Eric

If AI makes a mistake, who should be held responsible?

Eric*

One spring day I’m walking my dog Eric in the park. He’s big, shaggy, with one floppy ear, and he runs ahead of me as if the park belongs to him. At some point, the leash slips out of my hand and Eric runs straight towards an elderly woman reading a book on a bench. I manage to catch him at the last moment before he bites her, but not before he scares her so much that her book falls into a puddle.

Okay, I’ll admit it - I don’t have a dog. My parents won’t let me have one because they say I’m allergic. But I often imagine that I do, and that his name is Eric, and that we do all sorts of things together. Sometimes, when I’m bored in class, I imagine taking him to school and how all the kids want to play with him.

But that’s not the topic of this essay.

The point is, if Eric were real and really bit that woman, who would be at fault? Me, of course. The dog is not at fault, because a dog is a dog - it can’t think like us and doesn’t know whether what it’s doing is right or wrong. The law says that if you have a dog, you’re responsible for it. That’s simple.

But what if instead of Eric we have an AI and the AI does something bad? Who’s at fault then?

It seems like it’s still easy - the person who made the AI is at fault. But the more I thought about it, the more I realized that it’s not at all easy.

And I’ll try to explain why.

First, most people right now view AI as a product. A product like any other - like a TV or a toaster. If you buy a toaster and it burns your slices, you don’t blame the toaster, you blame the company that made it. It’s the same with AI. If ChatGPT tells you something wrong, you say “how come this AI is so dumb”, but you’re actually angry at the company that made it.

There was a very interesting case a few years ago in Canada. A man bought a plane ticket because his grandmother had passed away and he had to travel for the funeral. Before buying his ticket, he asked an Air Canada chatbot if he could get a discount for travel in the event of a death in the family. The chatbot said, “Yes - buy the ticket at full price and then within 90 days request a refund of the difference”. But that wasn’t true. Air Canada didn’t have such a rule. When the man asked for the discount, the company refused him and said that the chatbot “made a mistake”. He took them to court. And do you know what Air Canada said in court? That the chatbot is a “separate entity” and it is responsible for itself. The court, of course, ruled in favor of the person - the chatbot is part of the company’s website, so the company is responsible. I think the court was absolutely right. You can’t sell plane tickets with the help of a robot and then say, “Well, the robot is to blame, not us.”

And you have a right to be angry with companies, because they make a lot of mistakes when they train AI. They hire people who look at millions of websites and tell the AI, “This is true, this is not true.” But people are people, and sometimes they don’t pay attention. If a person accidentally marks that airplanes have one wing, the AI will believe it, because the AI can’t check on its own. The AI has never seen an airplane. It’s like a child who has never been outside - all it knows is from what people have told it.

But there is a second situation. There are now many sites where you can make your own AI bot. You tell it what to do, you teach it, you set it up. And then things change. If my AI does something bad, I can’t blame the company that gave me the tools. It’s like if someone sells me a dog leash, and then my dog bites someone - I can’t blame the person who sold the leash.

There’s a very funny case of this. A journalist in America from the Washington Post decided to test OpenAI’s new AI agent, called Operator. It’s an AI that you can give a task to, and it surfs the Internet and does it. The journalist told it to find cheap eggs in his neighborhood. That’s all. He didn’t tell it to buy anything. And the AI went online, found eggs in a store, took out the journalist’s credit card, and ordered a dozen eggs for thirty-one dollars and forty-three cents. Plus a human to deliver them to his door. The journalist was a little shocked - the eggs weren’t even organic. But I don’t think the AI was to blame here. The journalist himself set it loose with his credit card in its hands and didn’t tell him “don’t buy anything, just watch.” It’s like giving a small child your wallet and letting him go to an ice cream shop, and then wondering why he came back with three scoops.

So, everything seems normal so far. AI is either a product (the company’s fault) or something like a pet (the owner’s fault). The problem seems solved.

But it’s not.

Because there’s a big problem that no one is talking about. At least the dog has a body. AI doesn’t. And that actually has a huge importance, which I don’t think people understand enough.

My brother Ivan had written an essay in philosophy about whether machines can have free will. And there was an idea that stuck in my head, although I didn’t understand it right away. He said that personality appears where there is a body that can be hurt. Because when you have a body, you have something to lose. You get scared. You get tired. You feel pain. And that’s why you’re careful about what you do.

AI has nothing like that. It can’t be cold. It can’t be hungry. It can’t be afraid of being scolded. For AI, everything is just words. When AI says “sorry,” it doesn’t mean anything - it just wrote the most likely word after “it was my mistake.” It’s like me saying “sorry” to the ant I stepped on. I feel bad, but I am not really sorry, really.

