ARTIFICIAL INTELLIGENCE & HIGHER
EDUCATION
The Empty Classroom
When students outsource learning to AI and companies cut the engineers who know better, both ends of the talent pipeline fray at once. India is not watching from a safe distance.
Chuppala Nagesh Bhushan
.jpeg)
Professor Garcia’s diagnosis
was unambiguous. Students had used large language models to complete every
homework assignment, never internalising the material, and then sat
examinations in which no such crutch was permitted. The result was predictable
in hindsight and alarming in magnitude. The students had not cheated in the
traditional sense of copying a classmate’s work; they had done something
subtler and, in its own way, more damaging—they had outsourced the act of
thinking itself.
Both
ends of the pipeline
Berkeley is one data point, but
the dynamics it reveals are structural. The same technology that induced
students to stop learning is simultaneously being deployed by corporations to
reduce their dependence on experienced engineers. Alphabet, Meta, Microsoft and
dozens of smaller technology firms have, over the past 18 months, announced
waves of redundancies concentrated among mid-career and senior software
engineers, citing AI productivity gains as partial justification. The
arithmetic, presented in isolation, is seductive: if one engineer augmented by
AI can do the work of two, a workforce rationalisation looks like efficiency.
It looks rather less efficient
when viewed alongside the Berkeley data. Senior engineers are not merely faster
coders; they are, crucially, the people who know when the AI is wrong. They
hold the institutional memory that distinguishes a plausible-sounding output
from a correct one. Junior engineers, meanwhile, are emerging from universities
having used AI to satisfy course requirements without acquiring the underlying
competence those requirements were designed to build. The pipeline is being
hollowed from both ends simultaneously, and the companies spending hundreds of
billions of dollars on AI infrastructure have not yet connected these two
facts.
|
“The pipeline is being hollowed from both ends.
Companies are eliminating the engineers who know when AI is wrong, while
universities produce graduates who never learned to think without it.” |
The
Indian dimension
India has particular reason to
pay attention. The country produces roughly 1.5 million engineering graduates
annually—more than any other nation—and has staked much of its economic
ambition on translating that demographic advantage into technology leadership.
The National Education Policy of 2020 explicitly identified computational
thinking and digital literacy as core outcomes. Prime Minister Narendra Modi’s
government has invested heavily in IIT expansion and in schemes such as the
PM-SHRI programme to modernise school infrastructure. These are not trivial
commitments. They are, however, commitments that rest on an implicit
assumption: that the credentials conferred at the end of the educational
pipeline correspond to genuine competence at its exit.
That assumption is now
threatened. India’s student population is among the world’s most enthusiastic
early adopters of AI tools. A 2025 survey by EY India found that over 68% of
Indian university students reported using generative AI “regularly” for academic
work, a figure higher than the global average of 54%. This is not, in itself, a
problem. The capacity to use AI fluently is a genuine labour-market skill. The
problem arises when AI use substitutes for, rather than augments, the
acquisition of foundational knowledge. A student who uses ChatGPT to write an
essay about sorting algorithms has not learned to sort; a student who uses it
to check their own implementation has learned more efficiently than their
professors imagined possible. The distinction between these two uses is
everything, and it is a distinction that most Indian universities—and most
universities globally—have not yet built the capacity to draw.
The structural vulnerability is
compounded by India’s labour-market dynamics. A disproportionate share of
India’s technology employment is concentrated in services outsourcing—firms
that write, test and maintain software on behalf of clients elsewhere. This
sector, which employs roughly 5.4 million people and generates nearly $230
billion in annual exports, is precisely the segment most exposed to AI-driven
automation in the near term. Firms such as Infosys, Wipro and TCS have already
announced significant reductions in lateral hiring and have spoken openly about
AI replacing routine coding tasks. If the jobs that historically absorbed
India’s graduate engineers begin to contract precisely as the quality of those
graduates declines, the consequences for employment, social mobility and
economic growth could be severe.
What
is to be done
The policy response required is
neither a ban on AI tools nor an uncritical embrace of them. Both would be
mistakes. A ban is unenforceable, counterproductive and would deprive students
of skills they genuinely need; an uncritical embrace would repeat Berkeley’s
error at national scale. What is required instead is a rethinking of how
universities signal and verify learning.
Reconstruct assessment around
demonstrated understanding. The most
important reform is also the most straightforward: shift assessment weight
toward activities that cannot be outsourced. Proctored examinations, oral
defences of project work, live coding exercises and problem-solving conducted
under observation are not innovations; they are methods that long predate AI.
Their revival is overdue. If 70% or more of a student’s grade depends on
in-person demonstrated competence, the marginal value of AI-assisted homework
collapses without any prohibitive rule being required. India’s University
Grants Commission and the All India Council for Technical Education should
update accreditation standards to require institutions to specify, and justify,
the proportion of assessment that involves direct, unmediated demonstration of
skill.
Teach the foundations before
the tools. The linear algebra
anecdote from Berkeley is instructive. A student enrolled in a course with an
“open AI” policy for both homework and examinations emerged unable to perform
basic matrix operations. That student then failed the subsequent course. The
problem is sequencing: AI tools are most valuable to people who already
understand what they are doing, because such people can identify errors,
evaluate outputs and direct the tool usefully. Used by novices without foundational
knowledge, they are a shortcut that leads nowhere. India’s engineering
curriculum—set by bodies such as AICTE—should establish a principle of
“foundations first”: the first year of any technical programme should be
assessed entirely without AI assistance, building the substrate on which later,
AI-augmented learning can meaningfully rest.
