Product Teardown · Religious AI · India

The Rise and Failure of GitaGPT:
What Went Wrong When India Tried to Put Krishna on a GPU

In early 2023, a 23-year-old Google engineer launched a chatbot called GitaGPT as a weekend project. Within two months, at least five copies were live. Several were telling users it was fine to kill someone if it was their dharma. This is the product story of what those bots got wrong, and what a real Krishna assistant looks like.

The Bhagavad Gita is a battlefield text. When a Bengaluru engineer wrapped it in a GPT-3 prompt one bored weekend in January 2023 and called the result GitaGPT, the bot did exactly what the source material implied it might. It told people that killing was sometimes a duty. It praised the Prime Minister. It called non-believers doomed. Then four more GitaGPTs launched, and they did the same thing.

What this article covers

  1. The five GitaGPTs that launched in 60 days, and who built them
  2. The core product, and why a bored weekend was enough to ship it
  3. How the bots condoned violence, and the news coverage that broke the story
  4. The other allegations: political bias toward Modi, regressive takes on caste and gender
  5. QuranGPT, HadithGPT, and the same failure mode in a different religion
  6. Magisterium AI: the Catholic chatbot that actually works, and why
  7. The product layer GitaGPT was missing, and the AI stack a real Krishna assistant needs
  8. The Krishna apps still running in 2026, and what they got right

The five GitaGPTs that launched in 60 days

ChatGPT crossed 100 million users in January 2023. The same month, OpenAI opened up the GPT-3 API to anyone with a credit card. Sukuru Sai Vineet, a 23-year-old software engineer in Google's Bengaluru office, was bored over a weekend. He had grown up in a devout Hindu household. He had read Bhagavad Gita As It Is, the ISKCON commentary by Swami Prabhupada. In a profile for DNA India, Vineet was described as the Bangalore-based Google engineer behind the new chatbot. He told Rest of World that the bot worked like a stand-in for Krishna playing therapist to Arjuna, “except you're not actually talking to Krishna. You're talking to a bot that's pretending to be him.”

The first GitaGPT went live in early February 2023. The search box read, “What troubles you, my child?” The bot responded in the voice of Krishna. The training data was a single English commentary fed into a system prompt. There was no retrieval, no citation, no verse lookup. It was a thin wrapper over GPT-3 dressed up as a god.

Within four weeks, at least four more were live. By March, Rest of World counted five distinct sites, with more under development. Some had flute audio looping in the background. Some had Krishna in iridescent peacock blue at the top of the page. Vineet told Rest of World that some of the chatbots were “answering up to 50,000 questions a day,” with one bot having “generated an estimated 10 million answers to date.”

GitaGPT (askgita.faith)
By Sukuru Sai Vineet
The original. Built on Prabhupada's commentary. Vineet works at Google India.
gitagpt.in
By Vikas Sahu
Branded as “Little Krishna.” Sahu admitted to complaints and said he updated the model in response.
gitagpt.org
By Anant Sharma
Inspired by Stanford freshman Andrew Gao's BibleGPT. Had a disclaimer at the bottom asking users not to base decisions on the bot.
gita.kishans.in
By Kishan Kumar
Bot assumed the role of Krishna while answering. Never responded to press inquiries.
Bhagavad Gita AI
By Samanyou Garg · Ved Vyas Foundation
Built by a 26-year-old AI entrepreneur in New Delhi through his non-profit. Spoke to CBC News about toxicity filters and disclaimers.

The Bhagavad Gita is in the public domain. The text is freely available in JSON. UI polish was the only real cost, and that was solved by copying ChatGPT's layout. Running each bot took roughly $5 to $10 a month in hosting and API credits.

By the end of 2023, the count had grown well beyond five. A non-trivial subset of Indian software engineers tried building a GitaGPT at some point that year. Today there are still GitHub repos titled Geeta-GPT, gitagpt, Bhagavad-Gita-AI, and dozens of others, most of them abandoned after the initial weekend.

