What is sourced from primary texts, what is translated by humans, what is drafted with a language model, and what checks run before anything is published.
This page exists because the question is legitimate. When a site publishes commentary on Sanskrit scripture, readers have every right to know who wrote it, what they read, and where a machine was in the loop.
The short answer: the translations on this site are the Wisdom edition, edited by Ankur Shukla, drawn from public-domain English sources. Commentary is drafted with a language model and reviewed by a human before publishing. Every verse citation is checked. Every named commentator's position is verified against a sourced edition before it goes up.
The longer answer is the rest of this page.
Every verse explanation on this site passes through the same sequence. Nothing is published that hasn't been through all five steps.
Sanskrit editions in the public domain. The Bhagavad Gita text used here follows established critical editions. For the Devanagari and transliteration, we work from standard published sources, not from model output.
The English translations are the Wisdom edition, drawn from public-domain English sources. Where a specific source is identifiable, it is named. New translations going forward carry full provenance. The AI does not translate Sanskrit. It works from the human translation as its source.
Which verses are published, which are held back, and how they are organized is Ankur's editorial call. The model does not select content.
A language model drafts the commentary (simpleMeaning, detailedMeaning, modernRelevance) from the human translation and editorial context. The model is given the translation, not the Sanskrit. It is not asked to translate; it is asked to explain. The prompt it works from is constrained to reduce the specific failure modes listed below.
Each piece is read against the source translation, checked for accuracy, and edited before it goes up. The review date on each page reflects the last time this check ran.
These are hard rules. If any of them are violated, the content is not published.
Language models can fabricate confidently. They can invent verses, misattribute commentaries, and smooth over real disagreements between traditions in a way that sounds authoritative. This is the failure mode we work hardest against.
Every verse citation, every named commentator's view, every cross-reference is checked by a human against a sourced edition before it ships. The model is not trusted to get these right on its own. The review step exists precisely because it doesn't.
The practical implication: if you find something on this site that misattributes a position, invents a verse, or flattens a real sectarian disagreement, that is a failure of the review step and we want to know about it. Write to work.ankurshukla@gmail.com. Errors that are substantive get corrected with a visible note on the page.
The disclosure is on every page of this site, in three places.
A single line in the footer of every page: Wisdom editorial translations. Commentary AI-drafted, human-reviewed. A per-post byline on each blog and verse page naming the editor, the AI-drafted commentary, and the last review date. And this page, reachable from every byline.
The decision to disclose is not complicated. When AI touches scripture, the risk is not only stylistic. A model can produce something that sounds like authentic commentary while quietly inverting the philosophical claim, or attributing a verse to the wrong source. Readers are right to want to know what was sourced, what was translated by a human, and what was drafted with a model in the loop. Hiding that pipeline is what causes problems. Showing it is what fixes them.
The framing the disclosure leads with is what is human, not what is machine. This is not spin. The five steps above are genuinely what happens. The AI drafts commentary from a human translation the editor chose, and the editor reviews what the model produces before it goes up. That is the accurate description of the workflow.
Every page on this site carries a last reviewed date. When a substantive error is found and corrected, the date is updated and a brief note is added at the bottom of the page describing what changed.
This is standard practice in scholarly digital editions. A visible errata record is more trustworthy than content that appears to have never been wrong.
If you find an error, please write to work.ankurshukla@gmail.com with the verse reference and the specific claim you think is wrong. Include a source if you have one. We will check it.
Wisdom is a solo project. The app, the website, the data pipeline, and the editorial work are all Ankur Shukla. The site is not a publishing house with a fact-checking department. It is one person trying to do this carefully.
The disclosure on this page is part of that attempt. The goal is not to appear trustworthy. The goal is to be trustworthy, which means being honest about what the process actually is, including the parts that involve a language model.
Report an error →