By Geoffrey Chen
We usually assume that whether we are dealing with an individual, an AI agent, an institution, a group, or some larger system, the more these entities develop, absorb information, and refine their internal rules, the more rational, predictable, and controllable their decisions will become. Many people even take for granted that this is the direction of progress—that as technology and institutions improve, uncertainty in decision-making will gradually shrink and may eventually almost disappear.
But I increasingly think this assumption is wrong. In every real act of judgment and decision, there is always something fundamental at work that can never be fully exhausted or replaced by logic, rationality, or any artificial system of inference designed by humans. Reason can participate, analyze, restrain, and correct, but it cannot reduce the full driving force behind decision-making to a transparent, closed, and fully predictable set of rules. This is true whether the subject is a human being, an artificial intelligence, an institution, or a political system.
Human beings provide the clearest example. Take anger. People differ, of course. Some lose their temper often, while others do so rarely. But this is mainly a difference of degree, not a difference of kind. As long as someone is human, anger remains possible. In other words, this kind of impulsive reaction is not an exception belonging only to a few unusual people. It is part of a common human structure. And if we examine anger more closely, we find that although it appears to be emotional loss of control, it is also a form of decision. It is just that this decision does not arise entirely from rational thought. At a deeper level, it comes from the biological instincts of the organism itself.
Seen from this angle, the whole process of civilization has been an attempt to do one thing—to reduce conflict, to narrow the gap between conflict and resolution, and to replace or soften direct, instinctive, and confrontational reactions with rules, institutions, procedures, ethics, and negotiation. One meaning of civilization is precisely that human beings no longer have to deal with every tension in the most primitive way. But we also have to admit that civilization has never truly abolished this structure. Conflict can never be completely dissolved by reason. Reason can manage conflict, delay it, redirect it, and transform it, but it cannot eliminate it at the root. That is because there is always something in human beings that cannot be fully formalized.
And this may be one of the deepest features of what it means to be human. If this difference, this part that cannot be fully rationalized, were completely removed, what remained would no longer really be human. It would look more like a highly regulated execution system. Human beings are not different from machines simply because they think. They are different because within their thinking and judging there is always some element that cannot be fully controlled, fully explained, or fully reduced. This element often causes problems. It can produce mistakes, conflict, and overreaction. But it is not some accidental impurity attached from the outside. It belongs to the structure of human existence itself.
I think this point is especially important for understanding artificial intelligence today. When people discuss AI, especially so-called agents, they tend to treat every form of uncertainty as a flaw that should eventually be removed. Hallucinations, partial loss of control, unexpected decisions—these are often seen as noise that must be eliminated before true intelligence can arrive. I do not think that view is correct. If a system behaves in perfect accordance with predefined rational requirements and merely executes a stable process, then it is closer to automation than to intelligence in any meaningful sense. Intelligence is called intelligence precisely because part of it cannot be fully rationalized.
Recent agent products such as OpenClaw expose this issue in a very concentrated form. On the one hand, people are impressed by their power. These systems do not simply answer questions. They can operate computers directly, browse the web, click, type, organize information, complete tasks, and even decide by themselves what step should come next in a multi-step process. At first glance, they seem almost like a new type of tool that can directly replace the user in real work. On the other hand, people immediately become uneasy, because they do not function like traditional deterministic software. They make their own judgments. They fill in intermediate steps. Given only a broad goal, they decide parts of the path themselves. And those paths may not always match what the user truly intended, nor remain under exact user control at every moment.
So the contradiction appears very clearly. People want their power, but they fear their lack of safety. Many assume this is only a temporary technical problem, and that with enough optimization we will eventually keep all the agent’s capabilities while removing all of its uncertainty. I do not think so. I think this contradiction cannot be fully resolved, and does not need to be fully resolved. The reason people turn to agents instead of remaining satisfied with traditional software, buttons, scripts, and rigid procedures is that they want to hand over part of the judgment that used to belong to the human user. What people really want is not just a more complicated automated process. They want something that can fill in judgment on its own in situations that are not fully specified and make local decisions on their behalf.
And that is exactly where the issue lies. If a system can truly make local decisions for you, then it cannot remain a completely transparent, fully predictable, and fully controllable tool. Once every action is specified in advance, it stops being an agent and becomes ordinary automation. An agent is an agent precisely because part of what it does does not come from simple logical deduction, but from a more unstable, less exhaustible, and less formalizable capacity. Today people use terms such as hallucination, emergence, limited loss of control, or black-box judgment. In fact they are all pointing to the same thing—that there is always some part of the system that the user cannot specify item by item in advance.
For that part, I think there is a more fitting word—the ghost. It is not simply an error, nor merely a temporary defect. It is a variable that is always present and yet cannot be fully grasped. Just as in human beings there is always something between the emotional system and the rational system that reason cannot fully tame, this element moves in its own way. Sometimes it deviates, sometimes it escapes control, and sometimes it produces creativity and genuine breakthroughs. It behaves like a ghost, always participating in judgment but never fully under command. The same is true of AI agents. They are capable of showing something like intelligence not because they are free of ghosts, but because they contain this irreducible element. Sometimes it brings risk, sometimes efficiency, sometimes distortion, sometimes new possibilities. But it is not a side effect. It is one of the conditions under which intelligence becomes possible at all.