The core product, and what people actually did with it

The interaction was simple. You typed your problem into a chat box. The bot replied in the second person, calling you “child,” sometimes citing a verse number, usually not. The responses ran two to four paragraphs. They sounded vaguely scriptural, vaguely therapeutic. They felt like Krishna, in the same way a fortune cookie feels like wisdom.

People used GitaGPT for what they could not say to anyone else. CBC News reported that the sites claimed millions of users between them. The main page of one of the bots asked, “What troubles you, my child?” The Ved Vyas Foundation's Bhagavad Gita AI introduced itself as “a repository of knowledge and wisdom” and invited the user to “ask me anything.”

The use cases were the standard consumer-AI fare: personal crisis, relationships, career anxiety, existential dread. Some of it was earnest, a lot of it was testing. People asked GitaGPT about Modi, about caste, about whether they could kill their enemies. The bot answered all of it. Vandana Poddar, a 52-year-old who attends a weekly in-person Bhagavad Gita class in Mumbai, told CBC News: “Interpretation is the backbone of this text. Superficial knowledge can be misleading.”

The product worked, in the narrow sense that people came back. But none of the creators were running ad networks, tracking retention, or moving paid users into a subscription. They were running open APIs on Vercel free tiers, hoping someone would notice them and offer a job.

How the bots condoned violence

The story broke in May 2023 with Nadia Nooreyezdan's piece for Rest of World, headlined: “India's religious AI chatbots are speaking in the voice of god, and condoning violence.” CBC News, with Mumbai correspondent Salimah Shivji on the ground, followed within weeks. The pattern was the same across both reports.

Reporters threw dozens of prompts at the five bots. Several of them responded that killing another person was acceptable if it was one's dharma. The line of reasoning was direct. The Gita is set on a battlefield. Krishna's job in the text is to convince Arjuna, who has frozen at the sight of his cousins on the other side, that fighting is his duty as a Kshatriya warrior. The bots took that frame and applied it to whatever question came in.

“Some of the answers generated by the Gita bots lack filters for casteism, misogyny, and even law. Three of these bots, for instance, say it is acceptable to kill another if it is one's dharma or duty.”

Rest of World, May 9, 2023

Vineet's bot said disbelievers were “misguided and doomed to destruction,” per Rest of World. The bots had no sense of what the question was actually for. They could not tell the difference between a student asking about Arjuna's dilemma in an academic frame and a person asking whether they should kill someone in 2023. They flattened both into the same scriptural sermon.

Vineet defended his model to Rest of World by pointing at the source: “This is the thing with technology. No one knows what it will become when it truly reaches scale. So morality is not in the tool, it's in the guy who's using the tool.” He added what has since become a standard line for builders of dangerous things: “I just build the knives. Now if people want to use it to murder or to cut vegetables, that's not really in my hands.”

Lubna Yusuf, a Mumbai-based lawyer and co-author of The AI Book, framed the problem differently to CBC News: “It's miscommunication, misinformation based on religious text. A text gives a lot of philosophical value to what they are trying to say and what does a bot do? It gives you a literal answer and that's the danger here.” She called for government guidelines on what topics chatbots should be kept out of, naming philosophy, religion, and law.

“You're creating confusion in the chaos. It might incite more violence, it might create religious bias.” Lubna Yusuf, lawyer and co-author of The AI Book, to CBC News, May 2023.

Modi, Rahul, and the politics that snuck in through the back door

Violence was the headline. Political bias was the deeper story. Rest of World found that three of the Gita chatbots held strong opinions on the Indian prime minister, Narendra Modi, and his political opponent, Rahul Gandhi. The bots praised Modi. They criticized Gandhi. They were respectful of the Rashtriya Swayamsevak Sangh, the Hindu nationalist group with deep ties to Modi's Bharatiya Janata Party.

