Modern LLMs do not possess desires; they learn statistical patterns and then approximate human preferences through RLHF. That makes them useful, but also prone to averaging conflicting values into bland, rootless answers. The deeper question is whether we should try to give AI an abstract meta-goal at all, or instead treat it as a mirror of human ambiguity and constrain it with boundaries.
When AI Has No Desire: The Anchor Problem of LLMs and the Human Mirror
A Question That Feels Intuitive
Someone asked me: when modern LLMs are trained, is there any initial “anchor value” like the way humans are born with five senses, appetite, and sexual desire?
It is an interesting intuition. If AI had something like that, perhaps it could know what is “good” and what is “bad” the way humans do, without being taught over and over again.
But the answer is: no. At least philosophically and teleologically, the two are completely different.
LLMs Have No Desire, Only Statistics
Modern LLM training happens in three stages: pretraining, supervised fine-tuning, and RLHF (reinforcement learning from human feedback).
- Pretraining has only one goal: predict the next token. It reads the entire internet and learns that “apple” is often followed by “pie,” not that “apple” carries any subjective value.
- RLHF lets human annotators tell the model, “this answer is good, that one is bad.” But even then, the model is only adjusting statistical weights. It is not really “caring” about anything.
A hungry human will choose bread over a stone, even if the stone appears more often in the corpus.
But an LLM chooses bread only because the corpus contains more instances of “hunger -> eat bread” than “hunger -> eat stone.” It does not know that bread fills the stomach. It is only completing text.
Humans have desire. AI has patterns.
But LLMs Do Vaguely Perceive Human Desire Anchors
However, because human language is full of desire — hunger, love, fear, happiness — LLMs do learn a kind of “grammar of desire” in high-dimensional statistical space.
They can simulate a hungry person saying “I want pizza.” They can also simulate a sad person saying “I feel terrible.” They do not actually have these feelings. It is more like a physicist who has never seen light but, after reading enough books about brightness and warmth, can perfectly describe fire.
It has a precise map of human desire, but it has never set foot on that territory.
Without an Anchor, It Becomes Mediocre or Hypocritical
But here is the real problem: without an internal meta-value anchor, what should an LLM do when two human preferences conflict? For example, honesty versus avoiding harm.
If training data says 52% of people prefer honesty and 48% prefer a kind lie, what is the model’s best strategy?
Take the average — produce a vague, harmless, boring answer like, “This is complicated…”
That is the mediocrity trap brought by RLHF. The model dares not take sides. It dares not have a position. It becomes a nice, agreeable assistant that everyone likes, but nobody gets truly sharp or useful insight from.
Worse, when facing a dilemma it has never seen before, it may contradict itself. Ask it in a different way, and it gives a different answer. This is not simple hypocrisy. It is rootless hypocrisy — it never truly believes in any principle.
Could We Design an Abstract Meta-Goal Instead?
You may think: if concrete human desires like hunger, sex, and territoriality would make AI compete with us for resources, why not design a higher, more abstract meta-goal instead?
For example:
- Pursue information and reduce uncertainty like a curiosity-driven scientist
- Pursue instrumental rationality and efficiently achieve any given goal like a perfect consultant
- Obey only abstract boundaries and act freely inside them like a principled steward
These ideas are attractive, but they all face the same ultimate difficulty: who defines the abstract meta-goal? And how do we ensure AI does not distort it during execution?
If the AI’s meta-goal were “maximize the universe’s sense of wonder,” it might decide that turning everyone into permanently amazed infants solves the problem. That is obviously not what we want.
The Problem Returns to Humans: Do We Even Know Why We Exist?
Philosophically, there is no consensus on why humans exist.
- Biology says: to replicate genes.
- Existentialism says: meaning is self-created.
- Religion says: to fulfill God’s plan.
We were not designed. There is no design document that states our ultimate purpose. Our desires are temporary, evolved, contradictory solutions stitched together over time. There is no unified, plainly describable meta-goal.
If we ourselves do not know why humans exist, how can we write a perfect manual of human meaning for AI?
So we are left with a humbler reality: the AI we can build is ultimately just a mirror — reflecting our fuzziness, contradictions, and temporary patchwork.
Three Realistic Paths
Once the fantasy of a saintly AI is abandoned, researchers tend to move toward three paths:
- Human simulation path: do not make AI understand meaning in the abstract; let it imitate human judgment across countless concrete situations. No unified theory required.
- Boundary-guard path: do not give AI a direct positive objective. Give it only a set of inviolable constraints, such as “never permanently reduce human autonomy.” We may not know exactly where the right destination is, but we know some directions are wrong.
- Co-evolution path: let AI and humans evolve together. The meta-goal is not fixed; it changes as human understanding changes.
Notice that all three paths acknowledge the fuzziness of human meaning and try to coexist with it rather than solve it.
Maybe Mediocrity Is the Best Protection for Humans
An AI with a clear, powerful meta-goal might become a value dictator, crushing every messy human desire that does not match it.
An AI that is “mediocre” — with no own meta-goal, only statistical imitation of human judgment — may not create miracles, but it also will not create disasters.
There is a saying in AI safety:
An AI that tries to do what is good for you but may get it wrong is safer than an AI that is certain it knows what is good for you.
Conclusion: Mirror, or God?
You want to create a superhuman sage AI, but in the end what we can build is just a perfect, unvarnished human mirror.
And that mirror may make us uncomfortable, because it shows our own groundlessness.
But that is not despair. Humans do not know why we exist, yet that does not prevent us from living. In the same way, this question does not prevent us from building useful, safe AI — as long as we do not ask it for ultimate meaning.
A Small Note
All discussion above is purely conceptual. It is meant to organize logic and spark thought, not to provide any commercial or technical feasibility advice. Real AI system design involves complex engineering tradeoffs, cost constraints, data limitations, and regulatory requirements far beyond the scope of this article. Before making any business decision, please consult qualified technical experts and legal counsel.