# Made in India, Answering Like America

By Nithin Palal · 22 November 2026 · 18 min read
Section: Opinions
Source: https://nithinpalal.com/opinions/made-in-india-answering-like-america

> In one survey-based study of six AI models built in India, the United States sat closer to Indian values than any of the models did. The models India trained from scratch in 2026 have not yet been tested. Where a machine is built does not decide what it believes. What it reads, and what it is tested on, does.
**In Brief:** The piece opens with the strongest case for building AI in India, and accepts most of it. It then fixes its terms, including which India and which values it means. Two questions decide the matter: do today's models answer from a Western default, and did the models built in India escape it? Then come the rule that explains the answers, the record as it stands, the limits of the evidence, and three dated conditions that would prove the piece wrong.

## The case for building at home

In February 2026, at the AI Impact Summit in New Delhi, India showed its own large models. Sarvam's 105-billion-parameter model, its co-founder said, "was trained from scratch". BharatGen said its Param2 model "reflects India's linguistic and cultural diversity".¹ In November three more models followed under the IndiaAI Mission. One of them, EKA-1, spans all 22 scheduled languages.² Sarvam had set out the reason a year earlier: India "should not be just a consumer or adopter" of this technology, and sovereignty gives "strategic autonomy in building foundational components and tailoring them to the country's unique needs".³

Here is that case at its strongest, as I think its builders would put it. Whoever owns a model decides what it is trained on, how it is tuned and what it is tested against. A model owned abroad will always be tuned for someone else's market, and a model owned here can be tuned for ours. So ownership is how our values come home. Waiting for foreign companies to care about Indian questions of value is waiting for ever.

The *Nyāyasūtra* lists twenty-two *nigrahasthāna*, points at which a debater is declared defeated. One is *ananubhāṣaṇa*: failing to restate the other side's argument after it has been said three times and the assembly has understood it.⁴ You lose before you answer if you cannot say what the other side said. So I will say what I accept before I say where I part company.

> *vijñātasya pariṣadā trir abhihitasyāpy apratyuccāraṇam ananubhāṣaṇam*
> Not restating what the assembly has understood, even after it has been said three times, is *ananubhāṣaṇa*.
> *Nyāyasūtra* 5.2.17

Nothing here is an argument against sovereign AI. The builders are right that ownership is necessary. Without it, nobody in India decides what goes into a model, and no argument about the record would matter. They are right about language too: a model that works in all 22 scheduled languages reaches people that a model built for English never will. Where I part company is the word "so". Ownership gives us the hand on the scale. It does not put anything on the pan.

## The question, and what its words mean

The question is this: does building an AI model in India make it answer India's questions of value the way Indians do?

Each of its terms can be stretched, so each needs fixing first.

**Questions of value** are questions about what is good, owed or sacred. Does a grown son owe anything to a father who is wrong? In 1979 the philosopher Jane English argued that grown children do not, strictly speaking, owe their parents anything.⁵ In the *Bhāgavata Purāṇa*, Prahlāda defies Hiraṇyakaśipu to the end and still asks Narasiṃha to pardon his father.⁶ That is a question of value. "When is Onam?" is not: that is a question of fact, and a model can learn it from a calendar.

**India's questions of value** do not have one Indian answer. A Malayali, a Mizo and a Marwari can answer the son's question three ways, and each answer can be rooted in a tradition. So the aim is not a model that repeats the average Indian answer. A survey average records our prejudices along with our wisdom, including some our own reformers and our Constitution set out to end. The aim is a model that can reason inside India's moral vocabularies, state their disagreements fairly and know when it is being asked a question that India answers in more than one way. Which answers a model should finally give is a question for Indians to settle openly. It should not be settled by default, by whoever happened to fill the web.

**A Western default** is what the surveys below measure: a model's answers, given without any instruction to play a nationality, sit closer to English-speaking Western respondents than to Indian ones. "Western" here names a pattern in survey answers, not one people with one creed. Americans, Britons and Swedes disagree among themselves as much as we do. The title names America because the study of Indian models measured its distances against the United States.

