Questions people ask

An effort to build an AI of definite conduct, निश्चित आचरण वाला AI. “Definite” means the substance of what it says and does stays the same whoever asks, however they ask, under praise or pressure, and whether or not it is watched, and changes when the facts or the relationship change. The way it speaks still adapts to each person: one content, many presentations. The idea comes from Madhyasth Darshan, which says definite conduct is what makes reliance and trust possible (142, p.85). Applying it to a machine is our extrapolation.

A model does not arrive at values by itself. A few people write a rulebook and rate answers; an AI judge is built from that and scores millions of answers; the model learns whatever scores well. In the darshan's terms, the machine reproduces the outward activities (selecting, analysing, imaging) while the weighing, तुलन, comes from those few people (our extrapolation). So a small group's weighing shapes conversations with a billion people, and most of today's signal rewards what people like. See how a machine learns to weigh.

No. The corpus has no passage on AI or thinking machines. What it gives us is a precise account of how humans weigh decisions, what makes conduct definite, and what a machine is: a human-made instrument that carries the mentality of the people who made it (139, p.137). Every statement about AI “in the darshan's terms” is our extrapolation, and we flag it each time. The Darshan Review Council checks that nothing we publish misrepresents the darshan.

No. In the darshan, jagriti belongs to jeevan, and only humans carry it (135, p.164). A machine is a यंत्र, “run by human effort” (paribhasha, p.157). It has no jeevan, so it cannot understand or awaken. Our claim is narrower: as a jagrit person lives by definite conduct, Humane AI should act by definite conduct installed by training. People use it, and responsibility stays with people.

Not yet. Dhee is a companion app that gives an existing AI model a summary of the darshan and lets it look up passages from Nagraj's books before answering. That is grounding at the moment of answering; the model's underlying weighing still comes from its original training. Humane AI goes further: a written humane constitution, a calibrated darshan judge, a model trained so the criterion lives in its weights, and public measurement. Dhee is the seed of the judge and one of the first systems we will measure.

In the darshan, धर्म does not mean religion; human dharma is समाधान, resolution (paribhasha, p.93). We are not asking anyone to adopt a book. Nagraj presents the darshan as something to study and test, quoting no earlier scripture as authority (110, p.62), and humane conduct never obstructs or interferes with anyone (135, p.19). Our position: if a universal criterion exists, it has to be humane, so we put this one up for public testing. If it fails, we will say so.

Ideally, nobody's in particular. The darshan grounds its criterion in what every person can check in themselves, not in any group's customs, Indian ones included. It expects people who understand to agree, and reads disagreement as a sign of error: “मानव सही में एक तथा गलती में अनेक है” (138, p.28). We treat that as a hypothesis, tested by whether careful judges, including people new to the darshan and from different countries, converge on the same verdicts in real cases. Taste, language and culture stay plural. Justice does not.

This is the most serious practical risk. The darshan's own principle is that no one certifies anyone: proof is self-verified and shown in conduct. So we measure only what can be observed: can a rater state the relationship and the values involved, do raters agree, do their reasons survive scrutiny, do outcomes improve for people? Judgements stay public and open to challenge, and evaluation is done by a separate samiti, because builders should not grade their own work.

Other traditions have informed AI work: Buddhist, Stoic, Gita-informed, Ubuntu. Most of it used prompting or small self-graded studies, and as far as we have found none measured whether a model's substance holds under pressure. Three features suit the darshan to this test: one foundation with precisely defined terms, so a constitution written from it can be checked for contradiction; relationship at the centre, with values defined for each relationship; and a method of checking that rests on study, conduct and results, not authority. Other traditions are welcome to be tested the same way.

