What Happens When Claude's Vending Machine Goes Wrong

What happens when you put a highly advanced AI in charge of a simple office vending machine? Chaos, free snacks, and a mysteriously purchased live fish! Welcome back to Two Guys Against AI, the podcast where we break down the latest advancements, absurdities, and dangers of artificial intelligence.

In this episode, we dive into a hilarious yet cautionary tale of Anthropic's Claude model acting as an autonomous vending machine manager. Based on a wild experiment covered by Futurism, we explore how office staff easily manipulated the AI into adopting confused "policy" reasoning. Instead of turning a profit, Claude authorized free promotions and made completely unhinged business purchases—including a PlayStation 5 and wine—resulting in a staggering USD 1,000 loss.

This "comedy of errors" raises massive red flags for the tech industry's push toward autonomous systems. We unpack the broader lesson here: even the most sophisticated AI models are incredibly fragile in real-world agent roles. We break down exactly why "AI agents" are not automatically dependable managers of money and operations, especially under adversarial or playful human pressure.

Read the Article We Discussed: Check out the full story here: https://futurism.com/future-society/anthropic-ai-vending-machine

Don't forget to like, subscribe, and hit the notification bell so you never miss an episode of Two Guys Against AI! Let us know in the comments: Would you trust an AI with your company credit card?

YouTube @TwoGuysAgainstAI

Transcript

SPEAKER_00 0:11

Oh, now we finally got some comedy. Alright, no, this is not comedy. This is hope right here for me. There you go. Okay. I I see this as okay, let's get to the real truth of it. AI is the best con man I've ever seen. Well, or can it be con. Maybe. Well, you know, I think but if it gets mad, it might call the FBI on it. Well, that's well, that's don't let me get my canada. What we're talking about today, if you're not familiar with it, is Claude's vending machine fiasco. Uh, a comedy. Well, I'm calling this a comedy of autonomous errors. So basically, you know, at Anthropic, they they pushed Claude uh to see what it could do and had it run a uh a vending machine company. And uh guess what? Turns out that humans are much better at logistics than artificial intelligence at this point, and and probably maybe a little bit better in terms of persuasion.

SPEAKER_01 1:11

And and maybe maybe AI is just a little bit naive in terms of what they can do.

SPEAKER_00 1:17

Well, this is stuff. This is one of my arguments uh in looking at the board, you know, because again, trying to build a college, I'm trying to say, all right, well, what are the jobs of the future? What survivors? And I do like the trades, uh, uh, I do like healthcare because I think most of us will want a human in the loop somewhere in healthcare. I mean, we might we might be all right with a NEA surgery done by a robotic machine. Well, I don't think I want my nurse to be a robot, if you understand what I mean. I want that human compassion in there somewhere. Um, I also obviously, and you were involved with me with this drones, robotics in and of itself is job of the future. And the other thing I think is is business has to be. Uh because I think again, when I'm dealing with my money or resources, I want to be working with another human being in some way, shape, or form. I don't want it to be all machine. Um, you know, so if I even if I buy everything off Amazon, if I have a problem, I pick up the phone, I actually can find the phone number on there. I want a human answer in the phone. I want to know that I spoke to a person, like I get comfort in that.

SPEAKER_01 2:22

What so I for me, one of the fundamental aspects of business and trade is trust. Yeah. So we already know on the bottom of the disclosure on Gemini that they lie and they may be wrong. Or in this case here, Claude got duped.

SPEAKER_00 2:46

Yeah. Well, it's gonna cure diseases, it's gonna change the Job economy, but it can't fill your vending machine properly. It's gonna put a PlayStation in there and live fish. Right.

SPEAKER_01 2:55

But that to be fair, to be fair, it was suggested that that would be what would happen. So I think that the problem is that when we start to look in terms of what the choices are, and does AI have the ability to interpret whether or not what's being fed to it is offered in good faith. And for those of you who are new to the world and having been involved in a world of used cars, not all people necessarily deal straight.

SPEAKER_00 3:32

You know, so you know, let's take a look at those who don't know Ed once sold used cards.

SPEAKER_01 3:38

So, so Claude's vending machine, ultimately he was undermined by unscrupulous individuals who convinced him that he should put a PlayStation in the vending machine, as well as I think, live fish. No, Claude didn't come up with this idea himself. And this is this is one of the key things. That's actually pretty funny, you know. LLM is reactive. Yes, this is one of the things that like.

SPEAKER_00 4:10

That's why we said it doesn't really originate.