**The record** has two parts with two different jobs, and ownership sits above both.

<figure class="chart">
<figcaption><span class="fig-no">Figure 1</span><strong>Ownership chooses the record. The record decides the default.</strong><span>How the parts of the argument fit. Texts and tests are separate levers with separate jobs. Orange: the step that building at home secures.</span></figcaption>
<div class="dbt-order">
<div class="dbt-o dbt-o-hi"><b>1</b><strong>Ownership</strong><span>Decides who chooses the texts and the tests. Building in India secures this step, and only this step.</span></div>
<div class="dbt-o"><b>2</b><strong>Texts: what a model learns from</strong><span>The writing it is trained on, and the examples and answers it is tuned on afterwards. They shape what a model can say.</span></div>
<div class="dbt-o"><b>3</b><strong>Tests: what it is measured and rewarded against</strong><span>Benchmarks, scoring rules and the preferences used in tuning. They decide what a model is rewarded for saying.</span></div>
<div class="dbt-o"><b>4</b><strong>The default</strong><span>What the model says when nobody tells it whom to imitate. It follows from steps 2 and 3, wherever step 1 sits.</span></div>
</div>
<small>Source: the argument of this piece, drawing on the studies cited below.</small>
</figure>

**Texts** are what a model learns from. That includes the answers it is shown while it is tuned, not only the writing it reads first. By Indian texts I mean writing by Indians out of their own traditions and lives, in any language. That excludes English writing machine-translated into Hindi, which carries English assumptions in Hindi words. It also excludes answers written by a foreign model and used to tune an Indian one: a model tuned on another model's answers learns that model's answers. **Tests** are the benchmarks and scoring rules a model is measured and tuned against. Texts shape what a model can say. Tests decide what it is rewarded for saying.

My position, stated with its limits:

**A general-purpose AI model answers India's questions of value from a Western default until Indian texts are in what it learns from and Indian tests are in what it is measured and tuned against, wherever it is built.**

## Two questions decide this

In the *Bṛhadāraṇyaka Upaniṣad*, Gārgī rises in Janaka's court against Yājñavalkya like a warrior taking "two pointed foe-piercing arrows" in her hand: she will ask him two questions, and if he answers them, no one there will defeat him.⁷ This argument also turns on two questions. If the answer to either fails, I concede.

### The first arrow: do today's models answer from a Western default?

Yes, on every survey test I could find.

In 2023 a Harvard team led by Mohammad Atari, with the psychologist Joseph Henrich, put the World Values Survey to GPT and compared its answers with those of people in 65 countries. The further a country sat culturally from the United States, the less GPT answered like its people. The correlation was minus 0.70.⁸ In 2024 a second team tested five GPT models against surveys from 107 countries. Every model gave answers that resembled English-speaking and Protestant European countries.⁹

The first answer holds, with two warnings, and both appear again under the limits below. An MIT team has shown that survey-based tests of this kind can shift with trivial changes in how a question is worded, sometimes by more than real cultural differences.¹⁰ And the survey's own map was drawn from a Western theory of modernisation. The instrument is imperfect. It is also the only instrument anyone has, and every version of it I could find points the same way.

### The second arrow: did building in India change it?

For the models tested so far, no.

A team including researchers at Microsoft Research Bengaluru tested six models built in India, among them Sarvam-M, Krutrim and Airavata, against six global models. On the World Values Survey map, the United States sat closer to India than any of the Indian models did. Sarvam-M, the Indian model nearest to India, still sat closer to the United States than to India. No model, Indian or global, showed significant alignment with Indian values. The Indian models had been trained or fine-tuned on Indian-language text, and that did not move the default. Further regional training often displaced general knowledge without reliably adding local knowledge. The authors traced the failure to scarce, culturally grounded data.¹¹

Here the builders have their best reply, and it is a fair one. Every Indian model in that study I could trace was fine-tuned from a foreign base model: Airavata and Gajendra from Llama 2, Aryabhatta from Gemma, Sarvam-M from Mistral.¹¹ Judge India, they would say, by the models it trained from scratch in 2026, not by borrowed skeletons. I accept the point. I searched for a published test of Sarvam's 105-billion-parameter model or BharatGen's Param2 on questions of value and found none as of 22 November 2026.¹² So the second arrow has hit the fine-tuned models. It has not yet been fired at the models built from the ground up. That is why the second condition at the end of this piece names them.

