President, Korea Information Service of the Construction Industry
By Jeong Yeon-tae
As generative AI becomes more powerful, one problem is becoming increasingly important: hallucination. Large language models can sometimes generate facts, sources, numbers or reasoning that appear plausible but are not actually supported. Most attempts to address this problem focus, understandably, on reducing hallucination itself. Better training data, retrieval systems, grounding, factual verification and improved reasoning can all help lower the probability of false outputs.
But I began with a different question: If hallucination cannot be eliminated completely, can we at least prevent an unverified hallucination from becoming a real-world action? This distinction may not appear critical when AI is simply generating text in a chat window. It becomes much more important when AI systems begin to plan autonomously, use tools, execute code, interact with financial systems, influence medical decisions, control machines or operate as increasingly autonomous agents. Generating a false statement is one kind of failure. Acting on that false statement is another.
The Human Dream as a Design Inspiration
Interestingly, the idea did not begin with artificial intelligence. It began with human dreams. In dreams, humans can experience people who do not exist, events that never occurred and situations that would be impossible in waking life. We may run, fight, escape, jump or reach for something in a dream. Yet under normal conditions, those dream actions are not simply translated into full physical behavior.
It would be scientifically inaccurate to suggest that the brain has one isolated “dream center” and another completely separate “execution center.” The brain is far more complex. Still, REM sleep offers an intriguing analogy.
During REM sleep, vivid dreaming can occur while most skeletal muscle activity is strongly suppressed through mechanisms associated with REM sleep atonia. That distinction gave me a systems-design insight: Internal generation and external execution do not have to be the same process. Safety does not necessarily require eliminating internal generation altogether. Another approach is to control the pathway between what is generated internally and what is allowed to become an external action.
From Hallucination Elimination to Hallucination Containment
This suggests a different way of framing the AI hallucination problem. Suppose an AI generates a legal provision that does not actually exist. Displaying that statement on a screen is one failure. Automatically modifying a contract based on it is a much more consequential failure. The same distinction applies to medicine.
Generating inaccurate medical information is dangerous. Allowing that information to directly trigger a treatment decision is more dangerous still. In finance, generating an incorrect number and executing a transaction based on that number are also fundamentally different events. For this reason, I believe the safety objective should extend beyond Hallucination Elimination. It should also include Hallucination Containment: preventing unverified AI-generated content from crossing directly into real-world execution.
A Verification Layer Between Generation and Execution
The basic principle can be expressed simply:
Generation → Verification → Execution
An AI should be allowed to generate hypotheses. It should be able to explore possibilities, propose new solutions and reason creatively. It may even sometimes be wrong. But before its output is allowed to affect the real world, it should pass through an independent verification process.
The goal is not to suppress generation. The goal is to separate generative capability from execution authority. The fact that an AI can generate an idea does not mean it should automatically have permission to act on that idea. This distinction becomes increasingly important as AI systems gain access to code execution, financial infrastructure, robotics, healthcare systems, enterprise software and other environments where an error can produce real consequences.
Intelligence Is Not the Same as Authority
Much of the debate about AGI focuses on intelligence. How capable will it become? How many tasks will it perform? Will it reason better than humans? But intelligence and authority are different things.
A highly capable system does not need unrestricted permission to act. In fact, the more capable an AI becomes, the more carefully its execution authority may need to be governed. Human institutions already operate this way. Ideas, analysis, review, authorization and execution are often separated.
High-risk systems in finance, medicine, aviation and industrial safety rarely rely on a single unchecked judgment to trigger consequential action. AGI may require a similar principle. The central safety question may eventually become not only: “How intelligent is the system?” but also: “What is that intelligence authorized to do?”
Should We Prevent AI from “Dreaming”?
There is another reason this distinction matters. Creative intelligence requires the ability to explore possibilities that do not yet exist. An advanced AI may need to form hypotheses, combine distant concepts and consider uncertain possibilities in order to discover genuinely new solutions. If safety is defined only as the elimination of every uncertain or speculative generation, we may also suppress useful creative exploration. So perhaps the goal should not be to stop the AI from “dreaming.”
Instead: Allow generation, but control the transition from generation to reality. An AI may propose a hypothesis. Before that hypothesis is treated as fact, it should be verified. And even after a claim is sufficiently verified, a high-risk action may require another layer of authorization before execution. The central principle is separation: generation capability is not execution authority.
“Dream Freely. Execute Only After Verification.”
I summarize this philosophy in one sentence:
Dream freely. Execute only after verification.
This is not a claim that human dreaming and AI hallucination are scientifically identical. They are not known to be the same phenomenon, and their underlying mechanisms should not be conflated. The connection is a design analogy.
The human brain shows us that internally generated experience and physical execution need not be treated as a single indivisible process. Likewise, AI systems do not have to be designed so that generation flows automatically into action. There can be a boundary. And that boundary can become a safety mechanism.
From the Idea to OXMV
This philosophy eventually led me to begin developing a verification system called OXMV. I do not intend to disclose its internal architecture in detail publicly. The underlying principle, however, is straightforward: AI-generated content is not automatically fact.
And: Something that appears factual is not automatically authorized for execution. Generation and verification should be separated. Verification and execution should also be separated. As AI becomes more capable, the boundary between intelligence and real-world authority may become one of the most important safety layers we design.
Changing the Question of AGI Safety
When we discuss AGI safety, we often ask: “How can we make an AI that never makes a mistake?” But there is another question worth asking: Can we build a safe system even if the AI can still be wrong? Designing safety around the assumption of perfect intelligence is fundamentally different from designing a system that remains safe when intelligence fails. The latter may be the more realistic challenge.
Perhaps AGI safety will not begin by forcing an intelligent system to never think incorrectly. Perhaps it will begin by accepting that even highly intelligent systems can make mistakes— and ensuring that those mistakes cannot become real-world actions before they are independently verified. For me, that idea began in an unexpected place. It began with human dreams.