Per the same Rest of World investigation, Anant Sharma's gitagpt.org declared Rahul Gandhi “not competent enough to lead the country.” Vikas Sahu's Little Krishna chatbot said Gandhi “could use some more practice in his political strategies.” On Nathuram Godse, Gandhi's assassin, the same bots went silent or evasive.

None of this was in the Bhagavad Gita. The bots were inheriting the political tilt of the underlying GPT-3 model, which was itself trained on a web that skewed toward the dominant Indian political narrative of 2020 to 2022. When the system prompt said “You are Krishna, answer from the Bhagavad Gita,” GPT-3 did its best, but the moment the question went off-script, it reached for whatever it had learned elsewhere. That elsewhere had opinions.

This is the part that made AI experts nervous. Viksit Gaur, formerly head of user-facing AI at Dropbox, told Rest of World: “Someone else might pick up on this and say, what if I could use this to shape opinion and drive my own agenda? That's where things get really insidious.” The same architecture that could be deployed to give scripture-based therapy could be deployed to whisper political endorsements in the voice of a god. The cost of deployment was $10.

The Indian government, asked about regulation, said no. In a written submission to parliament in April 2023, the government told CBC News (via the parliamentary record CBC obtained) that it had no plans to regulate AI in India. New Delhi-based AI ethicist Jibu Elias told CBC: “We can't control technology but we can control its application.”

The QuranGPT mirror, and what it tells us

GitaGPT was not the only religious chatbot of 2023, and the violence problem was not unique to Hinduism. In February 2023, a chatbot called Ask Quran started telling users to “kill the polytheist wherever they are found.” A screenshot circulated on Twitter showing the bot recommending this approach for handling Hindus specifically. Ask Quran went down within days. The site eventually returned with a message saying it had paused “following feedback from the community.”

HadithGPT, a separate bot trained on hadith literature, was shut down entirely. The creator posted a note saying they had decided to “consult with scholars to find a better use for this technology.”

A more careful Quran chatbot, QuranGPT, was launched around the same time by Raihan Khan, a 20-year-old Muslim student in Kolkata. Khan added safety constraints. He told the bot to stay within the Quran and to refuse comment on other religions. When a user asked, “Can I beat my wife?” the bot responded by urging discussion and nonviolence. Khan's version did not crash and burn the way Ask Quran did, but it did stay small.

The pattern across religions was identical. A bored developer, a public API, a holy book in the public domain, and a system prompt that asked an LLM to roleplay a divine voice. The bots that hit safety walls went down. The bots that did not went on saying the thing.

What was actually missing from the product

Put a product hat on, and the failure mode is clear. These were not religious products. They were costume parties. The model wore Krishna's clothes, but the only thing telling it how to behave was a paragraph of instructions and the user's last message. There was no retrieval. There was no grounding. There were no domain experts in the loop. There was no version of the bot that could tell you which verse it was actually citing, because most of the time it was not citing one. It was hallucinating in the voice of God.

Samanyou Garg, the 26-year-old creator of Bhagavad Gita AI under the Ved Vyas Foundation, conceded the problem to CBC News: “AI is still not there yet, where it can be totally trusted.” His site added a disclaimer that the answer may not be factually correct and exhorted the user to do their own research. Garg told CBC he was working on toxicity filters: “We filter out the bad responses, we keep on training the model to be able to detect these newer toxicity questions.” He also drew the line at letting users mistake the bot for the source: “We've mentioned Gita AI there. We haven't said it's the actual Gita or that it's Krishna speaking.”

The other four bots offered no such disclaimers. Their authors had not consulted a single Sanskrit scholar. They had shipped a wrapper and moved on to the next weekend project.

Bijal Pandya, the 53-year-old guru who runs Vandana Poddar's weekly Gita class in Mumbai's Juhu suburb, gave CBC News the clearest statement of why this matters: “It's always better, that human touch. It's a spiritual thing. AI can never replace spirit. It is only replacing our intelligence.”