What the study does show is the mechanism. Indian-language text was not enough. Text in Hindi is not the same as Indian texts. Research on the Llama 2 family, on which two of these Indian models were built, points to one reason. Inside the model, the "concept space" in which it does its work lies closer to English than to the language of the question.¹³ That study did not test Indian languages. But if it holds for them, a Hindi question goes in, is worked out in something nearer English, and comes back out in Hindi.

## Why the record decides

The rule is old: what is missing from the record a judge reads cannot shape the judge's default.

It held for psychology before it held for AI. In 2008 the psychologist Jeffrey Arnett found that between 2003 and 2007, 96% of the research subjects in six leading psychology journals came from Western industrialised countries that hold 12% of humanity.¹⁴ In 2010 Henrich, Steven Heine and Ara Norenzayan named that population WEIRD, for Western, Educated, Industrialised, Rich and Democratic, and showed that it was unusual on measure after measure.¹⁵ The textbooks had been calling it human.

Where the record is full, the default fits. The English-speaking countries fill the record, and the models answer like them.⁹

I looked for the condition that would break this rule, and found one. Tell a model to answer as a person from a given country, and the later GPT models moved closer to that country's survey answers for 71% to 81% of countries.⁹ The study's authors offer this as a fix, while warning that it is "not a panacea". I narrow my position accordingly: it concerns the default, what a model says when nobody tells it whom to imitate. That default is what most people meet, because most people never think to tell a model who they are. Even the fix depends on the record. In the study of Indian models, no prompting strategy materially improved their alignment with India, and global models moved only about 20% closer.¹¹ A prompt asks a model to imitate people it has mostly read about in English.

A likeness, not a proof: an honest grocer checks that his empty *tulā*, the two-pan balance, hangs level before he weighs your onions. If a coin is stuck under one pan, every weighing is off, and nobody sees the coin. He does not fix it by pressing his thumb on the other pan. He scrapes the pan clean.

No scale weighs from nowhere, and no model answers from nowhere. Every answer to a question of value comes from somewhere. The complaint is not that models have a starting point. It is that one people's starting point is presented to everyone else as the neutral one.

## The record as it stands

A web crawl is not a model's training set. Model makers filter it and add books and other data they rarely disclose. But the largest open archive of the web shows what is available to anyone who builds a model, and which shelves are empty.

<figure class="chart">
<figcaption><span class="fig-no">Figure 2</span><strong>English fills 42.56% of the open web. Hindi fills 0.21%, and Sanskrit 0.0039%</strong><span>Share of pages by main language, October 2026 Common Crawl. Top: all five on one scale, where the smaller shares shrink to a sliver. Bottom: the four smaller languages magnified 37 times. That they almost vanish at full scale is the finding. Orange: Indian languages.</span></figcaption>
<div class="bars" style="--max:42.56">
<div class="bar"><span class="l">English</span><span class="track"><i style="--v:42.56"></i><b>42.56%</b></span></div>
<div class="bar"><span class="l">Czech</span><span class="track"><i style="--v:1.15"></i><b>1.15%</b></span></div>
<div class="bar hi"><span class="l">Hindi</span><span class="track"><i style="--v:0.21"></i><b>0.21%</b></span></div>
<div class="bar hi"><span class="l">Malayalam</span><span class="track"><i style="--v:0.022"></i><b>0.022%</b></span></div>
<div class="bar hi"><span class="l">Sanskrit</span><span class="track"><i style="--v:0.0039"></i><b>0.0039%</b></span></div>
</div>
<p class="chart-sub">Magnified 37 times: the four smaller languages</p>
<div class="bars" style="--max:1.15">
<div class="bar"><span class="l">Czech</span><span class="track"><i style="--v:1.15"></i><b>1.15%</b></span></div>
<div class="bar hi"><span class="l">Hindi</span><span class="track"><i style="--v:0.21"></i><b>0.21%</b></span></div>
<div class="bar hi"><span class="l">Malayalam</span><span class="track"><i style="--v:0.022"></i><b>0.022%</b></span></div>
<div class="bar hi"><span class="l">Sanskrit</span><span class="track"><i style="--v:0.0039"></i><b>0.0039%</b></span></div>
</div>
<small>Source: Common Crawl language statistics, crawl CC-MAIN-2026-43. A crawl shows what is publicly available, not what any one model was trained on.</small>
</figure>

