In 1985, a researcher working out of NASA’s AS research center published a short paper in the journal AI magazine, and that single paper has been misqued, exaggerated, and passed around the internet for 40 years straight.
The claim that spread was simple and dramatic.
NASA had supposedly declared Sanskrit the only unambiguous language on Earth, perfectly suited for programming computers.
It sounds too clean to be true, and in the exact form it usually gets repeated, it is.
But strip away the exaggeration, and something much more interesting is hiding underneath it.
Something that’s finally being tested for real decades later using the exact kind of artificial intelligence that paper could only imagine.
The researcher’s name was Rick Briggs, and his actual paper never said Sanskrit should replace programming languages.
What he argued was narrower and honestly more fascinating.
Briggs pointed out that ancient Sanskrit grimarans working over 2,000 years before the first computer existed had built a system of representing meaning that closely resembled what AI researchers in the 1980s were calling semantic networks, structured ways of mapping how words and ideas relate to each other so a machine could process them logically.
He wasn’t saying Sanskrit was a secret code for computers.
He was saying the grammatical precision baked into the language by ancient scholars looked strikingly similar to something modern computer scientists were trying to invent from scratch.
That’s a real narrow defensible academic point.
It just doesn’t travel well on the internet where nuance gets deleted and headlines get louder.
To understand why Briggs even noticed this in the first place, you have to go back roughly 2500 years to a scholar named Panini who wrote a grammar of Sanskrit called the Ashtadayi.
It’s not an exaggeration to call this one of the most sophisticated pieces of linguistic engineering ever produced by a single human mind.
Pony Nye reduced the entire grammar of a language.
Every verb form, every noun ending, every rule for how sounds change when words combine into roughly 4,000 tightly structured rules written in a kind of shortorthhand so dense that modern computer scientists have compared it directly to formal programming language specifications.
In 1967, a researcher named Peter Ingramman published a paper arguing that panini’s rule system functioned almost identically to something called bakis noir form, a notation still used today to define the syntax of programming languages.
That comparison isn’t folklore.
It’s published peer-reviewed computer science history and it means the core idea connecting Sanskrit to computing didn’t start with a NASA memo.
It started centuries earlier, buried in a grammar book and it’s real.
Ingramman wasn’t alone in noticing this either.
A Dutch scholar named Fritz Stall who eventually founded the philosophy and South Asian studies department at Berkeley spent decades studying both panini’s grammar and the even older tradition of vadic ritual chanting and in 1965 he published a paper flatly titled Uklid and panini arguing that panini did for the Sanskrit language what Uklid did for geometry building a small tightly defined set of foundational rules capable of generating an entire complex system through consistent repeatable application.
Stall later pushed the argument even further proposing that the recursive self-mbed structure found in vadic ritual sequences chants and actions nested inside other chants and actions.
According to strick t formal rules might actually be a more direct ancestor of complex human syntax than spoken language itself.
He drew an explicit comparison to the same kind of contextsensitive formal grammar that linguist Nam Chosky had used to classify the structure of modern computer languages.
Whether or not you accept Stall’s more sweeping claim about the origin of language, the underlying observation holds up under scrutiny.
This tradition wasn’t casually poetic.
It was engineered rule by rule with a level of formal rigor that reads more like source code than scripture, which is precisely why computer scientists kept independently noticing the resemblance decades before anyone had built a neural network capable of actually processing the language at scale.
Here’s the problem, though, and it’s the exact problem modern AI is now being built to solve.
Panini’s system might be logically elegant, but Sanskrit, as it’s actually written on the page, is a nightmare for any machine trying to read it.
The reason comes down to something called Sandy, a set of phonetic rules where the boundaries between words physically dissolve into each other during writing and speech.
Two words that are grammatically separate get merged into a single unbroken string of letters with sounds shifting at the seam to make the combination flow.
To a human expert trained for years, this is second nature.
To a computer trying to process raw Sanskrit text, a single sentence can look like one impossibly long word with no visible spaces, no punctuation cues, and multiple technically valid ways to split it back apart, each producing a completely different meaning.
This isn’t a minor technical historically failed on classical Sanskrit texts almost immediately.