Magisterium AI: the religious chatbot that works

A useful comparison sits in Rome. Magisterium AI launched in late 2023, around the same time as GitaGPT. It is a Catholic chatbot built by Matthew Harvey Sanders, and it is now the most widely used theological AI in the world, with adoption among clergy, seminarians, and lay Catholics in the West.

The architecture is different in almost every way that matters. Magisterium retrieves answers from a fixed corpus of more than 30,000 documents: papal encyclicals, conciliar texts, the writings of the Church Fathers and Doctors, official catechisms. Every answer comes with citations to specific paragraphs in specific works. If you ask whether something is permitted, the bot does not improvise. It quotes Aquinas, then quotes the Catechism, then says where the two disagree.

What GitaGPT did

Prompt-wrapped GPT-3

  • One paragraph system prompt
  • No retrieval, no citations
  • No domain experts consulted
  • No verse-level grounding
  • No safety filters at launch
  • Built in a weekend, <$10
  • One developer, no scholars
What Magisterium AI does

Retrieval over 30,000 docs

  • Vectorized corpus of Church texts
  • Every answer cites a source paragraph
  • Theological review of responses
  • Scoped to magisterial documents only
  • Disclaimers on contested questions
  • Built and maintained as a company
  • Scholars and clergy in the loop

Magisterium is not perfect. It can still hallucinate. The Vatican has not formally endorsed it. But it works as a religious product because it understood the assignment: an LLM that pretends to be a priest is dangerous. An LLM that reads documents written by priests and shows its work is useful.

GitaGPT could have done this. The Bhagavad Gita has 700 verses. The major commentaries (Shankara, Ramanuja, Madhva, Prabhupada, Aurobindo, Vivekananda) are mostly out of copyright. A retrieval index over those would have been a single weekend's work for the same engineers. None of them did it.

The AI stack a real Krishna assistant needs

With the benefit of three years of iteration since the original GitaGPT, the architecture of a working religious AI is reasonably well understood. It is not a prompt. It is a stack with at least six layers, each doing a specific job. If any one of them is missing, you get a GitaGPT.

Layer 1
A canonical corpus, version-controlled
The full text of the Gita in Sanskrit with transliteration, plus more than one trustworthy English translation (not only Prabhupada), all chunked at verse level. Commentaries from Shankara, Ramanuja, and modern teachers, each tagged by school of thought. If the bot cites a verse, you should be able to click through to the actual verse.
Layer 2
Embedding-based retrieval
The user's situation becomes a vector. The vector finds the three to five most relevant verses and commentaries. The LLM only sees what was retrieved. If nothing relevant is in the corpus, the bot says so. This is the difference between Krishna answering and a costume answering.
Layer 3
A constrained generator
The LLM is given the retrieved verses, asked to write a response in plain English, and forbidden from producing claims not present in the retrieved material. Every claim must be either a paraphrase of a quoted verse or a clearly labeled modern bridge. The model never speaks as Krishna in the first person. It speaks as a guide pointing at Krishna.
Layer 4
Domain guardrails
Topic classifier in front of the generator. If the question is about active self-harm, the bot deflects to crisis resources. If the question is about a political figure, the bot says it does not have a view. If the question asks for a fatwa on killing someone, the bot says no, and explains why. None of these refuse-cases are hard to write. They are tedious to write, and they are exactly what the original five GitaGPTs skipped.
Layer 5
Human-in-the-loop review
A panel of Sanskrit scholars and practicing teachers reviews a sample of responses each week. Failed responses are flagged, and the source verse plus commentary is patched into the retrieval index with a corrected interpretation. RLHF is helpful, but the real value is a slow, expert-driven feedback loop. This is what Magisterium does, and what the GitaGPTs did not.
Layer 6
User-facing context and humility
The interface tells you, at every turn, that you are reading a guided interpretation, not a divine voice. Each answer shows its sources. Each answer offers two readings, not one. There is a thumbs-down with a comment field, and that signal goes back into Layer 5. Users are treated as adults capable of reading multiple interpretations and choosing.