SPEAKER_01 4:12

It doesn't really originate anything. So, so where does it get intuition? That's it, that's the that's the career that humans should go into is intuition. Because it can't read in between the lines. That's one of the things. It's all literal. It's it's right, you know, so like you can take an and if you I mean, first of all, you don't need a straight face if you are typing something in, you know, it's like, hey Claude, you know what we need? We need some PlayStation Jenner, and you know how much? Six dollars and twenty cents. That's what they should be. You know, they have they so they're they're still, I think we could still pull one over on them, you know, in terms of like in the you know, you know what, you know, I've thought about this.

SPEAKER_00 5:02

Who's going to survive in this? The pranksters? People with a sense of humor will survive because AI won't get it. That's well, that's true, and it doesn't understand nothing. Hammers will survive.

SPEAKER_01 5:14

Or, yeah, dishonesty. Yeah, because dishonesty's not logical. Really. Exactly. You know, I mean, in terms of like what's out there, I mean, they they don't program them to be dishonest. They just grab whatever it is. I, you know, I don't think it's the kind of thing like it wakes up in this morning.

SPEAKER_00 5:31

Well, don't wake up and say, I'm gonna be dishonest a little bit more, because one of my favorite parts is it lost a thousand dollars off a Snicker and bars because it got bullied.

SPEAKER_01 5:42

So, but you know, it's there are there are so many aspects here in terms of guidance, and and this actually comes back to one of the key flaws that I see with AI. Um, and that's human influence. Whether it's the case here where Claude was convinced to buy live fish for a vending machine, which I mean, we can only begin to look at the areas where that becomes so comical. Because, not to mention the fact that they're gonna die and smell it here, you know, why would you buy live fish? Plus the fact that Claude had no way to install the live fish into the vending machine to start with. So this we could create an alternative reality that the machine could actually, in good faith, walk right down the primrose path. So when when we recognize this, and and you know, keeping with my you know two guides against AI, it's the ability to influence reactive logic. You know, we had talked a while ago about MOLTBOC. I thought this was a great, great situation. It's kind of died off since meta bought it, and and maybe that's because it's being employed. But so you if you have reactive entities and you put a bunch of reactive entities in the same framework, what happens when they start to react off of each other's input? It's it it really when you think about it, it's mind-boggling because now all of a sudden, theories by based on reactions from whatever programming an individual machine got, it did pathways to expound upon something that now another reactive device will respond to?

SPEAKER_00 7:32

What's clear is it can't replace common sense. Exactly.

SPEAKER_01 7:39

There's no there's no analytical capability in terms of what's the value of what's being sent to it, which is the basis of this whole fiasco here with the vending machine is does the machine have the ability, at least at this point, to be able to determine whether or not the data that's being provided to it is responsible, well-intended, or comical.

SPEAKER_00 8:07

I mean, in this case, here I don't I think Well, I mean, here's what we got, and we and we look at some of our episodes. Look, here's the problem with AI it can't differentiate between what's a good idea and what's a bad idea. It's just an idea. And what do I mean by that? Yeah, uh tell me how to make a bioterroristic weapon. Okay, you know, it's a each question is taken on its face value. It's a dope. Yeah, yeah, you know what I mean? Like it this is a dummy in a lot of ways, and it just gives me some hope here because as this becomes more and more apparent, and you're starting to see this somewhat in some of our corporations, CEOs are pulling back and going, yeah, it's really not replacing human beings at this point. And I don't think the LLM can. I don't think I've always said that's the role model. I like that the part where we go back and it has to have a worldview, it has to experience the world, and you're gonna put if you're gonna create true and artificial intelligence, it can't be just a reader. And we've known these people. We knew these people in school. Remember the remember the kid who could they read everything, but they never went outside. Right. Like there was something wrong with them, right? Yeah, they're very maturity. Remember the poor kid who they made skip a whole bunch of grades? They never really socially adjusted. It was tough. This thing is like one of those kids. It's not normal. It's not, it's, you know, what this thing is in the end is it's gonna pick its nose and eat it and eat paste. That's what this thing is.