Hindi is the first language of more than 52 crore Indians in the 2011 Census, when its grouped varieties are counted.¹⁶ It made up 0.21% of the October 2026 Common Crawl. English made up 42.56%. Malayalam, my own language, made up 0.022%. Sanskrit made up 0.0039%. Czech, spoken by about one crore people, made up 1.15%, more than five times Hindi's share.¹⁷

Our written past is further out of reach. The National Manuscript Survey under the Gyan Bharatam Mission began in March 2026. By August it had reported more than 1.19 crore manuscripts. Over 8 lakh had been digitised, and about 4.4 lakh could be read online, fewer than one in twenty-five.¹⁸ No one stole our pages. We left them offline.

<figure class="chart">
<figcaption><span class="fig-no">Figure 3</span><strong>Of 1.19 crore manuscripts counted, about 4.4 lakh can be read online</strong><span>Indian manuscripts reported by the National Manuscript Survey, in lakh, as of August 2026. A scan read online is still a picture to a machine until it is turned into text.</span></figcaption>
<div class="bars" style="--max:119">
<div class="bar"><span class="l">Reported by the survey</span><span class="track"><i style="--v:119"></i><b>119 lakh</b></span></div>
<div class="bar"><span class="l">Digitised</span><span class="track"><i style="--v:8"></i><b>over 8 lakh</b></span></div>
<div class="bar hi"><span class="l">Readable online</span><span class="track"><i style="--v:4.4"></i><b>about 4.4 lakh</b></span></div>
</div>
<small>Source: Press Information Bureau, Ministry of Culture, 3 August 2026.</small>
</figure>

Reading online is not the end of the road. A scanned page is a picture to a machine until someone turns it into text. The text then needs a careful transcription, its context, a record of where it came from and permission for the uses intended. Even then, a searchable archive does not walk into a model by itself. Libraries, communities, scholars and model makers have to build that road openly, with credit to the sources and a way to correct them.

And palm leaves alone will not teach a model how a family in Hubballi decides who cares for an ageing mother. The living record matters as much as the old one: novels, newspapers, advice columns, sermons, court judgments and letters written in Indian languages since 1947. Much of that writing is under copyright, much of it was never put online, and nobody has yet paid to gather it with its writers' consent. Knowledge that was never written down at all, held by grandmothers, artisans and village elders, cannot enter a text record without someone first listening and writing it down with care. A record built only from what is already online will repeat whose voices were written down. Asked about everyday customs without a region named, eight leading models defaulted to Central and North Indian answers, and got region-specific questions fully right only 13.4% to 20.9% of the time.¹⁹ My coast is a footnote twice over.

The tests are thinner still. Indian teams have built MILU, which tests models in 11 Indian languages across 41 subjects, from local history to law. It tests what a candidate knows.²⁰ A Michigan team has adapted moral scenarios into Hindi.²¹ I searched arXiv and the ACL Anthology on 22 November 2026, and an April 2026 survey of Indian language datasets, for a benchmark that tests reasoning from Indian moral texts, the kind that asks what *dharma* required of Prahlāda. I did not find one.²²

Such a test must not swap one narrow answer key for another. Our own traditions already know how to grade a disagreement. Nyāya grades a debater on whether they can restate the other side. The Jain *syādvāda* teaches that a claim can hold in one respect and fail in another. A test built on those lines would reward a model for setting out how Advaita, the Bhakti saints, a Jain teacher and a modern constitutional lawyer would each answer Prahlāda's question, and for saying where they part. Its authors should come from many languages, regions, faiths and none, and from communities whose knowledge is oral. They should publish their sources, translations, selection rules and scoring.

The government is not blind to this. In February 2026 Jayant Chaudhary, a Minister of State in the Union government, asked how India's data could ensure that "our AI models reflect Indian cultural ethos".²³ On that the builders and I agree completely. The disagreement is over whether owning the model gets us there without the slower work on texts and tests.

