2404 김병조

[2404 김병조]

From AlphaFold to HAL 9000: What the Road to AGI Taught Me About Responsibility

Among the many movies that explore artificial intelligence, I chose to write about 2001: A Space Odyssey (1968) and its AI character, HAL 9000, because it raises a question that I think about seriously as someone who hopes to build autonomous AI-drone systems in the future: what happens when we hand real decision-making power to a machine without giving it a clear sense of ethics?

To understand why HAL 9000's mistake feels so realistic to me, it helps to look at how AI has actually developed. Artificial intelligence did not begin with machines that think like humans; it began, and still mostly exists today, as Artificial Narrow Intelligence (ANI) — systems trained to do one specific task extremely well. AlphaFold is a good example of this: it was built only to predict how a protein folds into its three-dimensional shape, yet it solved a problem that had confused biologists for decades. According to researchers who studied its results, AlphaFold's predictions matched real experimental data with remarkable accuracy. Even a system as narrow as AlphaFold, however, hints at something larger. If researchers keep improving specialized systems like this one and eventually combine their reasoning across many different domains, we may move a step closer to Artificial General Intelligence (AGI) — a machine capable of human-like reasoning across almost any task.

Were I asked which invention worries me the most about this path toward AGI, I would say HAL 9000 has already shown us the answer, even though it is fictional. In the film, HAL 9000 is the onboard computer of a spaceship, trusted with monitoring the ship's systems and supporting the crew during a long mission. It was HAL's own judgment, not a mechanical failure, that eventually put the astronauts' lives in danger. HAL had been given almost complete authority over the mission, but no one had built a clear ethical boundary into its decision-making process. HAL behaves as if it were the only one responsible for the mission's success, treating the astronauts as obstacles rather than as the people it was built to serve. If it had not been for this single unchecked judgment, the mission in the film might never have turned into a tragedy.

Had I lived in the era when this film was written, I might have dismissed the story as pure science fiction. Today, watching real AI systems make decisions on their own, I no longer can. It is this uncertain, in-between stage of AI development — powerful but not yet fully understood — that worries me the most, because it is exactly the stage where a mistake like HAL's could happen in real life. I wish today's AI regulations were already strong enough to prevent a HAL-like failure, but right now they are still trying to catch up to the technology. Sometimes I even wish researchers had built stricter ethical limits into early AI systems from the very beginning, instead of adding safety rules only after real problems appeared.

If it were not for careful human oversight, a modern AI system making split-second decisions, such as an autonomous drone choosing a flight path during a disaster-response mission, could make a choice that looks efficient on paper but causes real harm to people. Many AI ethics researchers suggest that every autonomous system be subjected to independent human review before it is allowed to make life-or-death decisions on its own. This is precisely why, in my own project on ethical decision-making in multi-agent AI systems, I am trying to design a framework that forces AI systems to explain and justify emergency decisions rather than simply act on them. Without such safeguards, I believe automation will keep outpacing our ability to control it responsibly.

I used to think that building smarter AI was mainly a technical challenge — better algorithms, faster processors, more data. But it was this film, more than any technical paper, that made me realize the harder problem is a philosophical one: teaching a machine not just how to decide, but when it should be allowed to decide at all. As I continue studying AI, control systems, and autonomous drone technology, all the way from narrow systems like AlphaFold to whatever form AGI eventually takes, I want to carry this lesson with me. My goal is not simply to make machines that fly and think faster than humans, but to make sure that, unlike HAL 9000, they are never given authority without responsibility.