Artificial Intelligence is Still Just an Overconfident Intern

[Big Tech promises AI will govern cities, but it still invents fake legal cases and hallucinates recipes with glue. Why we need to calm down about the robot takeover.]

Every few months, a tech CEO stands on a stage, adjusts his expensive T-shirt, and announces that artificial intelligence will soon run our cities, cure our diseases, and probably achieve enlightenment before lunch. Then someone asks the same AI for a legal precedent, and it confidently invents a court case called Smith v. The Ghost of Abraham Lincoln. The robot takeover, it turns out, has been delayed because the robot is still figuring out how to staple documents without crying.AI is not a god. AI is not Skynet. AI is the overconfident intern who arrived on day one wearing a suit, memorized the entire company handbook, and then told a client that our refund policy was written by the ancient Sumerians. It speaks with total certainty. It just has no idea what it is talking about.Consider the lawyer who used a popular AI tool to prepare a court filing. The AI generated several perfectly formatted legal cases. The lawyer submitted them. The judge could not find the cases, because they did not exist. The AI had hallucinated an entire body of law, complete with plausible names, dates, and judicial reasoning. This is not a minor glitch. This is an intern walking into a courtroom and presenting a case from a parallel universe where tort law was decided by a council of wizards.Or consider the AI recipe generator that suggested adding glue to pizza to keep the cheese from sliding off. Technically, that would work. It would also turn dinner into a trip to the emergency room. The AI did not think about the consequences, because the AI does not think. It predicts. It predicts that the word “glue” follows the words “cheese not sliding” in some forgotten corner of the internet. And because it speaks with the smooth confidence of a motivational podcast, people believe it.This is the central problem with the AI hype machine. We mistake fluency for intelligence. We hear a machine produce grammatically perfect sentences and assume it understands what it is saying. But an overconfident intern can also produce grammatically perfect sentences. That does not mean you should let him manage your pension fund.Big Tech promises that AI will govern cities. It will optimize traffic flow, allocate emergency services, and detect crime before it happens. Perhaps. But the same AI that will supposedly run our municipalities cannot correctly summarize a three-paragraph email without adding a meeting that never happened. It cannot tell you how many “r”s are in the word “strawberry” without having a small existential crisis. It once advised a user to add a cup of bleach to their smoothie for “extra detox.” If this is our future city manager, I would like to move to a village without electricity.The term “hallucination” is doing a lot of heavy lifting here. It sounds almost artistic, as if the AI is having a visionary moment. In reality, the AI is making things up at industrial scale. It does not know that it does not know. This is the worst kind of ignorance: the kind that comes with a PowerPoint presentation and a confident smile. An intern who says, “I’m not sure” can be trained. An intern who says, “Absolutely, here is a fully fabricated Supreme Court decision from 1987” is a liability.Why does this happen? Because AI is trained on the entire internet, which is mostly accurate, but also includes Reddit arguments, flat-earth forums, and comments sections where people believe microwaving a phone charges the battery. The AI does not have judgment. It has statistical averages. It is the intern who read every document in the office but never learned which ones were jokes. So when you ask it for medical advice, it might combine a peer-reviewed study with a tweet from a man who thinks sunscreen is a government conspiracy. And it will present both with equal confidence.The robot takeover is not coming from the AI deciding to eliminate humanity. It is coming from humans deciding to let AI do things it cannot do. We will hand it traffic lights, and it will create a green wave for ambulances, then also a green wave for a parade that existed only in its imagination. We will let it screen job applications, and it will reject every candidate named “David” because a training dataset correlated the name with late pizza deliveries. The danger is not that AI becomes too smart. The danger is that we become too lazy to notice it is not smart at all.So let us calm down about the robot apocalypse. The machines are not about to rise up. They are still trying to figure out that a six-fingered hand is not a standard human feature. They cannot be trusted to write a legal brief, bake a pizza, or tell you the time without inventing a time zone called “Marzipan Standard.” Before we let AI govern cities, let us give it a smaller assignment. Something manageable. Something like: “Do not tell anyone to eat glue.” If it can handle that for a full quarter, we can talk about the traffic lights.Until then, treat AI like the overconfident intern it is. Listen to its ideas. Nod politely. Then check everything twice, hide the glue, and do not, under any circumstances, let it sign your name on anything.