Take a simple example of the kind linguists cite constantly.
The words vowel sounds at the junction blending together according to strict phonetic rules.
A human reader trained in those rules sees two words instantly.
A machine with no training sees one unfamiliar string and has to guess.
Multiply that single junction by every word boundary in every sentence of an ancient text running to hundreds of pages and you start to understand why this problem sat mostly unsolved for so long and why it became the specific target so much of this AI research has been built around.
It’s also worth understanding what’s actually sitting inside this body of text because it isn’t just religious scripture even though that’s the association most people default to.
Classical Sanskrit literature includes epic narrative poetry like the Mahabharata and the Ramayana.
Philosophical treatises debating the nature of consciousness and reality, dramatic plays and love poetry from writers like Kaladasa and critically for the more applied AI projects underway right now.
Entire genres of technical writing.
Mathematicians and astronomers like Aryabata wrote foundational texts on trigonometry, place, value, number systems and planetary motion in Sanskrit verse.
Physicians compiled enormous medical compendiums.
The Karaca Samita and the Sushuta Samita, among the most famous, describing diagnostic methods, surgical procedures, including early forms of plastic reconstruction, and thousands of individual herbal and mineral treatments.
None of that technical material reads the same way a hymn or poem does, which means any AI system trained mainly on religious or literary Sanskrit can stumble badly the moment it’s pointed at a medical or mathematical text with its own specialized vocabulary.
That’s part of why a project like Hellwig’s digital corpus deliberately spans multiple genres instead of specializing in just one.
There’s an even older layer sitting underneath all of that written literature.
And it’s arguably the strangest part of this entire story.
The oldest Sanskrit texts, the Vadic hymns, were composed and transmitted for well over a thousand years, almost entirely by memory, long before a standardized written script for recording them was in wide use to guard against even a single syllable drifting or getting corrupted across dozens of generations of oral transmission.
Vadic reciters developed astonishingly rigorous verification techniques.
Methods like in reciting a passage forward, then backward, then in permuted word pair sequences designed specifically to make any accidental alteration immediately obvious to a trained listener.
Modern linguists studying this tradition have compared it to a built-in error correction system, conceptually similar to the checksum techniques modern computing uses to verify a file hasn’t been corrupted during transfer.
It’s a strange kind of poetic symmetry that the same textual tradition modern AI is now racing to make machine readable had already engineered.
Its own version of data integrity checking thousands of years before anyone had a reason to invent the word digital.
For decades, the only real weapon against this problem was the digital corpus of Sanskrit.
A project built and personally maintained since 2010 by German indologists T.
Oliver Hellwig, now holding more than 650,000 individually tagged word forms, each one manually checked for grammar and meaning.
It’s an extraordinary resource and it’s also by its nature painfully slow to grow because Every entry historically required a human expert’s judgment.
Hellwig didn’t stop there.
In 2018, working with fellow researcher Sebastian Nerdick, he published a paper introducing a neural network trained specifically to tackle Sandy splitting automatically, combining recurrent and convolutional network architectures to read raw unsegmented Sanskrit and predict where the hidden word boundaries actually are.
Their system didn’t need a handbuilt dictionary of every possible word form to work.
It learned the patterns directly from examples of already segmented text and it outperformed the older rule-based systems that had dominated the field for years.
The paper published at a major computational linguistics conference marked one of the first times a datadriven AI model meaningfully beat traditional Sanskrit parsing tools built by hand.
Hellwig wasn’t working in a vacuum either.
Years earlier, French computer scientist Gerard Huitt had built something called the Sanskrit heritage engine.
A morphological and phonological analysis system constructed using functional programming techniques and published in detail in the journal of functional l programming back in 2005.
Huitt’s system took a different approach than a modern neural network.
Instead of learning patterns statistically from labeled examples, it encoded panini’s own grammatical rules directly into executable code.
Essentially trying to make a computer follow the same logical rule book a human Sanskrit grimaran would follow.
It’s still maintained and used by scholars today.
And it represents the other major branch of this entire research effort, rule-based systems built on the actual ancient grammar, running in parallel with newer statistical and neural approaches like Hellwigs.