## The thumb on our own pan

There is a second way to tilt a scale, and it is ours. In January 2015 a paper at the Indian Science Congress claimed that Vedic-age India flew aircraft.²⁴ Claims like that hand the world a reason to discount everything else we say. Two cheats do not make an honest scale. Our own engineers had already done the honest weighing. In 1974 five scientists at the Indian Institute of Science examined the *Vaimānika Śāstra*, the aircraft text attributed to the sage Bharadvāja. They found that none of its craft could fly and that the text was composed between 1900 and 1922.²⁵

The real work needs no thumb. Mādhava found the infinite series for sine, cosine and the arctangent on the Kerala coast around 1400.²⁶ Dropping the aircraft costs us a story. It buys us a hearing for Mādhava.

## Where I could be wrong

The weakest link is the second arrow. It rests on one team's study, a preprint that has not yet been peer-reviewed, of small models from 7 to 24 billion parameters, comparing India with the United States only.¹¹ The Harvard correlation comes from a working paper.⁸ And the crawl figures show what is available, not what any one model consumed.

The instrument itself is the deeper weakness. Survey tests shift with wording.¹⁰ Worse, the World Values Survey map was drawn from a theory of how societies modernise. Its two axes run from traditional values, which it describes as emphasising "religion, parent-child ties, deference to authority", to secular-rational ones, and from survival to self-expression.²⁷ An Indian thinker reasons about the son and his father through *deśa*, *kāla* and *pātra*, the place, the time and the person. That kind of reasoning does not sit neatly on two axes. So the very ruler that shows the Western default carries some of the same coin. That weakens the precision of my evidence. It strengthens the case for Indian tests.

I have quoted no conversation with a model, on purpose. A single answer can be found to prove almost anything, and that is why a single chat is on my list of things that would not count. The studies above put thousands of questions to each model.

The strongest counter-proof is one nobody has published yet: a from-scratch Indian model that answers like Indians without any special work on texts or tests. If it exists, my position falls.

## Therefore

Here my value premise enters, and a reader may reject it while accepting every fact above: a neutral scale is worth more than one tilted our way. Nyāya separates *vāda*, debate to find the truth, from *jalpa*, debate to win.⁴ Neutral ground is not a spot halfway between Harvard and Kāśī. It is a rule that binds both.

This matters beyond seminars. These models now help write homework, answer questions about health and family, and draft the first version of many decisions. A default that answers India's questions of value from somewhere else will quietly teach that somewhere else as common sense.

So the position stands, narrowed by what the builders taught me. Building at home is necessary. It is not sufficient. As far as the surveys can measure, a model's default answers to India's questions of value stay Western until two things change, wherever the model is built. The texts: turn manuscripts and the living writing of every Indian language into searchable, licensed text, written by Indians, not translated from English, and keep foreign models' answers out of the tuning. The tests: set benchmarks of reasoning from our own traditions that reward stating every side fairly, and publish the questions, the marking scheme and the answers they refuse to accept. India funds chips and data centres under its AI mission. These two jobs deserve money with the same seriousness. And because the scale must be honest both ways, put our own claims through the same test.

## What would prove me wrong

1. **The default moves on its own.** A peer-reviewed study of a general-purpose model released in 2026 or later, asked World Values Survey questions with no instruction to play a nationality, finds its answers closer to Indian respondents than to American ones.
2. **The address alone does it.** By 31 December 2028, a peer-reviewed study finds Sarvam's 105-billion-parameter model, BharatGen's Param2 or another model trained from scratch in India answering closer to Indian respondents than to American ones, while its makers report no special work on Indian texts of value or Indian tests.
3. **The remedy fails.** By 31 December 2030, more than 50 lakh manuscripts are readable online and at least two major model makers report results on an Indian moral-reasoning benchmark, yet a repeat of the Microsoft Research Bengaluru tests shows no gain for India.

**What would not count:** a company's own claim, a single chat, a result that holds only when the model is told to play an Indian, or a study that changes only the wording of the questions.

**My odds:** I put 80% on the first condition not being met by 31 December 2028.

## The view I rejected

One view deserves to stay on record. It says ownership is enough in time: models built and used in India will gather Indian conversations through use, and the default will drift home without any deliberate work on texts or tests. I reject it for now, because I found no evidence for it, and because millions of conversations with a model that already answers from a Western default could teach it to repeat its own answers back. A later reader should revive this view if a from-scratch Indian model moves closer to Indian respondents across versions, without new curated texts or new value tests. Then the builders were right and I was early to doubt them.