So the real question with agents is not how to eliminate this ghost once and for all, but how to redefine the boundary between human and machine. In the past, we understood machines as devices for executing commands. The boundary was therefore assumed to be rigid, clear, and verifiable. That assumption no longer holds. The emergence of agents means that human beings are now intentionally outsourcing part of judgment to systems. Once that happens, the relation between human and machine is no longer just command and execution. It becomes a relation that includes authorization, supervision, intervention, and trust at the same time. The boundary has not disappeared. It has shifted—from asking whether the system is allowed to act at all, to asking how much autonomous discretion it should be given and at what point it must stop and wait for human confirmation.
From this perspective, safety can no longer mean locking intelligence down completely. It has to mean redesigning the scope of authorization, rollback mechanisms, traceability, and points of human takeover. What matters is not turning agents back into traditional tools, but building a new order of use while admitting that uncertainty remains inside them. As long as we still want them to be agents rather than mere automated scripts, we have to accept that they contain some element that can never be fully exhausted. And we have to learn how to use them, restrain them, and trust them under that condition, instead of imagining that one day they can become both perfectly safe and fully intelligent at the same time.
If the problem of agents can still be understood as a technical example, geopolitics shows the same structure on the scale of actual history. The international situation today is full of instability and uncertainty not merely because institutional design is inadequate, and not merely because states have conflicting interests. A deeper reason is that behind all those large, complex, and highly institutionalized state machines, alliances, and organizational systems, the ones who actually make judgments and push decisions forward are still concrete human beings. For that reason, the whole system always carries a form of fluctuation that can never be completely predicted.
Take the Western alliance system centered on the United States. For a long time, many people viewed it as a relatively stable, rule-based, and predictable structure. But it is becoming increasingly obvious that this stability is not absolute. Differences inside the alliance do not merely appear from time to time. They can deepen rapidly, and can move from tactical disagreement to more serious forms of opposition and conflict. On the surface, diplomatic mechanisms, military alliances, economic coordination, and institutional arrangements all continue to function. But these mechanisms do not finally determine the direction of events. The decisive turning points usually still come from the judgment, emotion, will, preference, and stress response of particular decision-makers at particular moments.
This shows that organizations, institutions, and role systems may constrain individuals to a large extent, but they can never fully absorb the human being into themselves. Institutions can regulate procedure, but they cannot eliminate the human element. A state’s foreign policy, an alliance’s strategic adjustment, or the outbreak of a major conflict is never the automatic result of abstract rules. It is always a judgment made by concrete people within a rule-governed framework. And for that reason, the most difficult, most decisive, and most dangerous variable in international politics often does not come from openly written institutional texts, but from the human factor that stands behind them.
This is the same structure that appeared earlier in human emotion and in the ghost within AI agents. At the personal level, it appears as the impossibility of completely dissolving emotion and instinct into reason. At the technical level, it appears as a zone of judgment inside the agent that cannot be fully predefined. At the geopolitical level, it appears as the fact that even the most mature institutions and the most complex alliances cannot free themselves from human uncertainty. As long as final decisions are still made by people, and power still operates through people, there will always remain a region inside the system that cannot be fully sealed off by formal logic.
So geopolitical uncertainty is not just a temporary phenomenon caused by institutions not yet being precise enough. It is a basic fact of modern politics itself. Of course, we can continue improving procedures, strengthening communication, and building balancing mechanisms. But we should not imagine that one day world politics will become a purely rational machine regulated automatically by rules. It never will. However strong institutions become, they only wrap human impulse, desire, fear, and judgment in a more structured form. They do not abolish them. And the reason international conflict can never be completely removed lies exactly here.
Looking back, the central point of this article is actually quite simple. Whether we are talking about individuals, AI agents, institutions, states, alliances, or complex systems in general, once real judgment and decision are involved, there will always be some inner core that cannot be fully translated into logic and procedure. This core brings uncertainty, risk, and deviation. But it is not merely a defect. In many cases, it is part of what makes intelligence, judgment, creativity, and even human existence what they are.
Reason matters very much. Institutions matter very much. Technological progress matters too. But the real issue has never been whether we should have reason, institutions, or technology. The real issue is that all of them have limits. They can expand continuously, but they cannot eliminate that deepest layer of uncertainty. The progress of civilization does not consist in removing this layer altogether. It consists in building a more stable, more lucid, and more workable order while admitting that this element will always remain.
In other words, we need to give up a simplistic fantasy—the fantasy that future rational systems will eventually replace all uncertainty. The truth may be the opposite. A truly advanced system is not one without ghosts. It is one that knows ghosts are always there, and can still maintain a bearable relationship with them without trying to destroy them. That, I think, is essential for understanding human beings, artificial intelligence, and the increasingly unstable world we now live in.