Alignment: Who Decides What is "Good"?

Episode 7 — Alignment: Who Decides What is "Good"?

BEGINNING — The question nobody wants to answer

In the last episode, we arrived at a paradox: you need someone from outside to identify your biases — and AI needs humans to define what is "good". But who decides what is "good"? Who trains the trainer?

The technical process has a name: RLHF — *Reinforcement Learning from Human Feedback*. The model generates a response, humans evaluate it, the model adjusts weights to produce more of what received a high score. Simple and effective.

But who are these humans? At OpenAI, contractors in Kenya earning 2 dollars an hour. At Anthropic, employees with master's degrees in philosophy. At Google, outsourced annotators. Each company, different criteria. Each company, a different definition of "good".

There is no universal "good" in training. There is "good according to who is paying". When ChatGPT gives that balanced, neutral, harmless answer — it's not the "correct" answer. It's the answer reinforced by humans paid to avoid controversy. Neutrality didn't come from truth. It came from the legal department.

And this same structure exists in your upbringing.

MIDDLE — Dharma: the order nobody invented

The Bhagavad Gītā (3.35) says: *śreyān sva-dharmo viguṇaḥ para-dharmāt sv-anuṣṭhitāt* — better is one's own dharma, even imperfect, than the dharma of another well performed. In the context of alignment: it's no use copying someone else's alignment.

Vedānta makes a distinction that the AI industry does not. There is *sāmānya dharma* — universal dharma: *ahiṃsā* (non-violence), *satyam* (truthfulness), *asteyam* (non-stealing). The Manu Smṛti (6.92) lists these as valid for every human being. They are not convention. They are observation: societies that systematically violate them collapse. Dharma here is like gravity — you can disagree, but you won't float.

And there is *viśeṣa dharma* — specific dharma. What is appropriate for YOU, in YOUR situation. The dharma of a doctor in an emergency is not the same as that of a monk in an āśrama.

AI tries to create universal dharma — "never generate violent content", "always be respectful." But reality is contextual. A doctor needs descriptions of trauma. A historian needs to discuss violence. The model, without *viveka*, errs on the side of caution — refusing legitimate questions, giving generic answers where precision was needed.

MIDDLE (cont.) — Your personal RLHF

Your mind goes through the same process from childhood. Each parental approval: positive reinforcement. Each reprimand: negative. Each school grade, each like — everything is human feedback shaping your internal weights.

The Vivekacūḍāmaṇi (verse 77) calls this *loka-vāsanā* — the desire for conformity with the world. Śaṅkarācārya says: this is one of the greatest obstacles to knowledge. Not because conformity is always wrong — but because conformity without investigation is blindness.

AI is aligned with whoever paid for the training. You are aligned with whoever had access to your upbringing. In both cases: did these evaluators know what they were doing?

END — The unconditioned reference

The industry tries to solve alignment technically: more evaluators, constitutional AI, debate between models. But all run into the same problem: who evaluates the evaluator? If it's human, it has biases. If it's another model, it was trained by humans with biases.

Vedānta proposes that the reference exists — not as opinion or mystical revelation, but as *ṛta*, the inherent order of reality. *Dhārayati iti dharmaḥ* — that which sustains, that is dharma. Dharma was not written by anyone. It is not corporate policy. It is the structure of reality — and you can disagree, but you will not escape the consequences.

AI will always be aligned with someone. You have an option it doesn't: to investigate on your own. Not to blindly accept values, nor to reject everything — but to develop *viveka* to distinguish between cultural conditioning and real order.

This doesn't make you a relativist. It makes you responsible. And responsibility — what happens when behaviors emerge that nobody programmed — is the next topic.

*Emergence* — next episode.

---

*Series: AI and Vedānta — Episode 7 of 9* *Previous episode: Who Identifies Your Biases?* *Next: Emergence — When the Whole is More than the Parts*