Mohammad Yasir Shah
Watch a child pick up a stone she has never seen before. She turns it over, taps it against another stone, and puts it in her mouth despite every warning not to. Nobody taught her to do this. Curiosity, in moments like these, looks less like a skill someone imparts and more like a basic condition of being alive. That raises an uncomfortable question for anyone building artificial intelligence today. If curiosity can appear in a child without instruction, could it ever appear in a machine the same way, without an engineer writing the code that tells it to explore?
It is not an idle question. As AI systems take on more open-ended tasks, whether they merely simulate curiosity or possess some version of it could shape how much we trust their judgment.
Answering honestly means separating two claims that get tangled together.
The curiosity engineers build
Most of what resembles curiosity in today’s AI systems is not spontaneous at all. It is engineered, deliberately and carefully. Researchers building reinforcement learning agents often add what is known as an intrinsic reward: a small internal bonus the system earns simply for encountering something new, or for being wrong about what it expected to happen next. Methods such as the Intrinsic Curiosity Module and Random Network Distillation exist for exactly this purpose, though they go about it differently. One rewards the agent for encountering situations its internal model predicts poorly, the other for encountering states a fixed reference network finds unfamiliar. The goal is the same: give the agent a reason to wander into unexplored territory even when no external prize is waiting there.
None of this happens because the machine wonders about anything. It happens because a human decided that novelty should count as a reward. Strip away that reward function and the exploring stops. What looks like curiosity is, at bottom, a carefully designed incentive structure, no different in principle from any other objective an engineer chooses to optimize.
When behavior shows up uninvited
There is a separate and more interesting phenomenon that has caught researchers off guard in recent years. Large models trained on broad, general objectives sometimes develop abilities nobody explicitly built into
them. A system trained only to predict the next word in a sentence has, in various documented cases, gone on to perform translation or basic arithmetic despite never being taught those skills directly. Nobody coded a grammar for French or a method for carrying digits. The capability turned up once the model and its training data grew large enough. Scientists call this emergent behavior, and it demonstrates something genuinely important: complex, unscripted abilities can arise from simple training goals once a system reaches sufficient scale.
It is reasonable to ask whether something resembling curiosity could emerge the same way, as a useful side effect of a broad training objective rather than a hand-coded instruction. In principle, it could. A system trained to minimize its own uncertainty about the world might start seeking out unfamiliar situations simply because that is an efficient way to reduce error, without anyone ever writing the word explore into its code.
But this observation answers only half the original question.
Doing versus feeling
There is a meaningful difference between a system that behaves as though it is curious and a system that actually experiences curiosity. The first is a matter of observable action. The machine seeks new information, tests unfamiliar situations, or asks questions it was never directly trained to ask. The second is a matter of inner experience, an actual felt pull toward the unknown, the sort of thing a person means when they say they are dying to find out how a story ends.
Behavioral curiosity is something engineers can build, and something that can also emerge unplanned inside a sufficiently large system. Experiential curiosity is an entirely different problem, tangled up with consciousness and subjective experience. Nobody has a reliable way to test for it, let alone build it. A machine that passes every behavioral test for curiosity tells us nothing about whether anything is actually happening inside.
This is not a philosophical technicality. It shapes what we are entitled to say when a system does something impressive and unplanned. Engineers can honestly report that a model explored, adapted, or asked an unexpected question. What they cannot honestly report, at least not yet, is that the model wanted to.
So what is the real answer?
Put simply, an AI can begin acting curious without ever being told to explore. That much is no longer in dispute. Emergent behavior has already shown that complex activity can appear from broad objectives rather than instructions written line by line. What remains unresolved is whether that AI ever wants anything at all. Every action a machine takes, however novel or surprising it looks from the outside, is still ultimately in service of the objective it was trained on, even when that objective is broad enough to make the resulting behavior look spontaneous.
The next real breakthrough will not come from an AI that explores more convincingly than the last one. It will come, if it ever comes, from a system capable of setting its own goals instead of optimizing a goal someone else supplied. That is a far harder problem than anything solved so far, and by most credible accounts, remains entirely open.
Until then, what looks like curiosity in a machine is best understood as an extraordinarily convincing performance of it, built by people who were curious enough to design it that way.
(The author is an AI researcher and inventor with interests in emerging technologies, robotics, and scientific innovation. He can be reached at developeryasir39@gmail.com)