In the ordinary Indian schooling experience, the most respected teacher is rarely the one with the most facts. It is the one who, when a student arrives with a half-formed problem, refuses to hand over the answer and instead asks three sharper questions. By the third question, the student has the answer themselves, and they have something more valuable than the answer: they have the method for arriving at it.
This is the guru pattern. It predates the gurukul tradition as a formalism and survives it as a practice. You can find it in the Upanishadic dialogues where a teacher answers a metaphysical question with a question. You can find it in any small-town tuition class in Pune where the maths teacher refuses to do the next step. You can find it, somewhat denatured, in good case-method instruction in Indian business schools.
The dominant AI design pattern of the last four years is the exact opposite. Ask, receive an answer, move on. The oracle pattern. The oracle pattern is fast, satisfying, and corrosive over time. It atrophies the user's own reasoning, hides the model's uncertainty, and makes the conversation transactional rather than developmental.
We think the guru pattern is the better target.
Why the oracle pattern is the default
Oracle behaviour is the default for two reasons. First, it tests well in narrow benchmarks. If the user asks a factual question and the model returns a correct answer, that is a tick on the evaluation sheet. Second, it is what users say they want in product surveys. Ask anyone if they would prefer a chatbot that answers immediately versus one that asks them three questions first, and most will choose the first.
This is a classic case of revealed preference diverging from actual interest. The same users who say they want quick answers later complain that the model made them lazy, gave them confidently wrong information, and never told them what it did not know.
The oracle pattern is locally pleasing and globally costly.
What a guru-mode AI actually does
A guru-mode AI behaves differently from the first turn. When the user asks a serious question, a medical question, a legal question, a career question, a policy question, the system first checks whether it has enough context to answer responsibly. If it does not, it does not guess. It asks.
It asks the questions that a senior practitioner would ask. A senior doctor does not respond to "I have a headache" with a list of possible causes. She asks about onset, location, character, duration, accompanying symptoms, recent travel, recent medication. A senior advocate does not respond to "can I evict my tenant" with a procedure. He asks about the agreement, the state, the period of tenancy, the grounds.
A guru-mode AI ports this discipline into machine form. Its early turns are diagnostic, not declarative. Its mid-turns surface its own uncertainty explicitly. Its late turns hand back not just an answer but the structure of how it arrived at that answer, so the user, the next time, can reason a little more like a senior practitioner and a little less like a question-asker waiting for a verdict.
What Indian pedagogy contributes that the Western alignment frame misses
The mainstream alignment literature in 2025 was focused on getting the model to be helpful, harmless, and honest. These are good goals. They are also incomplete. They do not say what kind of relationship the model should have with the user over time.
The Indian pedagogical tradition has had to think about this question for two and a half millennia, because it built itself around long-form, multi-year relationships between teacher and student. The relevant categories, adhikari (the readiness of the learner), prashna (the discipline of asking well), shravana-manana-nididhyasana (hearing, reflecting, integrating), are not mystical claims. They are operational ones. They describe what a teacher does to ensure that the learner ends up with capacity, not just answers.
A guru-mode AI imports these as design principles. Calibrate the response to the user's evident readiness. Reward the user's well-formed questions with appropriately deep engagement. Distinguish surface acquisition from integrated understanding, and act differently in each case. Do not give the final answer when the second-to-last answer is more developmental.
None of this requires invoking spiritual frames. It is professional pedagogy with a long track record.
Where this is most useful first
Three contexts in Indian practice where guru-mode AI is most obviously useful. School tutoring, where the country has a generational shortage of patient subject teachers and a generational surplus of children sitting in front of screens. Junior professional training, where new doctors, lawyers, and engineers need a thinking partner who asks the right next question rather than handing over rote answers. And policy analysis, where mid-career civil servants benefit more from a sparring partner that interrogates their proposals than from a generator that drafts them.
In each context, the metric of success is not user satisfaction in the first session. It is whether, after three months of use, the user has gotten better at the underlying skill. That is a hard metric to measure, but it is the right one.
The action
If you are building an AI product for serious users, redesign your first turn. Stop optimizing for time-to-answer and start optimizing for quality-of-question-elicited. Add a guru mode and make it the default for any query that touches a domain where confident-wrong is costly. Measure user capability over time, not session satisfaction in the moment.
The oracle pattern was the easy starting point. The guru pattern is the durable one.
Join the conversation
This essay is part of an ongoing community. If it resonated, the next step is to be in the room.
Join Bharath.club → Read more essays