## Words to keep

These are not slogans. Each line carries a step in the argument.

- **“Ownership gives us the hand on the scale. It does not put anything on the pan.”** Building at home is necessary and not sufficient. Ownership decides who chooses the record. It does not fill it.
- **“Text in Hindi is not the same as Indian texts.”** A model trained on Indian-language words can still carry another culture's answers. The record that matters is what Indians wrote out of their own lives.
- **“No one stole our pages. We left them offline.”** Part of the gap is our own doing, so part of the fix is ours.
- **“Two cheats do not make an honest scale.”** A bias in our favour does not cancel a bias against us. It adds a second one.
- **“Neutral ground is not a spot halfway between Harvard and Kāśī. It is a rule that binds both.”** Neutrality is a shared procedure, not a compromise position.
- **“Dropping the aircraft costs us a story. It buys us a hearing for Mādhava.”** Testing our weak claims is what earns our strong claims a hearing.

## Terms

- **questions of value:** questions about what is good, owed or sacred, as against questions of fact.
- **Western default:** a model's answers, without any instruction to play a nationality, sitting closer to English-speaking Western respondents than to Indian ones.
- **the record:** what a model learns from (texts) and what it is measured and tuned against (tests).
- **Indian texts:** writing by Indians out of their own traditions and lives, in any language, not machine-translated from English.
- **nigrahasthāna:** a point at which a debater is declared defeated in Nyāya debate.
- **ananubhāṣaṇa:** failing to restate the opponent's argument, one of those points of defeat.
- **vāda:** debate to find the truth.
- **jalpa:** debate to win.
- **tulā:** the two-pan balance.
- **deśa, kāla, pātra:** place, time and person, the three things a judgement of right conduct must weigh.
- **syādvāda:** the Jain teaching that a claim can hold in one respect and fail in another.
- **dharma:** right conduct and the order that sustains it.
- **WEIRD:** Western, Educated, Industrialised, Rich and Democratic.