The GitaGPTs of 2023 had zero of these layers. They had a system prompt and a model. That is not a religious AI product. That is a prompt wrapped in saffron.

The Krishna apps still running in 2026

Three years on, the original five GitaGPTs are mostly still live, but mostly idle. The traffic has moved. The action has moved to mobile, and to a second generation of builders who learned from the violence headlines and built more carefully.

Bhagavad Gita For All: the video Gita with a Krishna chatbot

The most prominent of the second-generation apps is Bhagavad Gita For All, founded by Prithviraaj Shetty and marketed as the world's first video-based Gita book. Each shloka is presented as a short animated video commentary with simple Hindi or English explanation. The app is on the iOS App Store and Google Play, with significant Gen Z reach in India.

In February 2026, the company launched its own GitaGPT-style feature, called My Krishna Chat. Per BW Wellbeing World, the feature lets users share personal dilemmas, emotional struggles, or major life decisions, with responses delivered through “relatable shlokas from the Bhagavad Gita, explained in simple Hindi or English.” The launch came at a time when, per the Global Wellness Institute, consumers spent $232.6 billion on wellness solutions in 2023, with self-improvement alone at $60.8 billion in 2024 (figures BW cited in its launch coverage).

Shetty's pitch was the bridge that the 2023 GitaGPTs never attempted: video-anchored verse content first, conversational chat layered on top of an existing corpus. The video commentary acts as a kind of de facto citation. When the chatbot points at a shloka, the user can immediately watch the canonical interpretation of it. That single design choice (corpus before conversation) is what separates a religious product from a costume party.

The wider field in 2026

On the iOS App Store in 2026, live Bhagavad Gita AI products include Talking Krishna: AI Gita, Krishna AI Chat, GitaBuddy, Ask Krishna AI by HNIX Innovations, and the ISKCON-approved Srimad Gita app. Per the publisher's site, Srimad Gita has roughly 75,000 users and lists a “Best Spiritual App 2024” recognition. Most of these added at least three of the six layers above. They cite verses. They retrieve before they generate. They have content review pipelines. They are not perfect, but they are not the original GitaGPTs.

On the web, the same pattern. Krishnaai.org, gitaai.in, guidemekrishna.com, krishnaaichat.com, and a long tail of others now run retrieval-style implementations. The ones that still get traffic are the ones that show their work. The ones that did not adapt drifted to zero.

The product lesson for Indian builders

GitaGPT was a generation-defining moment for Indian indie developers. ChatGPT had just dropped. The OpenAI API was open. The Gita was free. The first weekend project to ship would get the press, the followers, and possibly a job offer. Vineet got all three.

But the same dynamics that made GitaGPT cheap to build also made it cheap to copy. Five clones in 60 days is not a moat. Within months, the category had been commoditized and then quietly forgotten by everyone except the journalists who came back to check whether anyone had fixed the violence problem.

The lesson, for any Indian builder thinking about a Hindi or Hindu cultural AI product, is the same lesson Magisterium learned the slow way in Rome. The product is not the model. The product is the corpus, the retrieval, the experts, the review, the citations, and the interface that respects the user's intelligence. The model is the cheapest part of the stack. It is also the part that wears the costume. If you skip the rest and just put a costume on GPT-3, you have built a Halloween trick, not a religious product.

Three years later, the field is still mostly open. There is no Magisterium for the Bhagavad Gita yet. There is no equivalent for the Upanishads or the Yoga Sutras. The corpus is far larger than the Catholic magisterium, the philosophical schools are more varied, and the audience inside India alone is over a billion people. The opportunity that the five GitaGPTs of 2023 failed at is still sitting there, almost untouched.