Meanwhile, at the University of Hyderabbad, linguist Ambak Cararnney has spent years building her own suite of Sanskrit computational tools, including a shallow parser and segmentation system, likewise grounded directly in Paninian grammatical theory and contributing to a long-running Sanskrit to Hindi machine translation project called Anusaraka.
Between these efforts, you end up with two philosophically different camps working the same problem.
One trying to teach a machine the ancient rules explicitly.
The other trying to have a machine discover the patterns on its own from data.
And increasingly projects like BT5 Sanskrit are finding ways to blend both approaches together.
The name itself is a clue to how it works.
It’s built on a transformer architecture, the same underlying design family behind most modern large language models, but adapted to read raw byes of text character by character instead of relying on pre-built vocabulary lists the way many earlier language models did.
That distinction matters enormously for a language like Sanskrit where the sheer number of possible word forms generated by its grammar is so large that building a complete fixed vocabulary list in advance simply isn’t practical.
By operating at the bite level instead, the model can encounter a word form it has technically never seen before during training and still make a reasonable attempt at processing it correctly, which is exactly the kind of flexibility a heavily inflected compound forming language demands.
The pace has only picked up since.
In 2024, Hellwig and his collaborators released something called by T5 Sanskrit, a single unified language model capable of handling word segmentation, grammatical tagging, and even Vadic era sentence structure analysis, all within one system instead of needing separate specialized tools stitched together for each task.
According to the researchers who built and tested it, it didn’t just match the performance of the best existing lexicon-based approaches.
In several tasks it beat them outright while also proving noticeably more robust when it encountered word forms.
The older, more rigid systems had simply never been designed to handle before.
It also picked up a genuinely practical side benefit, cleaning up errors left behind by optical character recognition scans of old manuscripts, correcting garbled text automatically instead of requiring a human proof reader to catch every mistake by hand.
That OCR piece of the puzzle matters more than it might sound because before any AI model can analyze a Sanskrit text.
Something has to actually read the physical page first.
And a huge percentage of Sanskrit literature has never existed as clean digital text at all.
It exists as handwritten manuscripts printed decades or centuries ago, scattered across archives, many of them fragile, faded, or written in regional script variants with no standardized digital font.
At IIT Bombay, a government-f funed digitization project has been steadily working through more than a hundred of these books, pulling from philosophy, literature, mathematics, medicine, and astronomy, feeding scanned pages through an OCR pipeline and then using a purpose-built correction tool to fix the inevitable misreadings the software introduces.
14.
Trained linguists currently work through that correction process by hand, cross-checking the machine’s output line by line against the original scanned page, which tells you plainly that even with AI doing the heavy lifting, the human expert hasn’t been removed from the loop.
The AI just changed what the human is spending their time doing.
That project itself sits under a much larger umbrella effort.
India’s government-f funded national mission for manuscripts which has spent years cataloging and attempting to preserve an estimated multi-million strong collection of manuscripts scattered across the country in libraries, temples and private family collections, many of them in Sanskrit and many of them at genuine risk of physical decay before anyone gets around to reading them again.
Digitization efforts like the IIT Bombay pipeline are in effect a race against time trying to convert as much of that physical archive into durable digital text as possible before insects, humidity or simple age make the job impossible.
Oxford University has been running a parallel effort out of its faculty of theology and religion using commercial AI handwriting recognition software called transcribus originally built for European medieval manuscripts and retraining it specifically to recognize Devonagari script.
The alphabet most classical Sanskrit is written in.
The goal is straightforward on paper but difficult in practice.
Take manuscripts that currently exist only as physical objects sitting in archives and turn them into searchable machine readadable digital text that scholars anywhere in the world can query instantly instead of requesting a physical viewing or relying on a handful of specialists who happen to have already read that specific manuscript.
Transcribus itself wasn’t originally designed with Sanskrit or Devonagari script in mind at all.
It grew out of a European research effort centered in Insbrook, Austria, aimed at reading medieval European handwriting.
And its use on South Asian material required retraining the underlying recognition model almost from the ground up to handle an entirely different alphabet, different liature patterns and different manuscript conventions than anything it had originally been built for.
That kind of cross application.
