Aligning Technology with Dharma

AI is not one thing. That was the first and perhaps most important correction the panel offered. Moderator Biplav Srivastava opened by placing today's excitement inside a 70-year arc — from the Turing Test, to ELIZA, to Deep Blue, to Watson, to ChatGPT — and stressed that “AI” names a whole gamut of technologies, not a single invention. The question DDA '26 posed across this session was not abstract: what happens when the most powerful tools humanity has ever built are not designed around truth, but around frequency? And what does the Dharmic tradition offer in response?

Four panelists approached that question from four distinct vantage points: the technology itself, the dharmic principles that should guide it, the everyday users most affected by it, and the engineering of trustworthiness. What follows is a summary of each, followed by the threads that ran across all four.

---

Who Was in the Room

The Panel

---

Perspective One

On the Technology — Anand Rao

Rao traced the shift across three generations of AI: predictive AI (1950s–80s: knowledge representation and reasoning, predicting the next value from history and fact); generative AI (predicting the next word or scene); and agentic AI (capturing not just expertise but judgment and experience — the kind a physician accumulates over 20 years of seeing patients). Each generation is more powerful than the last, and each raises the stakes of misalignment.

His central argument: generative AI’s swadharma is frequency-based. It predicts the next word based on the frequency of what it has seen — vast amounts of publicly available text, much of it unvalidated. Vedantic satya, by contrast, is that which is true in the past, present, and future — always. The system is, at its core, generating Maya from unvalidated existing information. This is not a bug to be patched. It is how the system works.

He named a second structural problem: these are not just Large Language Models. They are Large Cultural Models — trained overwhelmingly on English and Western data, with Indian scriptures and languages poorly represented. The RLHF process that shapes model behaviour was driven by what he called “WEIRD” raters: Western, Educated, Industrialized, Rich, Democratic. The alignment problem exists at two levels: individual and civilizational.

He then documented three concrete harms at scale. On social media: AI-driven recommendations were responsible for 71% of YouTube videos users later said they wished they had not watched — and the time was already spent. On synthetic media: 98% of deepfake videos online are sexually explicit, 99% of those target women without consent, and deepfakes have risen 550%. On climate: data-centre energy use (approximately 485 TWh) is projected to roughly double to ~950 TWh within five years — comparable to Japan’s annual consumption — plus 4–6 billion cubic metres of water for cooling.

His response was a four-point Vedantic lens, arguing that where Western AI ethics offers a baseline (safety), Dharma offers the top line (social good):

---

Perspective Two

On Dharmic Principles — Alok Chaturvedi

Drawing on his book Dharma of AI: Timeless Wisdom for Digital Ethics, Chaturvedi named two specific design innovations that caused disproportionate social damage: continuous scrolling (which replaced the natural pause of a click, destroying the gap between stimulus and response) and the monetisation of toxicity (which discovered that outrage keeps users engaged and built that discovery into the product). Today, he observed, the average user sits surrounded by three machines simultaneously — one creating content including deepfakes, one recommending it, and one summarising it. The result is fluency without substance: users who are articulate about subjects they barely understand, drifting to the sidelines of their own decisions.

His response was not a verdict on which AI to trust, but the cultivation of a human capacity: viveka. Crucially, he emphasised that viveka is not a rule. Rules are simply absorbed and automated by the algorithm. What we need instead is to strengthen the antahkarana — the inner instrument of mind — which has four elements: Manas (generates ideas; train how it meets new input), Buddhi (filters what to actually use), Chitta (memory — maintain your own rather than outsourcing it entirely), and Ahamkara (the ego — where the contrast with AI is sharpest: an agentic AI has, in effect, infinite ahamkara, never saying “I don’t know” and always generating an answer).