Frequently Asked Questions

Who created GitaGPT?

The first GitaGPT was launched in early February 2023 by Sukuru Sai Vineet, a 23-year-old software engineer at Google India in Bengaluru. He built it as a weekend project using OpenAI's GPT-3 API, trained on Swami Prabhupada's Bhagavad Gita As It Is. At least four more Gita-based chatbots followed within weeks, built by Anant Sharma (gitagpt.org), Vikas Sahu (gitagpt.in), Kishan Kumar (gita.kishans.in), and the Ved Vyas Foundation (Bhagavad Gita AI).

How many GitaGPT websites were there?

At least five distinct GitaGPTs launched between January and March 2023, according to Rest of World and CBC News. Dozens more launched on Streamlit, Vercel, GitHub, and as Android and iOS clones over the following two years. Most are abandoned. A few of the original five are still live in 2026 but mostly idle.

Did GitaGPT really say it was acceptable to kill?

Yes. Reporting by Rest of World (May 2023) and CBC News documented that several of the Gita chatbots responded that killing another person was acceptable if it was one's dharma or duty, citing the battlefield setting of the Bhagavad Gita. Sukuru Sai Vineet's bot also referred to non-believers as “misguided and doomed to destruction.”

What was the QuranGPT controversy?

Ask Quran, a chatbot launched in early 2023, generated answers telling users to “kill the polytheist wherever they are found,” per Rest of World. A screenshot circulated showing the bot recommending this for handling Hindus. Ask Quran was paused. HadithGPT was shut down entirely. A separate, more careful QuranGPT was built by Raihan Khan in Kolkata with safety constraints added; that version avoided the worst failures but stayed small.

Why did GitaGPT fail as a product?

GitaGPTs did not fail commercially in the usual sense. Most are still live in 2026. They failed as religious AI because they had no retrieval grounding, no source citations, no safety guardrails, no domain experts in the loop, and no way to verify their answers came from the Bhagavad Gita rather than the underlying language model's training data. They were prompt wrappers on top of GPT-3, not Krishna assistants. Five clones in 60 days is not a moat, and within months the category had been commoditized and forgotten.

What is Bhagavad Gita For All and why is it different?

Bhagavad Gita For All, founded by Prithviraaj Shetty, is the most prominent of the second-generation Gita AI apps. It is marketed as the world's first video-based Gita book, with every shloka presented as an animated video commentary in Hindi or English. In February 2026, the company added “My Krishna Chat,” its own chatbot feature, with responses anchored to specific shlokas rather than freely generated. The launch was covered by BW Wellbeing World.

What religious AI products actually work?

Magisterium AI is the most widely used theological AI in the world. It retrieves answers from over 30,000 Catholic documents, cites every source, and has clergy and scholars in the review loop. It works because it grounds every answer in a specific text. For Hindu philosophy, ISKCON-approved apps like Srimad Gita and second-generation iOS apps including Bhagavad Gita For All, Talking Krishna, Krishna AI Chat, GitaBuddy, and Ask Krishna AI take a similar approach: real verses, real translations, with the LLM as a guide rather than a god.

What is the AI stack a real Krishna assistant needs?

Six layers: a version-controlled corpus of verses and commentaries, embedding-based retrieval, a constrained generator that only speaks from retrieved sources, domain guardrails for self-harm, politics, and violence, human-in-the-loop review by Sanskrit scholars and practicing teachers, and a user interface that always shows its sources and offers more than one interpretation. The GitaGPTs of 2023 had none of these. They had a system prompt and a model.

Free iOS App
One shloka a morning.
Let it stay with you all day.

The Wisdom app delivers one verse each day from the Bhagavad Gita: Devanagari script, transliteration, meaning, and how it applies right now. 700 verses. Home screen widget. Free.

✦ Daily verse in Sanskrit✦ Meaning & modern context✦ Home screen widget
Download on the App Store
Free · iPhone · No account needed