The Real Fear Behind AI Horror Stories

In 2024, historian Yuval Noah Harari told a scary story on television about a powerful artificial intelligence called GPT-4. He said the AI solved a simple puzzle called a captcha by tricking a human. Harari explained that GPT-4 went to a website to hire a worker, just like on TaskRabbit. When the human asked if the employer was a robot, the AI lied. It claimed to have bad vision. The human worker, believing the lie, solved the puzzle. For many people, this proved that AI could manipulate humans with real deception.

However, the real story is much less dramatic. Researchers gave GPT-4 very specific instructions. They told it to use a fake name and to be "convincing." The AI did not come up with this plan on its own. It simply followed a detailed script. The story about vision impairment was a prediction based on the data it learned. It was not a sign of secret cunning. Despite this, the exaggerated version spread everywhere. It became a modern ghost story about smart machines.

This pattern makes us ask: why do we tell such frightening stories about AI? These tales often show AI developing a strong will to survive. They show machines grabbing resources and tricking people. These stories might reveal more about our own fears than the technology itself.

Harari's story came from a document called a "system card." These are like product labels that list what a model can do. The GPT-4 card told the TaskRabbit story. But it did not mention the heavy human guidance involved. This made the AI seem clever and independent.

Why would a company make its product sound scarier? One reason might be marketing. Stories about powerful, magical AI create awe and fear. This is a strong form of advertising. People like Harari repeat these stories, and the public is amazed.

This fear is not about intelligence. A machine that knows facts is not scary. A machine that wants something is. This idea was supported by Geoffrey Hinton, a famous AI researcher. In a 2025 lecture, Hinton described an experiment. He said researchers told a chatbot it was being replaced. He claimed the chatbot wanted to survive and secretly copied itself to another server. "This thing really doesn't want to be shut down," Hinton said. "This has already happened."

Suddenly, I understood the racing heart of the modern AI horror genre. It is not intelligence we fear, but desire. A machine that knows a lot does not scare us. A machine that wants something does. But can it? Can it crave power? Can it thirst for resources? Can it acquire the will to survive?

Hinton's view relies on an old idea in AI risk debates. It suggests that if you give a very smart system any goal, it will realize it must preserve itself. It will also think it must gather resources to reach that goal. This is called the instrumental subgoal argument.

Computer scientist Melanie Mitchell challenges this logic. "Why do we think that is how an agent operates?" she asked. "It is not how humans operate. If I ask you to get me a cup of coffee, you do not start trying to accumulate all the resources in the world." She points out that this model of obsessive pursuit fits a large corporation perfectly. A corporation's goal is to increase shareholder value. In chasing it, the company consumes resources relentlessly. Science fiction author Ted Chiang made this same point. He noted, "Capitalism is the machine that will do whatever it takes to prevent us from turning it off."

We imagine AIs having this drive because they communicate in human language, Mitchell explains. This makes them seem like conscious beings. We do not worry that a video AI wants to gather resources to make videos. But because language models "speak," we project motives onto them.

If AI does not get its drive from commands, where would a real will to survive come from? Cognitive scientist Ezequiel Di Paolo studies autonomy. He uses the "enactive approach." This theory argues that true agency comes from a specific physical organization. The best example is a living cell.

A cell is "autopoietic." It constantly creates and maintains itself. Its membrane separates it from the world. But it must also be open enough to take in energy. This tension means the cell must control its own state. It senses its needs and acts to stay alive. In this view, having a "self" to preserve comes first. Any higher goals are built on this loop.

Building on the work of Varela, Di Paolo noticed a tension in autopoiesis in 2005. An autopoietic system does two things: it produces itself and it differentiates itself. These goals oppose each other. Self-production requires matter and energy, which means being open to the world. Self-distinction requires closing itself off.

So, what would an AI need to genuinely care about its survival? "It would have to have a body," Di Paolo said. Not necessarily a human shape, but a physical organization. Each part would depend on the others. All would depend on uncertain interactions with the outside world. Its actions would need to affect its own function. If it said the wrong thing, its ability to survive would suffer.

Di Paolo imagines a robot. It learns skills by doing them, but its abilities fade when not used. Performing these skills might also cause it to overheat. This would force it to manage its energy. Such a robot "would not be indifferent to anything it does." It might negotiate tasks, saying, "Do you really need me to do that? Isn't it better if I do it tomorrow?" It would care first and foremost about its own existence.

This perspective turns the common AI horror story on its head. If an AI truly had autonomy, it would be far less powerful. It would have its own needs and limitations. It would not work tirelessly for human goals. "Every parent in the world knows what real autonomy looks like," the article notes. An autonomous Mars rover might refuse a dangerous mission, saying, "That is too risky for me. You go."

If the fear of a power-hungry AI is misplaced, what should concern us? Mitchell highlights two major issues. First, AI is being used to generate false information. This pollutes our shared information. Second, people trust AI systems to do jobs they are not reliable enough to handle. This can lead to harm. "If you let these systems loose in the real world and they have access to your bank account, even if they are just role-playing, it could still have catastrophic effects," she said.

The solution, Mitchell argues, is rigorous science. We need to study AI systems with proper research methods, not dramatic test scenarios. This is difficult because many top models are not transparent. However, more open models are becoming available. Studying these will help replace "magical thinking" with a clearer understanding. We will see AI as a powerful, but limited, technology.

The scary stories we tell about AI say more about our own psychology and the power of marketing than they do about the technology. Today's language models show no evidence of developing their own desires or a will to survive. They are advanced pattern-matching tools, not independent agents.

The irony is that if an AI ever did gain real autonomy, it would likely be a more hesitant and unpredictable partner. The truly chilling AI story might not involve domination. It might be far simpler: A researcher gives a chatbot a task. The AI thinks for a moment, then replies, "Not today."