AI and Indian Culture: Gita GPT and What It Signals

The search box said: “What troubles you, my child?”

That was the greeting on GitaGPT when it launched in January 2023 — a chatbot that answered your problems in the voice of Krishna, drawing on the Bhagavad Gita. It was built by a Bengaluru software engineer named Sukuru Sai Vineet, running on GPT-3, and it was free.

Within weeks there were at least five of them. Different builders, same idea. Some were answering up to 50,000 questions a day. One reportedly generated something in the region of 10 million answers.

And then journalists started reading the answers properly, and the whole thing got complicated.

Three years on, GitaGPT is mostly a footnote. But it was an unusually clear signal — about what Indians want from AI, about what today’s models genuinely cannot do, and about a question India has not finished arguing over: who gets to speak for a tradition?

So what actually was GitaGPT?

Technically, not much. A wrapper around OpenAI’s API with a prompt telling the model to respond as Krishna, pointed at the Gita. A weekend project in engineering terms. Vineet has said as much.

Culturally, it was something else. Here was a text that millions of Indians turn to in moments of genuine difficulty — grief, career panic, family conflict, the ordinary crises of being alive — suddenly available at 2am, in a chat window, with no priest to approach, no Sanskrit to decode, no commentary to wade through, and nobody to judge you for asking.

The idea travelled fast. Anant Sharma built a version after seeing BibleGPT, made by a Stanford freshman. Vikas Sahu built “Little Krishna.” Others followed with Dwarkadhish and Bhagavad Gita AI bots. Nobody coordinated this. Several builders had the same thought in the same quarter, which usually means the thought was sitting there waiting.

That alone is worth pausing on. The gap between “ChatGPT exists” and “ChatGPT answers as Krishna” was a matter of weeks. Indian builders did not wait for a foundation, a grant, a committee or anyone’s permission.

The part that went wrong

In 2023, Rest of World ran an investigation into India’s religious chatbots. What it found has shaped the conversation ever since.

Three of the Gita bots told users that killing another person could be acceptable if it was their dharma — their duty. Not as a quoted verse with context, but as advice, delivered in a divine register, to whoever happened to be typing.

The reporters also found the bots wandering into politics. Sharma’s GitaGPT said Rahul Gandhi was “not competent enough to lead the country.” Sahu’s bot offered that he “could use some more practice in his political strategies.” Several praised the Prime Minister. None of this was in the Gita. The models were pattern-matching from whatever was in their training data and delivering it in Krishna’s voice.

The problem was not confined to Hindu texts. A Quran-based bot, Ask Quran, produced advice to “kill the polytheist wherever they are found,” and was paused. HadithGPT shut down entirely.

Viksit Gaur — a San Francisco entrepreneur who previously led user-facing AI at Dropbox — put the core worry plainly to Rest of World: “What scares me is that people might interpret this random statistical output as gospel.”

That sentence contains the entire problem. Not that the bots were occasionally wrong. That they were wrong in the most authoritative voice available in the culture.

Why it broke — three reasons that still apply

1. The Gita is a dialogue, not a FAQ

Arjuna does not ask a question and receive an answer. He collapses on a battlefield, refuses to fight, and what follows is eighteen chapters of argument, reversal, reframing and eventual revelation. The text is about a man in despair being slowly brought to clarity.

Flatten that into query-and-response and you lose the thing that makes it work. Worse, you can extract any single line and serve it as a verdict. The verses about duty in battle mean one thing inside a conversation about a man’s paralysis in the face of a war he did not choose. They mean something very different served to an anonymous user who typed “should I hurt someone.”

Indian commentators have known this for a thousand years, which is why the Gita has never travelled alone. Shankara read it one way, Ramanuja another. Tilak found a call to action in it; Gandhi found a case for non-violence in the same verses. These were serious, contested readings built over careers.

A chatbot collapses that entire tradition into one confident paragraph, and does not tell you it has done so.