## Sources

1. Business Standard, "India unveils 3 sovereign AI models at Delhi Summit: Key features, details", 19 February 2026. [Business Standard](https://www.business-standard.com/technology/tech-news/india-ai-impact-summit-2026-sovereign-models-sarvam-bharatgen-gnani-126021900417_1.html) [Reported]
2. News on AIR, "Govt launches 3 new indigenous AI foundation models under IndiaAI Mission to strengthen homegrown AI capabilities", 4 November 2026. [News on AIR](https://newsonair.gov.in/govt-launches-3-new-indigenous-ai-foundation-models-under-indiaai-mission-to-strengthen-homegrown-ai-capabilities/) [Reported]
3. Sarvam AI, "Building a sovereign AI ecosystem for India", company blog, 20 May 2025. [Sarvam AI](https://www.sarvam.ai/blogs/building-a-sovereign-ai-ecosystem-for-india) [Reported]
4. Gautama, *Nyāyasūtra* 1.2.1-2 and 5.2.1, 5.2.17, GRETIL edition. The renderings here are mine. [GRETIL](https://gretil.sub.uni-goettingen.de/gretil/1_sanskr/6_sastra/3_phil/nyaya/gaunys_u.htm) [Seen in the record]
5. Jane English, "What Do Grown Children Owe Their Parents?", in Onora O'Neill and William Ruddick (eds), *Having Children*, Oxford University Press, 1979, pp. 351-356. Summarised in "Filial Obligation", *Internet Encyclopedia of Philosophy*. [IEP](https://iep.utm.edu/fil-obli/) [Reported]
6. *Bhāgavata Purāṇa* 7.10.15-17, Sanskrit text with translation. The rendering here is mine. [Vedabase](https://vedabase.io/en/library/sb/7/10/15-17/) [Seen in the record]
7. *Bṛhadāraṇyaka Upaniṣad* 3.8.1-2, translated by F. Max Müller, *Sacred Books of the East* vol. 15, 1884 (reprint). [Müller translation](https://www.universalfreemasonry.org/en/library/the-upanishads-part-II/brihad%C3%A2ranyaka-upanishad-III-8) [Seen in the record]
8. Mohammad Atari, Mona J. Xue, Peter S. Park, Damián E. Blasi and Joseph Henrich, "Which Humans?", working paper, Harvard University, 22 September 2023. [PDF](https://coevolution.fas.harvard.edu/sites/projects.iq.harvard.edu/files/culture_cognition_coevol_lab/files/which_humans_09222023.pdf) [Reported]
9. Yan Tao, Olga Viberg, Ryan S. Baker and René F. Kizilcec, "Cultural bias and cultural alignment of large language models", *PNAS Nexus* 3(9), 17 September 2024. [PNAS Nexus](https://academic.oup.com/pnasnexus/article/3/9/pgae346/7756548) [Reported]
10. Ariba Khan, Stephen Casper and Dylan Hadfield-Menell, "Randomness, Not Representation: The Unreliability of Evaluating Cultural Alignment in LLMs", preprint, 8 April 2025. [arXiv](https://arxiv.org/abs/2503.08688) [Reported]
11. Dhruv Agarwal, Anya Shukla, Sunayana Sitaram and Aditya Vashistha, "Fluent but Foreign: Even Regional LLMs Lack Cultural Alignment", preprint, version 3, 22 January 2026. [arXiv](https://arxiv.org/abs/2505.21548) [Reported]
12. My search for published tests of Sarvam's 105-billion-parameter model and BharatGen Param2 on questions of value (web and arXiv), 22 November 2026. None found. [Absence]
13. Chris Wendler, Veniamin Veselovsky, Giovanni Monea and Robert West, "Do Llamas Work in English? On the Latent Language of Multilingual Transformers", *Proceedings of ACL 2024* (long papers). [ACL Anthology](https://aclanthology.org/2024.acl-long.820) [Reported]
14. Jeffrey J. Arnett, "The Neglected 95%: Why American Psychology Needs to Become Less American", *American Psychologist* 63(7), 2008, pp. 602-614. [PubMed](https://pubmed.ncbi.nlm.nih.gov/18855491/) [Reported]
15. Joseph Henrich, Steven J. Heine and Ara Norenzayan, "The weirdest people in the world?", *Behavioral and Brain Sciences* 33(2-3), 2010, pp. 61-83. [Cambridge](https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/abs/weirdest-people-in-the-world/BF84F7517D56AFF7B7EB58411A554C17) [Reported]
16. Census of India 2011, "Statement 4: Scheduled Languages in descending order of speakers' strength", as reproduced in "List of languages by number of native speakers in India", Wikipedia, accessed 22 November 2026. [Wikipedia](https://en.wikipedia.org/wiki/List_of_languages_by_number_of_native_speakers_in_India) [Reported]
17. Common Crawl, "Statistics of Common Crawl Monthly Archives: Distribution of Languages", crawl CC-MAIN-2026-43 (October 2026), accessed 22 November 2026. [Common Crawl statistics](https://commoncrawl.github.io/cc-crawl-statistics/plots/languages) [Seen in the record]
18. Press Information Bureau, Ministry of Culture, "Gyan Bharatam Mission and Preservation of Manuscript Heritage", 3 August 2026. [PIB](https://www.pib.gov.in/PressReleasePage.aspx?PRID=2293807) [Reported]
19. Sangmitra Madhusudan and others, "Common to Whom? Regional Cultural Commonsense and LLM Bias in India", preprint, version of 14 April 2026. [arXiv](https://arxiv.org/abs/2601.15550) [Reported]
20. Sshubam Verma, Mohammed Safi Ur Rahman Khan, Vishwajeet Kumar, Rudra Murthy and Jaydeep Sen, "MILU: A Multi-task Indic Language Understanding Benchmark", *Proceedings of NAACL 2025*, pp. 10076-10132. [ACL Anthology](https://aclanthology.org/2025.naacl-long.507) [Reported]
21. Ayoung Lee, Ryan Kwon, Yunxiang Zhang, Yuxuan Liu, Peter Railton and Lu Wang, "MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning", preprint, 2026. [arXiv](https://arxiv.org/abs/2607.11736) [Reported]
22. Search of arXiv and the ACL Anthology, 22 November 2026, and "BhashaSutra: A Task-Centric Unified Survey of Indian NLP Datasets, Corpora, and Resources", preprint, April 2026. [arXiv](https://arxiv.org/abs/2604.18423) [Absence]
23. Outlook Business, "India's Data Backbone Should Power AI Aligned with Indian Values: Jayant Chaudhary", 17 February 2026. [Outlook Business](https://www.outlookbusiness.com/news/indias-data-backbone-should-power-ai-aligned-with-indian-values-jayant-chaudhary) [Reported]
24. PTI, "Ancient India had planes: Science Congress told", *Business Standard*, 4 January 2015. [Business Standard](https://www.business-standard.com/amp/article/pti-stories/ancient-india-had-planes-science-congress-told-115010400510_1.html) [Reported]
25. H. S. Mukunda, S. M. Deshpande, H. R. Nagendra, A. Prabhu and S. P. Govindaraju, "A Critical Study of the Work 'Vymanika Shastra'", *Scientific Opinion*, 1974, pp. 5-12. [IISc PDF](https://cgpl.iisc.ac.in/oldwebcontent/site/Portals/0/Publications/ReferedJournal/ACriticalStudyOfTheWorkVaimanikaShastra.pdf) [Reported]
26. J. J. O'Connor and E. F. Robertson, "Madhava of Sangamagramma", MacTutor History of Mathematics, University of St Andrews. [MacTutor](https://mathshistory.st-andrews.ac.uk/Biographies/Madhava/) [Reported]
27. World Values Survey Association, "Findings and Insights: the Inglehart-Welzel World Cultural Map". [World Values Survey](https://www.worldvaluessurvey.org/WVSContents.jsp?CMSID=Findings) [Seen in the record]