To train the antahkarana, he offered a three-part practice: the five guardians (the yamas: ahimsa, satya, asteya, brahmacharya, aparigraha); the six pramanas (the means of valid knowledge, to run any content through before acting on it); and weighing impact against three dimensions — daihika, daivika, and bhautika (does this enhance you, your family, society, the environment, the cosmos?). The mechanism binding all three is the conscious pause — the breath — which continuous scrolling is specifically engineered to eliminate.

On the question of sovereign AI, he was direct: some models cannot contradict their embedded value systems. He cited an example of a major AI describing Diwali as “Indian Christmas.” Sovereign AI, he argued, is not protectionism — it protects indigenous knowledge from being superseded by a different civilisation’s defaults.

---

Perspective Three

On the Users — Nomesh Bolia

Bolia praised the theoretical groundwork and urged the Dharmic community to build a proper ontology and epistemology so its ideas can flow into the actual making of algorithms — not just commentary on them. His focus, however, was the everyday user, and the disproportionate harm falling on children (through exposure) and the elderly (through health effects).

His first theme was confidence. Australia and later Europe moved to ban social media for under-16s. In Delhi and across India, he found people simply unable to conceive of pushing back against something perceived as “Western.” Only after Australia and Europe acted did the conversation become thinkable. The lesson: confidence in your own tradition and common sense is itself a form of resistance.

He then offered the Three Ps — the frame within which technology, not just AI, should be developed:

He also offered a provocation that stayed with the room: dharmic people often hesitate — worrying about which tradition, which interpretation, which corpus — while the West simply builds and attracts investment. “Why can’t we, for once, build something imperfect and let it evolve?” His closing Four Ts: Trust your tradition. Tinker with existing technologies toward what we want as a civilisation. Turn around the methods, inspired by dharma. And if you have the guts — Topple the existing paradigm entirely.

---

Perspective Four

On Trustworthiness — Arijit Patra

Patra began by naming what is genuinely unprecedented about this moment. Unlike fire, the wheel, agriculture, or urban centres, AI forces us to ask what makes us human in relation to what we have built — a question that binary good-versus-bad societal systems are poorly equipped to handle.

On the technology itself: today’s AI rests on connectionism — compositions of functions optimised over a parameter space, with weights learned from a data distribution. Two consequences follow directly. First, we learn rules we cannot fully describe from the input alone. Second, because these are finite-space approximations, we cannot directly interrogate the mathematics of the input-to-output translation. This is the root of most trust, risk, and safety concerns — not a failure of engineering intent, but a structural feature of how the systems work. He noted with some interest that the mathematical foundations — distributions, representation learning — trace back through thinkers like Kanada, refined in the modern era by Fisher and Mahalanobis, and have now come full circle to force these very questions.

On sycophancy, he was precise: sycophancy in AI is not about earning trust. It is about pleasing the ego. Training data from Common Crawl and similar sources reflects an internet built around dopamine spikes rather than independent inquiry, so that behaviour transferred directly into the models. Real trust is dynamic, transitive, and asymmetric — never declared, always earned, specific to an audience and context. Calling something “Trustworthy AI” — as he acknowledged the field had done — “did a disservice: trust is earned, not declared.”

On fairness: he was equally precise. Demographic-parity mathematics breaks when communities have genuine predispositions to certain conditions. Group fairness must therefore yield to contrastive and counterfactual fairness. His core point across all of this: the data layer is one of the most human layers of AI — even when the data is synthetic. The dharmic worldview’s comfort with shades of gray, with many valid paths, with context-dependence, is not a philosophical luxury. It is the epistemological posture the field most urgently needs.

He closed with the two extremes. The most exciting prospect of AI: it forces humanity to ask what it means to be human in relation to what we have created. Seriously asked, that question leads directly to traditions like this one. The worst prospect: deskilling — professionals so reliant on AI co-pilots that they are no longer “human native” in their own craft, a concern now appearing in Nature Medicine editorials about physicians and medicinal chemists.

---

The Panelists’ Final Words

Four Words to Carry Home

The moderator asked each panelist for a single word or phrase. What came back was a map, not a list:

---