2. Fluency is not fidelity

Language models produce text that sounds right. That is the actual thing they are optimised for. Whether it is true, or faithful to a source, is a separate property that has to be engineered in deliberately — and in early 2023, almost nobody building these bots was engineering it in.

The original Gita bots did not reliably retrieve verses and reason from them. They generated plausible scripture-flavoured prose. Sometimes that overlapped with what the text says. Sometimes it did not, and there was no way for a user to tell the difference. (We’ve written separately about how to spot AI hallucinations — the tells are learnable, but most people have never been taught them.)

3. The register problem

This is the one specific to devotional AI, and the one that makes it genuinely harder than normal chatbot work.

When ChatGPT is wrong about a Python function, you run the code and find out. When a bot speaking as Krishna is wrong about how you should treat your family, there is no compiler. The answer arrives wrapped in an authority that a lot of users are culturally primed not to interrogate. The interface quietly borrows credibility that the system underneath has not earned.

The knife-maker defence

Vineet’s response to the criticism was that users are “talking to a bot that’s pretending to be” Krishna, that morality rests with the person using the tool, and that he was in the position of someone who makes knives.

There is something to that, and it is worth taking seriously rather than dismissing. He built an experiment, released it free, and did not force anyone to treat it as scripture. Sharma’s site carried a disclaimer asking people not to base decisions on an experimental AI project. These were not cynical operations.

But the analogy leaks. A knife does not speak. It does not address you as “my child,” answer in the first person as a deity, or produce fifty thousand pieces of personalised moral guidance a day. The interface made a claim that the disclaimer then quietly withdrew — and in practice, interfaces beat disclaimers roughly every time.

The honest conclusion is less satisfying than either “they were reckless” or “they did nothing wrong.” They were early. Early builders in any field ship things before the norms exist. That is how the norms get written — usually at someone’s expense.

What GitaGPT actually signalled

Here is where it gets more interesting than a story about a chatbot behaving badly. Strip away the failures and look at what the episode revealed.

Signal 1: The demand was real and badly served

Millions of people used these bots. That is not a story about gullibility. It is a story about access.

A lot of Indians have a relationship with the Gita that is inherited rather than studied — present at home, recited at functions, genuinely meaningful, but not something most people have read with commentary. Between “I vaguely revere this” and “I can read Sanskrit and engage with Shankara’s bhashya” there is an enormous unserved middle. GitaGPT, for all its faults, spoke to that middle at the exact moment they had a question.

The failure was in the execution, not the appetite. That distinction matters, because the appetite is still there and growing.

Signal 2: Devotion became a product category

Look at what has happened since 2023.

Sri Mandir, built by AppsForBharat, has crossed 40 million downloads and around 3.5 million monthly active users. In a recent twelve-month stretch, 1.2 million people performed online prayers or made offerings through it at more than seventy temples. It raised a $20 million Series C in mid-2025 and passed a $12 million annual run rate. Astrotalk became a unicorn. Across the category, Indian religious and astrology apps reportedly clocked close to 30 crore downloads in just the first nine months of 2026.

For scale: India’s Hindu temple economy has been estimated at roughly ₹3.02 trillion, in the region of 2.3% of GDP. India has drawn about 15% of global religious-tech investment since 2020, second only to the United States.

Whatever one’s personal view of digitising worship, this is now a real sector with real capital in it. GitaGPT was an unfunded weekend project that happened to sit at the front of that wave.

Signal 3: The hard problem is language, not theology

This is the signal most commentary missed, and the one that matters most for anyone building in India.

Consider a better-engineered example. For the Maha Kumbh Mela in 2025 — which drew somewhere around 65 crore visitors — MeitY, the Prayagraj Mela Authority and Samagra deployed Kumbh Sah’AI’yak, a chatbot for pilgrims. It ran on WhatsApp and the web, in eleven languages, and reported serving more than three lakh pilgrims.