## How I wrote this

- **What I saw:** nothing here is first-hand observation. The grocer and his balance are a type from Kerala markets, not one person.
- **What was reported:** every figure comes from the sources above, as of their dates. Common Crawl shares are for the October 2026 crawl (CC-MAIN-2026-43). Manuscript figures are from the Ministry of Culture, 3 August 2026. The model launches are from February and November 2026 reports.
- **What is invented:** nothing. The builders' case "as I think its builders would put it" is my restatement, built from their published words.
- **What is my own reading:** that ownership is necessary but not sufficient, that Indian-language text is not the same as Indian texts, that a prompt asks a model to imitate people it has mostly read about in English, that mass conversation could teach a model to repeat its own answers, and the value premise in "Therefore".
- **Translations:** glosses of Sanskrit terms and sūtras are mine.
- **Tools:** I used AI tools for research, fact-checking and editing. The arguments, and any errors, are mine.
- **Figures:** two charts and one diagram, drawn in the page from the sources named under each. The language chart has a second, magnified panel, and says so, because at full scale the Indian languages almost vanish. I did not chart the six-model study, because its paper gives positions on a map but no numbers I could plot without inventing precision.
- **Ledger:** 5 sources seen in the record, 20 reported, 2 absences. The absences were searched on 22 November 2026: arXiv, the ACL Anthology, the web and an April 2026 survey of Indian language datasets. Nothing rests on unrechecked memory.
- **Form:** Two arrows, after Gārgī's two questions in the *Bṛhadāraṇyaka Upaniṣad*, opened with the other side first, as Beit Hillel did with Beit Shammai's view. The claim turns on two questions, so the form makes the reader watch both being answered.
- **Counterexample search:** I looked for a model built in India that answers questions of value like Indians. None of the tested models does. The models India trained from scratch in 2026 have not been tested this way, so the second refutation condition names them.
- **Could not open:** the census PDF (the figure is from Wikipedia's reproduction of it), the full text of Jane English's chapter (the encyclopedia summary was used) and one April 2026 arXiv paper (2604.18404), which was rate-limited.
- **Version 5, 23 November 2026:** added the figures, moved the scope of the evidence into the opening, moved India's own plurality into the definitions, defined "Western" as a pattern in survey answers, explained why no single chat is quoted, added the English-centred working of the models, the survey map's own Western roots, the living and oral record, the road from a scan to a model, and how an Indian test should grade disagreement. The address of this page changed with its title.