting for your own work registered as a cost to Learner Identity.

What this means for practice

• Instructors. State your position on AI for every course and every task in writing: participants read the same silence as permission or as prohibition, and 15 of 23 deferred to whoever awarded the marks.
• Instructors. Grade AI-assisted work through an oral defense. Participants invested in understanding where they expected to explain their work in person, and feared being unable to answer for what a model wrote.
• Administrators. Negotiate AI access for students instead of leaving it to company promotions and telecom bundles: no participant's institution supplied tools, and the newest model subscriptions were put at 15,000 to 20,000 rupees.
• Employers. Test candidates working with AI as well as without it. Participants called AI-permitted build tasks "more of a Reflection of what's going on in real life," and treated CGPA cut-offs as weaker evidence than they once were.
• Researchers. Interview recruiters and hiring managers before concluding that the signals have weakened. This study hears only the student side of the gate.

Limitations

• The 23 participants were recruited by purposive sampling supplemented by snowball referrals, which the authors note may explain the sample's clustering in a few institutions. All studied at established institutions of good standing, mostly in Chennai, so rural colleges, less-resourced institutions and regional-language-medium programs are absent.
• Only two participants were women, so the accounts largely reflect men's experience.
• All accounts are self-reported recollections of AI use, and graduates' descriptions of college were retrospective; no placement outcome, recruiter decision or employer assessment was observed.
• No faculty, administrators, recruiters or managers were interviewed, so institutional positions appear only through students' perceptions, and the claim that screens "do not hold" rests partly on participants' own estimates rather than on proctoring data.
• The first author graduated in computer science in 2024 from one of the institutions represented and works as an AI consultant; she conducted all interviews and coded the transcripts, with reflexive memos and regular consultation with the second author as the stated check.
• Interviews ran September to October 2026, most lasting 30 to 45 minutes (some Tamil–English ones about 20), six partly or mainly in Tamil and translated by the first author, who notes that some nuance may have been lost; the themes are interpretive constructs, not frequency counts.

Citation

Goutham, S., & Gu, Z. (2026). From AI-Free Gates to AI-First Work: How Engineering Students in India Navigate University Education and Campus Placements Without AI Guidelines. arXiv preprint arXiv:2610.29140.