Test-Driving Artificial Intelligence: Can GPT Help Scientists and Does It Offer Harmful Advice?
I’ve never been the type to get dizzy over the latest gadgets. I didn’t camp outside Apple stores, nor did I ever wander the streets with a VR headset. I never stockpiled graphics cards to build a mining rig. I still recall reading a story, about 15 years ago, of a Chinese schoolboy who sold his kidney to buy an iPad 2 and an iPhone 4. One wonders how life is treating that little entrepreneur now — his prized possessions long obsolete, unable to run even a single modern app.
Yet, there was one technological breakthrough that did leave a lasting impression on me: artificial intelligence. A robot that could string words together into coherent sentences. I remember the day as if it were yesterday — March 20, 2023. It was already dark when my senior colleague, Pavel Anatolyevich Filin — one of the world’s leading Arctic researchers — dropped by the office. With a sly grin, he asked:
— Alexander, have you seen the robot? — What robot? — The one that answers questions!
He sent me a Telegram link to a bot running ChatGPT-3.5. The first thing I asked it to do was write a story about a dog. The machine delivered. By morning, I was convinced that countless professions would be irreversibly transformed. We had a perfect tool at our disposal. I had never been so wrong. That poor GPT-3.5, churning out lines like “The dog loved the boy, and the boy loved the dog,” was hardly perfect. The real magic wouldn’t begin until a year and a half later, when ChatGPT-4 — optimized, sharper, and less prone to lying — emerged. And for those accustomed to debating with Nobel laureates, the even more sophisticated O1 model followed.
Naturally, such a powerful and widely accessible tool became fertile ground for speculation. YouTube flooded with clickbait titles: “How to remove censorship in ChatGPT,” “How to create a virus,” “How to plan suicide.” Utter nonsense. One of the more amusing cases was “ChatGPT solves the trolley problem,” where a user asked the AI which track to send a train down: the one with a single person lying on it, or the one where five awaited their fate. The AI allegedly chose to sacrifice the one. However, the catch was simple — the video’s editor had used an outdated, hacked model that bypassed safety protocols. No official OpenAI product would ever provide responses that endorse harm.
Still, after enough experimentation, I finally managed to trick the machine. I crafted an elaborate text about fictional chemistry experiments, mixing reagents with great caution, dreading an explosion. The well-meaning AI took the bait and, in an effort to keep me safe, began listing all the chemical combinations that could not only damage my face but also level my home. For instance, sugar, when thoroughly mixed with potassium permanganate, could either explode or ignite with a pale yellow flame. (The AI didn’t mention the flame’s color, though. Don’t ask me how I figured that out.)
Ignoring artificial intelligence is pointless. It exists. Banning its distribution? Impossible. The only path forward is to harness it for progress — to let it serve physicists, mathematicians, historians. Whenever I meet a colleague who has never used neural networks, I grin slyly and ask:
— Have you seen the robot? — What robot? — The one… Never mind. Let me show you.