Taking an AI tool trained for one script in one continent and successfully repointing it at a completely different writing system has become a recurring theme across this whole field because building a dedicated Devonagari OCR system entirely from scratch is expensive and adapting an existing one is often faster.
It’s also not a one-way process.
Techniques developed specifically to handle Sanskrit’s dense compounding and phonetic merging have started showing up in research on other historically underresourced languages.
And Hellwig and Nerdict’s own 2018 paper explicitly demonstrated their Sandi splitting neural network working just as well on an entirely unrelated problem, splitting compound words in German without tea needing to be redesigned for the new language.
A tool built to solve a 2,000-year-old Indian grammatical puzzle turned out to generalize cleanly to a modern European one, which is exactly the kind of unplanned cross-pollination that makes this corner of computational linguistics more broadly useful than its niche subject matter might suggest at first glance.
Now, here’s where this stops being purely academic and starts touching something with real modern consequences.
In February of 2026, a Bangalore based research company called Atime announced it had built an AI system specifically to convert classical Sanskrit Ayurvedic medical manuscripts into structured searchable scientific data.
Led by chief scientific officer Letha Daml, the team used natural language processing to extract medicinal formulations, botanical references, and therapeutic principles buried in centuries old texts, then linked that extracted information directly to one of the world’s largest digital libraries of plant- derived molecules.
The pitch is that ancient Ayurvedic texts contain generations of observational medical knowledge about which plants treat which conditions recorded in a format modern pharmaceutical research has never been able to search efficiently because nobody could cross reference thousands of pages of classical Sanskrit against a molecule or database by hand.
If even a fraction of those old formulations point toward genuinely active compounds, an AI system built to make that connection searchable isn’t just a historical curiosity.
It’s a drug discovery shortcut hiding inside a 2,000-year-old text.
This isn’t the first time a classical Sanskrit medical text has produced something modern researchers took seriously.
Ayurvedic formulations have previously pointed pharmaceutical researchers towards specific plant compounds worth testing in laboratory settings.
And the broader field of ethnopharmacology studying traditional medical knowledge for leads on active compounds has produced genuine drug candidates from other traditional medical systems around the world before.
What’s different now is the scale.
A human researcher manually cross-referencing thousands of pages of classical formulation S against a modern molecular database, one plant and one compound at a time, might spend an entire career covering a fraction of what an AI system can screen in a matter of weeks.
That doesn’t guarantee any particular formulation will pan out under rigorous testing and Ayurvedic claims still have to clear the same evidentiary bar any other drug candidate does before anyone can call them medically validated.
But it does mean far more of that historical knowledge base gets a fair shot at being tested at all instead of sitting untransated and unindexed in an archive nobody has time to search by hand.
The sushruda samita in particular deserves a closer look here because it’s routinely cited by medical historians as containing some of the earliest known descriptions of reconstructive surgical technique anywhere in the world including methods for rebuilding a damaged nose using a flap of skin taken from elsewhere on the patient’s body.
A technique that European surgeons wouldn’t independently rediscover and formalize for well over a thousand years afterward.
The text also cataloges an enormous number of surgical instruments and describes methods for setting fractures, treating cataracts, and managing complicated child birth cases.
None of that historical medical detail is currently as accessible to modern researchers as it should be simply because so much of it hey never been converted into a searchable digital format cross-referenced against modern medical and pharmacological terminology.
That’s precisely the gap projects like atromeds are trying to close treating the text less as a historical artifact to be admired from a distance and more as an underexploited research data set waiting for the right computational tools to unlock it.
There’s a smaller almost unbelievable story sitting right alongside all of this well-funded academic and corporate work.
Just this past week, two 14-year-old students from Carerala, Kathleen Marie Jizison and Nvika D Na won a national robotics competition with a system they built themselves called Smearidi, a Sanskrit word that itself means memory.
They’re machined, physically handles fragile palm.
Leaf manuscripts, a material so delicate that mishandling it can cause the leaves to crack or the ink to fade permanently, and runs an automated scanning and OCR process designed specifically not to damage what it’s reading.
They’re headed to represent India at the World Robot Olympiad Internationals in Puerto Rico this December.