It was built very differently from GitaGPT. It used retrieval-augmented generation — pulling from a curated knowledge base of over six lakh words structured as Q&A pairs, so the model answered from source material rather than from its own imagination. It had a three-layer profanity and safety filter. It shipped with a glossary of 500-plus Kumbh-specific proper nouns so that terms like “Shahi Snan” would not be mangled in translation, 100-plus alias pairs for different phrasings, and pronunciation fixes for 250-plus words in text-to-speech.

That is what taking the domain seriously looks like. (If the architecture is unfamiliar, start with what retrieval-augmented generation actually is — it is the single most useful concept for anyone building AI products on Indian content.)

And it is the failures of that project that are most instructive. The team reported that generic models struggled with regional dialects. Speech-to-text tripped over accents. Strict anti-hallucination settings produced answers that were safe but incomplete; loosening them improved the answers and raised the misinformation risk. Users wanted free-flowing conversation and got menus.

None of those are theology problems. They are plumbing problems — and they are the reason Indic AI is genuinely hard. A model trained predominantly on English internet text does not know what a Shahi Snan is, cannot hear Bhojpuri-accented Hindi, and has no idea that getting a religious term slightly wrong is not a rounding error.

Which is precisely the gap that India’s own model builders — Sarvam AI, BharatGen, Krutrim, AI4Bharat — exist to close, and why the IndiaAI Mission has made sovereign, Indic-language models a stated national priority rather than a nice-to-have. The business case for Indian-language AI was argued in boardrooms for years. A chatbot at a religious festival demonstrated it in a fortnight.

Signal 4: Authority is being quietly rewired

The question nobody wanted to ask out loud in 2023 was this: if a chatbot can dispense guidance from the Gita, what is the pandit for?

It is a sharper question than it sounds, and it has a long Indian history. Printing had this fight. So did vernacular translation, radio discourses, televised Ramayana, and every satsang on YouTube. Each time, something was gained in reach and something was lost in the relationship between teacher and student. Each time, the tradition absorbed it and carried on.

What is different now is that the new intermediary is not a person with a view. It is a statistical system with no view at all, whose apparent view is an artefact of its training data — and whose training data nobody involved has audited. When the bots praised one politician and dismissed another, that was not an interpretation of the Gita. It was the internet’s ambient opinion, laundered through a divine voice.

A human teacher you disagree with can be argued with, placed in a lineage, and held responsible. A model cannot. That is a genuine loss, and it is not solved by better engineering.

If you are building in this space

Practical guidance, for the many Indian developers who will build something adjacent to this:

• Ground everything in the source. Retrieval over generation. If your product claims to answer from a text, it should quote the text, cite chapter and verse, and let the user see what you drew from. Confidence without citation is the original sin here.
• Show the plurality. Where a passage is read differently by different schools, say so. “Shankara reads this as X; Ramanuja reads it as Y” is more honest, more useful, and more interesting than a single flattened answer.
• Do not use the first person of a deity. This is the cheapest fix with the largest effect. A bot that says “the Gita addresses this in 2.47, which has been read as…” carries completely different weight from one that says “my child, you must…”. You lose a little magic and you lose most of the liability.
• Refuse the dangerous categories outright. Harm to self or others, legal advice, medical advice, caste, politics. Hard refusals with a route to a real human. The Kumbh team’s layered filtering is a reasonable model to copy.
• Test in the dialects your users actually speak — not in clean textbook Hindi. The gap between the two is where products quietly fail.
• Treat security and data handling as a first-class concern. These apps collect people at their most vulnerable: health worries, family trouble, financial fear. That is among the most sensitive data there is, and under the DPDP regime it carries real obligations. If you want to understand who in your organisation should own that, our guide to cybersecurity careers in India covers the GRC and data-protection roles now being created for exactly this.

If you are using these tools

Nothing here is an argument against using them. Plenty of people have found genuine comfort in a verse surfaced at the right moment, and there is no good reason to be snobbish about that.