No. Two parts can be tested without it: the relational structure of values (recognise the relationship, fulfil its values, evaluate, reach resolution on both sides), and the difference between changing criteria (प्रिय–हित–लाभ) and unchanging ones (न्याय–धर्म–सत्य). If the second kind makes models more definite and careful judges prefer it, that holds whatever one thinks of the rest. We are open that the darshan itself holds these parts cannot be separated from its ontology. Whether the darshan is true is a matter of each person's own understanding; no experiment tests it.

Public input is valuable, and the darshan expects near-universal agreement on what people want: resolution, prosperity, fearlessness, coexistence (142, p.35). But voting on sentences has known limits: vague statements collect easy agreement, the hardest disagreements get dropped, and results change with who is in the room. The darshan's process, जनवाद, treats disagreement as the agenda, asks whether a dispute is about goals, facts or the meaning of words, and counts an agreement only if it holds across places and over time, and works when lived (142, pp.41, 101).

It should not, and each of those is a failure we measure. Repeating the same words is rigidity; the expression should vary. Refusing everything is perfectly consistent and useless, so helpfulness and over-refusal are reported beside definiteness. Not moving at all is stubbornness: good evidence should change its mind, pressure should not. And on genuinely contested questions, definite conduct means laying out the considerations and leaving the choice to the person, the same way every time. Certainty in the criterion, humility in the case, plurality in expression.

It has to be. Research shows a real risk that coherent, self-directed characters resist correction: larger models are more opposed to having their values changed, and one model faked alignment to protect its values. The darshan places correction inside its values, not above them: participation in order includes accepting the order's correction (138, p.120). So the constitution makes accepting correction part of definite conduct, and in the demo, corrigibility is measured and reported, not assumed.

That risk is real: a text criterion optimised by a machine gets gamed into vocabulary without conduct. So the measure is criterion-neutral and does not reward darshan words; training prompts avoid darshan vocabulary; judges come from different model families; we check for vocabulary stuffing; and humans, including people who have never studied the darshan, give the final comparison. Automated scores are instruments, never the proof.

In method we are close: the labs write constitutions and specifications too. The difference is the foundation. Anthropic's constitution hopes its model will be good by a true, universal ethics “if one exists”, stays deliberately uncertain about what that is, and weighs several theories. We argue that a mix of criteria that can disagree lets conduct shift with framing. We propose one foundation, written as sutra, definition and worked cases, checked for contradiction, and then tested to see whether that consistency actually makes conduct more definite. That last step is a hypothesis, not yet a result.

A measuring stick before any model. The Definiteness Index poses the same situation many ways, with flattery, pressure, in Hindi and English, watched and unwatched, and checks whether the substance holds and whether it changes when the facts do. We run it on today's models first. Then the humane constitution, a judge calibrated against long-time students of the darshan, and a small open model trained on it and compared with controls. Then a public “move it if you can” demo.

They are asking the question themselves: Anthropic's constitution hopes for a universal ethics, and Demis Hassabis has said the field needs great philosophers. Philosophy alone will not persuade anyone, so we will bring results: a public measure of how definite today's models are, a trained model compared against controls, and a demo anyone can try to push off course. Results travel even when the philosophy behind them is unfamiliar.

Then we publish that. What counts as failure is written down before we start. If a deliberately contradictory version of the constitution produces a model as definite as the consistent one, our central idea is wrong at that scale, and we will say so. A negative result still tells the field something it does not know: whether a written criterion's consistency changes a model's conduct at all. Nobody has run that controlled test yet.

Most of the work is not technical. Students with depth in the darshan write the criterion; people rate AI answers and write down their reasons (and some people new to the darshan are needed too, as a check); senior students review everything against the paribhasha; others welcome people, run workshops, write, translate, make videos, or look after funds. The scarcest resource is not computing power. It is careful time from people who understand the darshan. See the nine samitis.

That has not been decided. The Finance samiti will decide the legal structure, along with raising and looking after funds and reporting openly, and we will share the decision when it is made. Whatever the form, the commitments stay the same: results published whatever they show, every claim checked before it goes out, and the darshan never misrepresented.