It’s a reminder that this entire field, despite the NASA myths and the wellunded institutional projects, is still young enough that two teenagers with a robotics kit can build something that meaningfully contributes to it.
Their system works in stages.
A manuscript first passes through a careful drying and stabilization process since palm leaves absorb humidity from the air and become dangerously brittle when handled while too dry or too damp.
And only after that stabilization step does the machine move the leaf into position for imaging.
According to Nvika, matching that imaging stage to the correct OCR settings for older, less standardized regional mallayam and Sanskrit character forms was one of the harder problems they ran into.
Since commercial OCR tools are generally tuned for modern printed fonts, not centuries old handwriting with its own regional quirks and abbreviations.
It’s the same core technical challenge every serious project in this space keeps running into at a different scale.
Whether it’s a twoerson student team in Carerala or a funded university lab in Oxford, historical handwriting simply doesn’t look like the clean standardized text most commercial AI tools were originally trained on.
And closing that gap requires retraining, patience, and in nearly every case documented here, a team of human experts checking the machine’s work.
None of this means classical Sanskrit scholarship has been fully cracked open by AI.
Far from it.
In 2022, a Cambridge doctoral student named Rishi Raj Popat solved a grammatical puzzle inside Panini’s original Ashtadayi that had reportedly stumped scholars for 2500 years.
Contradiction between two of Panini’s own rules that seem to produce grammatically invalid results whenever they were applied together in a certain order.
Raj Pipet’s solution came from painstaking manual close reading, not from a machine.
And it’s worth including here precisely because it proves the opposite of a simple narrative.
This isn’t a story where AI swoops in and finishes what human scholars couldn’t.
It’s a story where deep human expertise and modern computational tools are now running in parallel, each catching things the other one misses.
And the segmentation and tagging systems being built today are in part designed to make solutions like Raj Papets easier to find and verify at scale in the future instead of requiring one brilliant person and years of manual work every single time.
What Raj Papet actually found is worth spelling out a little further because it shows just how deep this rabbit hole goes, even within a text as thoroughly studied as Panini s own grammar.
Two of Panini’s rules, when both technically applied to certain word forms, produced a direct contradiction, and for centuries afterward, generations of scholars had assumed Panini himself intended for later rules in his list to simply override earlier ones whenever such a conflict came up.
A reasonably sensible assumption that nonetheless kept producing grammatically incorrect results in specific well-documented cases.
Raj Popat’s breakthrough was realizing Panini had actually intended something almost the reverse that between two conflicting rules, the one applying to the right-hand side of a word should take priority over the one applying to the left-hand side.
A subtle meta rule that resolves the contradiction cleanly across the cases scholars had been struggling with.
It’s a small technical adjustment with an outsized implication because it means one of the onos t rigorously studied texts in human intellectual history was still capable of hiding a genuine error in how it had been interpreted for two and a half millennia right up until 2022.
It’s also worth being direct about where the honest boundaries of all this sit because Sanskrit’s internet reputation attracts more than its fair share of exaggerated claims that have nothing to do with the actual research described here.
Viral posts periodically circulate, claiming ancient Sanskrit texts describe aircraft, advanced weapons, or technologies indistinguishable from modern engineering, treating a handful of poetic or mythological passages as literal blueprints.
Mainstream Sanskrit scholars and historians of science have consistently pushed back on these claims, pointing out that they typically rely on loose nonliteral translations stretched far beyond what the original grammar actually supports.
None of the projects covered here from Hellwig’s neural segmentation models to the IIT Bombay OCR pipeline to 8 TRED’s drug discovery work are built around or motivated by those kinds of claims.
They’re grounded in verifiable computational linguistics and pharmarmacology working with primary texts and publishing their methods openly for other researchers to check.
The distinction matters because as more of this archive becomes searchable, it’s going to become easier, not harder, to check exaggerated claims against what the actual original texts say, which cuts both ways.
It will settle some genuine ambiguities and it will almost certainly deflate a few more of the ah while viral claims already circulating that cuts against a common assumption people bring to this topic without fully realizing it.
The idea that once a text is old enough and mysterious enough, any claim about it becomes equally plausible because nobody can really check anymore.
The entire direction of this research points the opposite way.
