The AI Talent Gap Is a Judgment Gap, and Universities Must Close It | Ashwin Krish | Director of Academics | Newton School of Technology

India will need roughly 1.4 million AI professionals this year and has about 400,000, according to a talent report released in Bengaluru last week. TeamLease Digital estimates that only about one in six of our IT professionals has any AI skills at all. Read those numbers alone and the fix looks obvious: train more people in AI.

Now read the other set. Stanford’s 2026 AI Index cites US payroll data showing employment among software developers aged 22 to 25 falling nearly 20% from 2024 levels, while older cohorts were far more stable. Industry is short of AI talent and is also hiring fewer of the young people who would become that talent.

Both findings are true, and they describe the same shift. Companies aren’t short of people who have heard of large language models. They’re short of people who can judge what a model produces: decide what’s worth building, notice when a confident answer is subtly wrong, and own a system when it fails in production. That ability used to be grown on the bottom rung of the job, in routine tickets, first services and small bug fixes that a senior reviewed line by line. Those tasks were never valuable for their output. They were the apprentices, and a model now finishes them in seconds.

So the bottom rung is disappearing, and the demand for people who have already climbed past it keeps rising.

The contract that lapsed

For decades, engineering education in India ran on an unwritten agreement with the industry. Colleges taught fundamentals; a graduate’s first two years at a company finished the training. Nobody expected a fresher to make design decisions. They expected someone teachable, and the job did the teaching.

That agreement has quietly ended. If companies no longer have junior work to train people on, a student has to graduate already standing on the second rung. The apprenticeship has moved inside the degree, whether universities have noticed or not.

The usual institutional response is to add a course: an elective on generative AI, a module on prompt engineering, a specialisation with “AI” in its name. I understand the instinct, because I’m on the side of the table that defends curricula to a university. But the tools change every few months, and a syllabus revised through a one-to-two-year approval cycle can’t chase them. What lasts is the ability underneath: reading code you didn’t write, predicting what a system will do before you run it, and knowing which of three plausible answers is wrong. Those habits get taught, or don’t, inside the courses we already run.

What it looks like in a classroom

I learnt judgment the slow way, through years of competitive programming before I ever shipped production code at CRED, and most of what I now teach is an attempt to give students that practice on purpose. In the algorithms course and programming lab at our Bangalore campus, we’ve made a few changes.

Students predict before they run. Every algorithm is traced by hand first, step by step, and only then written or generated. The prediction is where the learning happens, and running the code checks it.

AI tools are allowed, and the constraint sits in how we assess. Bans teach students to hide their tools. Our assessments ask for what a tool can’t fake: explain this trace, find the single bug in this generated solution, say what breaks when the input doubles. A student who leaned on a model all semester without building understanding finds out in a viva, early, while it’s still cheap to fix.

We also teach students to hunt for silent errors. Code that crashes teaches its own lesson. Code that runs and gives a plausible wrong answer is the kind that reaches production, so our practice problems increasingly hand students working code that is wrong and ask them to prove it.

[MARKETING TO FILL: one cleared outcome or student line, with source. Example shape: “In the first half of this semester, X% of our second-years…” Leave out if not cleared.]

None of this is new to good engineering teams; it’s how they train people. The difference is that universities now have to do it themselves.

What industry owes back

Universities can’t close this gap alone. Employers can start with their own interviews: if a candidate can pass by pasting model output, the interview is measuring the wrong thing, and debugging, prediction and trade-off questions measure what the job needs. They can keep a real entry rung, even a small one, because a company that waits for someone else to train its juniors will find no seniors in ten years. And they can come into the classroom. Practitioners who run mock interviews and review student projects change a graduate’s readiness more than any curriculum revision I can push through a committee.

The AI talent gap is usually described as a shortage of people who know AI. Seen from a classroom and an interview room, it’s a shortage of people who can tell when the AI is wrong. That is teachable, and it’s the oldest thing engineering education was meant to teach. We just can’t leave it to the first employer any more.