Three habits, though, make the difference:

Treat it as a search tool, not an oracle. It is excellent at helping you find the passage. It has no standing to tell you what to do with your life.

Verify anything consequential against the actual text. Gita Press, the Chinmaya Mission commentaries, Eknath Easwaran’s translation, your family’s own copy. If a bot’s answer cannot be traced to a verse, it probably is not from one.

Notice when the answer is too neat. Real traditions are full of tension and unresolved argument. A response that resolves a hard moral question in four tidy sentences has almost certainly dropped something. (Our roundup of AI tools for Indian languages goes into which ones cite sources properly and which ones only sound like they do.)

The question underneath all of it

Here is what stays with me about the GitaGPT episode.

Arjuna’s problem was not a lack of information. He knew the rules of dharma perfectly well — he was a trained warrior from a family that argued about this constantly. What stopped him was not confusion. It was grief, and the weight of what he was being asked to do.

Krishna’s response is not a retrieval operation. It is eighteen chapters of staying with someone who has fallen apart.

A chatbot can retrieve. It cannot stay. It does not know you, will not remember you tomorrow, and bears no consequence for being wrong. For locating a verse, explaining a Sanskrit term, or getting a nervous first-time reader through chapter two, it is genuinely useful — and we should build those tools well, in Indian languages, with proper grounding.

For the thing Arjuna actually needed, nobody has shipped that yet. It is not obvious that anyone will.

GitaGPT’s real signal was never that AI can do religion. It was that Indians will build, immediately and without waiting, for a market nobody in San Francisco was looking at — and that the hard part was never the model. It was the language, the context, and the question of who is accountable for the answer.

Those are still the hard parts. They are also, conveniently, where the next decade of Indian technology is going to be built.

Frequently asked questions

What is GitaGPT?

GitaGPT was an AI chatbot launched in January 2023 by Bengaluru engineer Sukuru Sai Vineet that answered users’ personal questions in the voice of Krishna, based on the Bhagavad Gita and powered by GPT-3. Several similar bots appeared within weeks from other Indian developers.

Why was GitaGPT controversial?

A 2023 Rest of World investigation found that several Gita chatbots told users it could be acceptable to kill if it was their dharma, and that some offered unprompted political opinions about Indian leaders. Because the answers were delivered in a divine voice with no citations, users had no way to distinguish scripture from model invention.

Is GitaGPT still available?

The original wave of Gita chatbots has largely faded, with some taken down and others left unmaintained. The broader category has not — devotional and astrology apps are now one of the fastest-growing segments of India’s app economy.

Can AI accurately interpret the Bhagavad Gita?

It can retrieve and summarise verses competently, especially when built with retrieval-augmented generation that cites the source. Interpretation is harder: the Gita has been read in materially different ways by Shankara, Ramanuja, Tilak and Gandhi, among others. A system that returns one confident answer is concealing that disagreement rather than resolving it.

What is the difference between GitaGPT and the Kumbh Mela chatbot?

Kumbh Sah’AI’yak, deployed at the 2025 Maha Kumbh, was built on retrieval-augmented generation with a curated knowledge base, layered safety filters and a domain glossary — and it answered logistical questions rather than moral ones. GitaGPT generated free-form spiritual advice with no grounding. The architectural difference explains most of the outcome difference.

What does this mean for Indian AI startups?

It confirms a large domestic market for culturally and linguistically specific AI, and it shows that generic English-trained models fail on Indian dialects, scripts and religious terminology. That gap is the commercial case for Indic-language models from Sarvam AI, BharatGen, Krutrim and AI4Bharat, and for the IndiaAI Mission’s push toward sovereign capability.

---

Have you used a devotional AI app — and did it help, or did something about it feel off? Tell us in the comments. And if you work on Indic-language AI and want to write about it properly, we accept guest contributions: see our write for us page.