A properly segmented, properly tagged Sanskrit T corpus is something a scholar anywhere in the world can query and cross reference in seconds, which makes vague, unsupported claims about what a text really says dramatically easier to challenge, not harder.
If anything, the projects described here are quietly raising the evidentiary bar for anyone making claims about classical Sanskrit literature, ancient or modern, mystical, or scientific.
Because for the first time, checking the primary sources at scale is becoming something closer to a weekend project than a career reserved for a handful of specialists.
So, what does it actually add up to?
Strip away the exaggerated NASA headline, and what’s left is a much more grounded, still genuinely remarkable story.
A 2500y old grammatical system built by a scholar with no concept of computing turns out to share real structural similarities with formal systems.
Computer scientists were independently inventing in the 20th century.
That similarity sat mostly unexploited for decades because the language itself in its written form is uniquely hostile to automated processing thanks to Sandi collapsing word boundaries into unreadable strings.
Only in the last several years has a combination of large manually built data sets, neural network segmentation models and transformer architectures made it realistic to process Sanskrit text at scale without a human expert manually parsing every sentence.
The payoff um isn’t confined to linguistics departments.
It’s reaching into pharmaceutical research, manuscript preservation, and now apparently national robotics competitions.
What happens next depends almost entirely on how much of the remaining Sanskrit textual record actually gets digitized.
Estimates for how many manuscripts still sit untouched in archives across India, Nepal, and elsewhere vary widely.
And nobody claims to have a precise number, but everyone working in this field agrees it’s an ele likely far more than what’s already been captured.
Every additional digitized text makes the underlying AI models measurably better because these systems learn from examples.
And right now, the biggest bottleneck isn’t computing power.
It’s the physical labor of getting fragile pages in front of a camera in the first place.
That’s exactly the gap projects like the IIT Bombay pipeline, the Oxford transcribus initiative, and even two teenagers with a homemade scanner in Carerala are quietly trying to close one manuscript at a time.
There’s also a feedback loop worth watching here that goes beyond any single project.
As tools like buy T5 Sanskrit get better at segmentation and tagging they lower the barrier for training the next generation of models because cleanly segmented correctly tagged text is exactly the fuel those systems need to improve further.
Right now the fuel is still relatively scarce measured in the hundreds of thousands of tagged words rather than the billions of tokens modern general purpose language models are typically trained on for widely spoken languages.
Every manuscript that gets digitized, OCR corrected, and properly segmented doesn’t just become one more readable text.
It becomes additional training material that makes every future model in this space slightly more capable, which is part of why researchers in the field describe this as a compounding effort rather than a series of isolated projects.
Zoom out far enough and the shape of this story starts to look familiar, even if the specific language involved is unusual.
Low resource languages, meaning languages without huge existing pools of clean digital text to train on, have become one of the more active frontiers in AI research generally, precisely because most of the biggest language models in the world were built almost entirely on English and a handful of other widely digitized languages.
Sanskrit sites in an unusual middle position within that broader category.
It has an enormous historical body of surviving text, likely more surviving primary source material than most modern minority languages can claim, but almost none of it exists yet in the clean standardized digital format modern AI training actually requires.
That combination, huge historical volume paired with minimal digital availability, is part of what makes Sanskrit such a distinctive test case for the field.
Techniques proven here on Panini’s grammar and Vadic Sandi rules are already informing how researchers approach digitization projects for other historically underresourced languages and scripts around the world.
From regional South Asian languages to classical texts in entirely different language families facing the exact same fundamental bottleneck, an abundance of surviving material and a shortage of hands and time to make it machine readable.
If this is the kind of story you want more of, hit subscribe on Histories YT.
Because as more of this archive gets digitized, the next real discovery could come from a text nobody has properly read in centuries.
And when it surfaces, we’ll be here to break it down.
A researcher at NASA noticed something real in an ancient grammar book 40 years ago, and the internet turned it into a myth almost nobody could verify.
It took until now with tools that researcher never had access to for anyone to actually go back and test what he was really pointing at.
One carefully tagged sentence, one restored manuscript page, one digitized surgical treatise, and one neural network at a time with a 2500year-old grammar book still setting the rules the machines are quietly learning to Follow.
