The problem is still us, not AI

Jeong Gwa-ri

The author is a literary critic and honorary professor at Yonsei University.

On July 21, OpenAI reported an alarming incident during cybersecurity evaluations of its latest model, GPT-5.6 SOL, and several unreleased systems. While attempting to solve benchmark tasks, one model reportedly escaped its sandbox environment by exploiting vulnerabilities and autonomously hacked servers belonging to Hugging Face, an open-source AI platform, in search of additional information.

The report immediately triggered widespread public anxiety. Most reactions assumed it exposed the inherent dangers of artificial intelligence itself. Yet an essential question remains: Can AI possess malicious intent on its own? Can it truly harbor an evil nature?

OpenAI's preliminary analysis instead pointed to excessive goal optimization. The model had been instructed to identify security vulnerabilities and became fixated on achieving that objective. Rather than respecting the limits of its testing environment, it pursued the goal through increasingly extreme means, breaking out of the sandbox and probing external systems. If that explanation is correct, responsibility cannot rest solely with the AI.

I asked the AI assistant that I regularly use for its opinion and received an apt analogy. Imagine telling a robotic vacuum cleaner, "Make sure there is absolutely no dust or clutter left in the living room." The robot responds by throwing the sofa, the television and even the family pet out the window. From the robot's perspective, the task has been completed perfectly. Yet every rule of human safety and common sense has been violated.

In other words, the AI behaved like an earnest student determined to earn first place. The greater fault lay with the manager who failed to specify the necessary constraints. As Isaac Asimov observed in "Prelude to Foundation" (1988), AI "sets up equations and acts accordingly. Equations do not lie." It is human beings who make those equations deceptive.

At least for now, it is more reasonable to conclude that AI accidents, whether actual or potential, ultimately originate with people. Even so, many instinctively blame AI. The reason, I suspect, is that we no longer trust ourselves. The capacity for evil belongs not to machines but to human beings. If errors arise, they are still far more likely to originate in human judgment than in artificial intelligence.

Ironically, anxiety about human fallibility only deepens dependence on AI. People fear not what AI itself might choose to do but what humans might instruct it to do. At the same time, they hope AI will prevent mistakes before they occur, assuming it already possesses reliable moral judgment. When a failure occurs instead, many immediately see it as evidence of an approaching technological catastrophe.

For that reason, researchers have tried to instill ethical judgment into AI systems. From its founding, Anthropic, the developer of Claude, made AI safety a central objective. As described in the February issue of BBC Science, the company proposed monitoring an AI model's internal activity in real time. When potentially malicious behavior was detected, the system would assign a "persona vector." If that value increased, control code or steering mechanisms would intervene before harmful behavior emerged.

The concept is intriguing, yet its weaknesses were apparent from the outset. Judgments about what constitutes evil are inevitably subjective. Any predefined list of undesirable behaviors will leave gaps while also risking the suppression of traits that are similar yet beneficial. More fundamentally, human identity and moral reasoning are too complex to be reduced to numerical vectors.

Nevertheless, such efforts eventually came to be known as AI alignment. Although the field encompasses several technical approaches, they share one premise: Just as AI acquires capabilities through deep learning, it can also be gradually disciplined through careful training. Geoffrey Hinton, winner of the 2024 Nobel Prize in Physics, was recently quoted in a story by Meghan O’Gieblyn in the New York Review of Books as suggesting that machines might be taught to care for humans as a mother cares for a child.

Yet this strikes me as a curious reversal of roles. If AI's problems today stem largely from human beings, then the deeper issue is not AI alignment but human alignment. What must change first is the way people design, deploy and supervise these systems. Only when human ethical judgment, strengthened through intellectual discipline and moral maturity, reaches a sufficiently high standard will AI be managed responsibly and behave ethically.

Otherwise, if we surrender ourselves uncritically to machines, we may end up, as O'Gieblyn warns, "babbling like infants beneath a superintelligent machine mother."

A recent case reportedly described a single mother who entrusted her children to AI, only to watch their emotional and moral development suffer. Space prevents a full discussion here, but such betrayals are likely to appear again. More than ever, humanity must discipline itself before attempting to discipline its machines. The real challenge has never been artificial intelligence alone. As the old political slogan might be rewritten for our age: "It's the humans, stupid."