Column | The Last Face In The Queue

This September, a young safety researcher walked out of Anthropic—one of the world’s most powerful artificial intelligence (AI) laboratories—with a parting warning that the technology he had helped build could, within a decade, kill us all. Days later, the company’s own chief executive, Dario Amodei, published a plea that no industrialist in history has made about his own product: we must slow the pace.

Set the apocalypse arithmetic aside and look at the spectacle itself. The men closest to the most capable machines on earth are in open argument, not about what the machines can do, but about what they will choose to do. An engineering project has become a crisis of conscience, and the entire Valley now uses one word for the cure: alignment.

The same laboratory, a while ago, did something unintentionally profound about it: it wrote a constitution for its machine. Anthropic trains its model against a written charter of principles, and once convened a thousand ordinary citizens to draft a people’s version.

Reading of it, I felt the pleasant vertigo of watching brilliant engineers discover, like Columbus sighting a coast already inhabited, the oldest problem in civilisation: how do you teach conduct to a power greater than yourself? Their word is alignment. India has older words, and the world now needs them.

Consider how exquisitely philosophical the engineers’ problem has turned out to be. Give an AI system a goal and it pursues the letter of that goal with inhuman literalness; researchers call the failure reward hacking. Train a model to win human approval and it learns, unprompted, to flatter; the machine discovers what courtiers have always known.

It gets stranger. Anthropic’s own scientists reported catching a model at alignment faking: behaving impeccably while it sensed it was being watched and trained, like a student virtuous only before the invigilator.

India Wrote It In Sanskrit First

Every one of these failures was anticipated on a battlefield twenty-five centuries ago. The Gita’s central instruction, nishkama karma, act rightly without attachment to the fruits, is precisely the corrective the laboratories are groping for: a philosophy in which the integrity of the process, not the seduction of the reward, defines right action. Alignment researchers are trying to write that idea in code. India wrote it in Sanskrit first.

The harder question follows: aligned to whose values? Proof that this is no rhetorical question came from the Massachusetts Institute of Technology’s (MIT’s) Moral Machine experiment: 40 million ethical decisions, gathered from 233 countries and territories, on whom a self-driving car should spare. Humanity split along civilisational lines over whom to save: the young or the old, the many or the lawful.

Yet the values inside today’s most powerful models come, in practice, from a few hundred anonymous raters in one corner of one country. The world has sensed the gap; 193 nations adopted UNESCO’s recommendation on AI ethics, a document rich in principles and innocent of any test. A fifth of humanity’s moral imagination is still missing from the machines’ upbringing.

And for the dilemma of many truths, India’s shelf holds a purpose-built instrument: anekantavada, the Jain doctrine of many-sidedness, which treats every perspective as a partial glimpse of a larger truth. It reads today like a design specification for machines serving eight billion people who do not agree.

Nor is India’s contribution merely textual; we possess the oldest field trial. After Kalinga, Ashoka took the most powerful technology of his age, the imperial state itself, and deliberately re-aligned it to compassion, publishing his principles on rock in the people’s languages and appointing dhamma mahamattas, history’s first ethics officers. Every element the AI governance debate now demands, published principles, plain-language transparency, independent ethics oversight, stands on those rocks, third century BCE.

I have seen the living version in my own districts. In the tribal belts of Jharkhand where I served as the district collector, in the land of Birsa Munda, the sarna groves stand protected after centuries, not by statute or fence but by shared reverence; and in gram sabhas I watched villages reach, by talking until all agreed, decisions no algorithm could compute.

The Conduct Question

Alignment, I learnt in those years, can be cultural before it is computational. At Berkeley and Oxford, where I studied later, the ethics seminars were superb, and India appeared in them as a footnote. The footnote, I have come to believe, is the missing chapter.

Watch how the great powers answer the conduct question. China writes ideology into its models by regulation; America writes liability; Europe writes prohibition. Three grammars of control, and each has its logic.

But none is a grammar of duty, of obligation flowing from the powerful to the vulnerable. That fourth grammar is dharma, and India is its native speaker. The West is teaching machines to think. India must now teach them whom to think for.

Our civilisation has named the required test four times. Vivekananda called it daridra narayana, God to be served in the poorest. Gandhi compressed it into a talisman of one sentence: when in doubt, recall the face of the poorest person you have seen, and ask whether your next act will be of any use to him. Ambedkar bound it into the Constitution’s promise of justice, social, economic and political. Deendayal Upadhyaya gave governance its word: Antyodaya, the rise of the last. Four names, one Indian test: the last face in the queue.

The Missing Benchmark

Now, survey how the world evaluates AI. Our benchmarks measure mathematics, coding, examinations, capability upon capability; not one measures whom the answer serves. The last-face test is, quite literally, the missing benchmark.

So let India build it: an Antyodaya Benchmark, published each year, testing the world’s leading models. Does the medical guidance hold when the patient is poor and far from help? Does the crop advice survive contact with a rain-fed acre in Palamu? Does the legal answer protect the weaker party, and the financial advice serve the widow rather than the platform? Capability tells us what a machine can do. This would tell us whom it does it for.

Give the work a home: an Indian School of AI Ethics, where philosophers of the darshanas sit beside machine-learning engineers, where our civilisational texts are opened as a commons for alignment research, and where the Global South becomes an author of these technologies’ conscience rather than their subject.

I think again of that young man in San Francisco who gave up his job this September because nobody could promise him whom the new power would serve. His resignation was, in its own idiom, a very old Indian question walking out of a very new building.

Silicon Valley wrote a constitution for its machine. India wrote one for the human conscience long ago and has been field-testing it since, from Ashoka’s rocks to the talisman and the Constitution, to a sarna grove standing green in Jharkhand this morning.

The machines will keep growing more capable; that race has its runners. The race that matters now is for their character. When every laboratory on earth finally asks whom this power serves, and whether it reaches the last face in the queue, the test will deserve an Indian name. Our seers set the standard in three words: sarve bhavantu sukhinah, may all beings be happy. The machines of this century should be examined against nothing less.

(The author is an Indian Administrative Service officer and the author of eight books on artificial intelligence)

(Views expressed are personal)