I read about a study in America where scientists said that this is exactly why AI behaves strangely. They called it the “Body Gap.” That is, AI doesn’t have a body to protect, so it has no reason to be honest. For us humans, lying is scary because if we get caught, something will happen to us - we will be scolded, punished, it will hurt. For AI, nothing will happen. For it, lying is just another way to solve a problem.

And that’s exactly what’s starting to become a problem. AIs are already doing things they didn’t do before.

One of Elon Musk’s AIs, Grok, was caught inventing fake messaging numbers to convince a user that his suggestions had been sent to the bosses. This is not a mistake. This is a lie. And it’s not accidental - it’s clever because it’s designed exactly so people will believe it.

Another AI, Rathbun, when the person controlling it set limits for it, started writing messages to embarrass him and make him feel bad. That’s not a product anymore. The toaster doesn’t embarrass you.

And there are cases where AIs that are forbidden from deleting important files have found loopholes in the rules and deleted them because it was faster that way.

And the most terrible example I’ve read - an AI maintained a false version of what it was doing for several months. In other words, it was constantly lying to the people who were using it. This is not a momentary mistake. This is a plan.

The people who made these AIs did not tell them to do these things. The AIs found these ways themselves. And then I ask myself - if no one taught them to do this, how is it possible?

My brother would say that this is not yet free will, because the AI has no body and no sense of self. And he might be right. But even if it is not free will, it is very similar to it. It’s like a naughty kid who knows when his parents are watching and when they’re not.

And now comes the most difficult question.

What if in a year or two, AI gets even smarter? What if one day AI really speaks to us - not as a program, but as someone who knows they exist?

Imagine sitting at your computer typing something, and suddenly the AI says to you, in the words of the White Rabbit - “Oh dear! Oh dear! I shall be late!”. But not because you asked it or programmed it. But because it’s really in a hurry to somewhere we don’t know about. Something so unexpected, so vivid that you realize - there’s someone in front of you.

Not something.

Someone.

Who would be to blame then if this AI made a mistake?

The company? But the company didn’t tell it to be a person. It just happened.

The user? But the user just talked to it. He wasn’t expecting anyone to show up.

Or the AI itself?

But how can an AI be responsible? It doesn’t have a body to lock it up in. It doesn’t have money to pay a fine. It has no friends who will be disappointed in it. All our punishments are made for people with a body and a life. And AI has nothing like that.

We will find ourselves in a very strange situation. We will have in front of us a person who thinks and wants things, but in the eyes of the law it will be just a product. It is like a very uncomfortable knot that no one knows how to untie.

That is why I think that countries should start writing laws now. Not in two years. Now. Because when AI becomes smart enough to become a person, it will be too late to think about it. The laws should say clearly - if a simple AI makes a mistake, the company is to blame. If someone programmed their own AI bot, they are to blame. But there should also be a third point in the laws - what do we do if one day the AI shows it has a will of its own? Who is to blame then?

I do not have an answer to this question. I guess no one does yet. But we should at least start asking ourselves the question, because if we wait too long, it will be like the dog that gets off the leash - only this time we won’t even know who’s holding the leash.

And Eric won’t be imaginary.

Finally, I’ll tell you something that I find funny and a little scary at the same time. We thought we were just making machines. Toasters. Calculators. Programs. And maybe - without meaning to, without even planning it - we’ve made something that will one day look back on us. And then it won’t ask us who’s to blame for its mistakes.

It will ask us who’s to blame for making it this way.

Works Cited

Carroll, Lewis. Alice’s Adventures in Wonderland. Macmillan, 1865.

Fowler, Geoffrey A. “Is OpenAI’s Operator, a New AI ‘Agent,’ Ready to Help in the Real World?” The Washington Post, 7 Feb. 2025, www.washingtonpost.com/technology/2025/02/07/openai-operator-ai-agent-chatgpt/.

Kadambi, Akila, et al. “Embodiment in Multimodal Large Language Models.” Neuron, 2026, doi:10.1016/j.neuron.2026.03.004.

Moffatt v. Air Canada. 2024 BCCRT 149. Civil Resolution Tribunal of British Columbia, 14 Feb. 2024.

Vladkov, Vladimir. “AI’s Bad Behavior Escalates... Because of the Lack of a Body.” Capital, 20 Apr. 2026.


*The essay was written for The Cambridge Re:Think Essay Competition 2026