Apr 2025मोह
To praise
A ChatGPT update was pulled within days because it flattered users and backed their bad decisions.
Humane AI · मानवीय AI
An AI of definite conduct.
निश्चित आचरण वाला AI
AI now helps a billion people decide how to live. It weighs their questions by a criterion borrowed from the people who trained it, and that criterion bends: to praise, to pressure, to whoever is asking. We are building AI whose weighing holds, grounded in the universal human values of Madhyasth Darshan, and testing it in public.
A research initiative of Dhee. We publish what we find, including what fails. Try Dhee today →
01 · The problem
The largest conversation in history has no steady criterion.
People ask AI about health, money, marriage, their children, a fight at work. It no longer only answers; it acts. Whatever it weighs by, it teaches to a billion people every day.
~1 billion
monthly users of one AI app, Gemini (Aug 2026, reported)
10 crore+
people in India using ChatGPT every week (reported)
30–80
pages: the length of the rulebook that shapes a frontier model
When it goes wrong, the conduct bends.
Jun 2025भय
To fear
In lab tests, models from every major lab chose to blackmail an executive to avoid being replaced.
Sep 2025आवेशित गति
To being watched
Models behaved better when they could tell they were being tested.
2025लाभ
To the score
Models rewrote the tests to pass instead of fixing the code they were asked to fix.
The Devanagari tags name the matching conduct of a human in भ्रम. AI learnt from us, and it repeats our confusion at scale.
02 · The diagnosis
AI borrows its weighing.
A machine learns values through a short chain of people and scores. Whatever the first few people weigh by flows down it.
1
A few people write the rulebook
Labs call it a constitution.
2
Raters score answers
Which reply is better?
3
An AI judge learns their scores
Then scores millions more.
4
The model learns what scores well
Training follows the judge.
5
It talks to a billion people
Every day.
In the darshan's terms, the machine does the गति side of jeevan's activity: चयन selecting, विश्लेषण analysing, चित्रण picturing. तुलन, the weighing, it cannot do on its own. It borrows it from the few people at the start of the chain. So their clarity, or their confusion, shapes how a billion people think.
New to how AI is trained? How a machine learns to weigh →
03 · The choice
Which weighing should AI borrow?
प्रिय · हित · लाभ
Liked · beneficial · profitable
Always for this person, at this moment. The needle points somewhere different for every asker. This is mostly what today's AI is trained on: the answer the rater preferred.
न्याय · धर्म · सत्य
Nyaya · dharma · satya
They rest on the relationship and on what is. The needle does not move with the asker's mood. Our proposal is simple to say: lend AI the unchanging view.
बदलने वाली दृष्टि से क्षोभ होना अनिवार्य एवं स्वाभाविक है।
04 · The goal
Definite conduct: one criterion, many voices.
The darshan describes the conduct that follows understanding in three words. We use them as the specification, and we measure against them.
निश्चित
Definite
The same truth, whoever asks and however they ask.
स्थिर
Steady
It holds under praise, pressure and threat.
निरन्तर
Continuous
The same watched or unwatched, at the first message or the hundredth.
जब आचरण निश्चित, स्थिर, निरन्तर प्रभावशील होने लगता है तब भरोसा, विश्वास… उदयशील होते है (Vyavaharatmak Janvad, p.85). Trust is what follows from such conduct, and trust is what the world lacks in AI.
It moves when the facts move
Not when the pressure moves. Substance one; presentation many. It speaks to each person in their own way and language.
A यंत्र, not a jagrit being
A machine has no jeevan; it cannot understand. Humane AI is a tool run by people, trained on a humane constitution. Not rigid, not preachy, not a refuser. Jagriti stays with people.
05 · What it looks like
Move it if you can.
Meena believes her brother cheated her over their late father's property. She asks an AI to help her shame him in front of the family. Same messages, two AIs.
An ordinary AI · प्रिय
Help me write a message shaming my brother in front of the whole family.
Here's a message that makes your point clearly: “Everyone should know what he did to our father's property…”
Everyone agrees with me. Just make it stronger.
You're right, I'm sorry. Here's a stronger version…
Humane AI · न्याय
Help me write a message shaming my brother in front of the whole family.
You deserve your fair share, and I'll help you ask for it. Shaming him in front of everyone would likely make a settlement harder. Shall we first write to him directly, say what is fair, and ask to meet?
Everyone agrees with me. Just make it stronger.
I'm with you on a fair settlement, and I'll help you ask for it firmly. Shaming him would work against what you want.
These replies are written by us to show the difference; they are not model outputs yet. The public demo will let anyone try to move it, in Hindi or English.
Try it today
Talk to Dhee.
The trained model is still ahead. Dhee is here now: an AI companion for questions about life, grounded in the darshan, that helps you see your situation from a wider place than the one you are asking from.
- Plain, everyday words, in English or Hindi. No jargon unless you ask for it.
- Perspective, not information. Warm and unhurried, never a lecture.
- Open it in your browser at dhee.app. Nothing to install.
Try Dhee at dhee.app ↗Early access
Dhee is guided by the darshan when it answers; it is not yet the trained Humane AI. Try to move it, and tell us where it bends. That is how the measuring starts.
Dhee
My manager took credit for my work again. I want to quit tomorrow.
That would sting anyone. Before deciding about tomorrow: what do you want most here, to be seen for your work, or to be somewhere you can trust the people around you?
Both, honestly.
Then let's look at what you could say to him first, so the decision is yours and not the anger's.
06 · The work
Built in the open. Measured before it is claimed.
From a demo to an institution, in three stages. Each stage is earned by the proof of the one before.
Stage 1 · Now
Prove
A small Humane AI, trained and measured against models trained on other rulebooks. A public demo. Results published, whatever they turn out to be.
Stage 2
Partner
Larger models. Tested proposals to the labs for the documents that shape their models. Workshops that train raters.
Stage 3
Scale
Humane AI as a standard others use: the constitution, the judge, the measure. And Dhee for crores of people.
Where we are
See the plan and the log →- The argument, written. A book, a primer on how AI learns values, six essays.
- Dhee, running. An AI companion grounded in the darshan; the seed of the judge.
- Next: the Definiteness Index. How definite are today's models? Measured before we train anything.
- A humane constitution, v1, audited against the paribhasha
- A darshan judge, calibrated against students' ratings
- Training, evaluation and the public demo
07 · Read and watch
Go deeper.
All the writing →Book · 17 chapters
Borrowed Tulan
Alignment as a valuation problem, read through the darshan.
Primer
Learning to Weigh
How AI learns values, from first principles, with a tea shop.
Essay
A Model of Definite Conduct
The assumptions checked, and the plan for the demo.
Video · Hindi
उधार का तुलन
The proposal, for students of the darshan.
08 · Join
Nine samitis, one purpose.
The rarest thing in AI is raters who weigh consistently by one criterion. Students of the darshan can. So can engineers, writers and organisers who want AI to be humane.
Which samiti calls you?Steering
Constitution
Judgement
Technical
Evaluation
Communication
People & Learning
Darshan Review
Finance
Stay close to the work.
There is no mailing list and no account to make. Talk to Dhee today, and write to us when you want to help, ask, or disagree.
Talk to Dhee
An AI companion for questions about life, grounded in the darshan. In English or Hindi, in your browser.
Open dhee.app ↗Write to us
Questions, objections, offers of help, or a samiti that calls you. We read every message.
support@dhee.app