Reorient industry hiring
toward verified competence. Indian
technology companies have historically relied heavily on branded credentials as
hiring proxies—IIT, NIT, BITS. If those credentials become noisier signals of
actual ability, firms will need to develop more robust assessment
infrastructure of their own. Several are already doing so: Infosys’s internal
“Infosys Springboard” programme retests and retrains new graduates before
deployment. This should become an industry norm rather than an individual corporate
decision. The National Skill Development Corporation is well-placed to convene
standards for sector-wide competency certification that sits alongside, rather
than replacing, university degrees.
Protect the
knowledge-holders. The Berkeley
episode is partly a story about what happens when experienced mentors become
inaccessible—Professor Garcia’s empty office hours are as significant as his
failing grades. In the Indian context, this translates to a concern about the
erosion of experienced faculty and senior engineers. India’s IITs and NITs
already face significant faculty shortages at the associate and full professor
levels; the private engineering college sector is considerably worse. A
teaching workforce populated primarily by junior instructors, who are
themselves recent graduates of the AI-assisted cohort, will struggle to
transmit the tacit knowledge—the intuition about why things go wrong, the
debugging instinct, the design judgement—that no language model reliably
replicates. Retention incentives for experienced faculty deserve attention from
the Ministry of Education.
Invest in AI-literacy, not
AI-dependency. There is a version of
this story that ends well for India. A graduate who understands algorithms, can
identify a model’s errors and knows when to trust an AI output is more valuable
than one who relies on AI for everything or one who refuses to use it at all.
The NEP’s emphasis on critical thinking and computational reasoning, if
implemented faithfully, is precisely the right framing. The risk is that
implementation slips into surface-level “AI literacy” courses—teaching students
to use ChatGPT—without ensuring that the deeper mathematical and logical
foundations are secure. The former without the latter produces students who are
fluent in the interface and ignorant of the substance.
The
longer view
There is a temptation to view
the Berkeley data as an American problem—a consequence of the particular
culture of a highly selective university in a tech-saturated city. This
temptation should be resisted. The incentives that led Berkeley students to outsource
their learning are not localised: they exist wherever examination pressure is
high, foundational courses are hard, and AI tools are accessible. They exist,
in other words, in Mumbai and Chennai and Hyderabad as surely as they do in
California.
India has an additional reason
for urgency that Berkeley does not. The country’s demographic dividend—the
unusually large cohort of young workers entering the labour force over the next
two decades—is a geopolitical asset only if those workers are genuinely
skilled. A demographic dividend populated by graduates who hold credentials but
lack competence is not a dividend; it is a liability. The window to get this
right is not infinite. The cohort now entering India’s engineering colleges
will form the backbone of its technology workforce in 2035 and beyond. The
decisions made now about how they are assessed, what they are required to
demonstrate, and which skills are treated as non-negotiable will determine
whether India leads or follows in the AI era.
Professor Garcia’s empty office
hours are a small and local fact. The dynamic they represent is neither small
nor local. The classroom is not merely a place where credentials are conferred;
it is where a society’s future capacity is built, one difficult problem at a
time. When students stop showing up to grapple with difficulty, the loss
compounds quietly, then suddenly.
Sources: Berkeleytime; UC Berkeley EECS
Grading Guidelines; EY India Student AI Survey 2025; NASSCOM IT Industry Report
2025; All India Survey on Higher Education 2023–24; National Education Policy
2020.
Analysis & Recommendations
This is a real and compounding crisis.
Here's how I'd break down the problem and what should actually be done:
What's Actually Happening
The chart shows three separate but
alarming signals:
- CS 10
(intro, broad audience): 35.3% F-rate — 5× the typical 7% benchmark
- CS 61A
(core CS majors): 10.6% — still 1.5× typical
- EECS 127
(upper-division): 16.8% — more than 3× the 5% benchmark
The pattern is consistent: Spring 2026 is
a dramatic outlier across all three courses, not just one. This isn't grade
inflation reversal — it's a structural collapse in demonstrated learning.
Recommendations by Stakeholder
For Universities / Departments
- Shift assessment weight heavily toward in-person,
proctored work. If 70%+ of the
grade comes from exams and live coding sessions, AI-assisted homework
becomes mostly irrelevant to outcomes
- Redesign homework itself — move from "produce a solution" to
"explain, extend, and break a given solution." Much harder to
outsource
- Mandatory oral components for projects in upper-division courses — a
5-minute explanation of your own code reveals understanding instantly
- Restore office hours as a grade signal — participation in guided problem-solving
sessions could count toward grades, reversing the empty-office-hours
problem Garcia observed
- Differentiate between AI literacy and AI
dependency — teach students when
and why to use AI, with explicit no-AI foundations first
For Students
- Understand the compounding trap: AI does the
homework → you pass the course → you fail the next one → you graduate unable to do the job → you get fired in year two
- The credential without the skill is a delayed
failure, not a success
- Use AI the way you'd use a calculator in a math
class — only after you understand what it's computing
For Companies
- Stop cutting senior engineers to fund AI tooling — you're eliminating the people who know when the
AI is wrong
- Reintroduce skills-based technical interviews
rather than credential screening, since degrees are becoming weaker
signals
- Invest in internal training pipelines rather than
assuming universities will fix this quickly
For the Industry / Policymakers
- Academic integrity policies need updating —
"open AI" policies on homework and exams (as the linear
algebra example shows) are pedagogically reckless
- Accreditation bodies should require demonstrated
competency standards, not just course completion
The Core Tension
The real problem isn't that AI exists —
it's that the incentive to use AI to pass courses is perfectly rational in
the short term and catastrophic in the long term, and current university
structures haven't closed that gap. The fix is making demonstrated
understanding unavoidable, not just discouraging shortcuts.
The workforce erosion at both ends —
senior engineers cut, junior engineers underprepared — is a genuine structural
risk that won't self-correct without deliberate intervention from universities,
companies, and students simultaneously.
Comments
Post a Comment