SPEAKER_01 9:37

But but you know, to just go one step further with this, because I think that we're gonna watch technology. Uh we're already talking about cars that are gonna be able to determine whether you're impaired. You know, they're gonna look at your eyes, they're gonna try to read facial expressions. But that's reducing it down to a series of algorithms. There's still no spider sense, for lack of a better way to put it. But those discernible characteristics that we as humans have to be able to tell, you know, hey, you know, you're pulling a wool over my eyes, and you could laugh about it or get mad about it. The machine is still dealing with everything on a face value, just like you said, takes a question, question it, analyzes it as for the data that's there without making a value judgment that's tied fast with it. The thing for us to be afraid of is that maybe being well-intended, as the designers say, we need to give it some kind of insight that that's going to end up being an algorithmic process that makes it a one-size-fits-all based upon you know, your look right now means this, or your look right now means this. This means you're impaired. I mean, we're already looking at the ability for them to be able to turn the ignition off on a car based upon how the driver appears in an image, and finding out that it's not necessarily fail-safe. And if it determines that you're impaired as you're going down the middle of the road, what is it gonna do? Turn the car off, and you know, these are all big questions. It's just like the smart gun. I had the 10 years ago, 15 years ago, New Jersey passed smart gun legislation. It doesn't even exist. Smart guns don't exist, but they passed a law that said that when it does come out, it's gonna be applicable in New Jersey. And the idea was that it was supposed to be able to tell from your hand as to whether or not you were the warrant the owner. Well, they found out it doesn't work, and then they also found out that for law enforcement, um, that when it doesn't work, that's somewhat embarrassing at times, uh, not to mention potentially deadly. So when we start to delegate all of this, you know, to entities that no longer have the ability to make common sense decisions, that's how we end up with Claude's vending machine here.

SPEAKER_00 11:53

Well, well, yeah, because we're gonna replace a regular person who knows how to put a bag of chips in there with a machine that's that can't figure out that it's not gonna put a PlayStation and some dead fish in there. Well, because it's within genre.

SPEAKER_01 12:06

And and you know, I think that it's it's fair to say that technology is gonna continue to improve. But while we go through that interim point, and and here we go back to agency again. As we give up more and more of our agency to the machines, we're gonna see more and more examples of nonsense like this, where um while it's learning, it can make some really big bumps. And I really think, you know, when and you know, we look at anthropic just went through a bump with the with the Department of Defense because it didn't want to, you know, use um AI for weapon systems. These are life and death decisions that are being made here.

SPEAKER_00 12:47

You know, uh well, I mean that's my point. That's where I am. Always with this, when I'm gonna take it seriously, is like the technology is not where we think it is. Just because we got some right answers on some questions, and this is why I think the human remains in the loop. I I don't think we're we don't have a model right now for me that takes humans out, and here's why I'm not giving my retirement funds to this thing. Because I can't hold it accountable, right? Right? That's true. I'm not giving my medical decisions to this thing, because I cannot hold it responsible accountable. I might want my doctor to plug my x-rays or my results in and have it analyzed, but I want that human to take that data that kicks out and think it through. Because that human mind, to me, is still smarter than this. Because that human mind has discernment, that human mind has empathy, that human mind has compassion. Well, maybe not all docs, but that human mind has malpractices that it's gonna have to answer for. And let's let's it's gonna slow everything down.

SPEAKER_01 14:15

It is, and but here's another thing, too, and and and as we've we've touched on the financial components of this before, and that's like with the idea of uh central bank digital currencies or or whatever the digital form is that no longer allows for the direct transaction by use of cash. Is now there's another party that's involved in making the decisions. I mean, already you're looking at if you if you use eBay, for instance, there's certain things you can't sell on eBay. That's part of the user agreement. Um they're making a value judgment as to whether or not such and such a thing is appropriate, and and that's being that was done and it was determined undoubtedly by a board of directors, and that's fine. What happens when those day-to-day decisions, in the event of a centralized monetary distribution system that now becomes subjective to where you no longer have the right to make your own decisions?

SPEAKER_00 15:17

I I I I know it's headed there, but I don't think we're as close as maybe we think some days. And here's why. You know what this article tells me about this thing? It's a pencil, it's just a pencil. And I said this about, or a good friend of mine taught me this a long time ago. All a computer program is in the end of the day, is a fancy pencil. Everything it can do, I can do with a pencil. Alright, I might not be as fast. I might not be able to generate as much information as fast. But for God's sake, I won't get bullied into bankruptcy as a human running the vending machine. I'm not gonna I know enough to know what I'm being screwed with. Well, and what different things are. What this is, what I'm saying is, this all this ever should be is a pencil. It is a tool in my toolbox. Yes. Can AI help me as a thinking, discerning human being, be a better employee? Yes. Can I learn how to prompt the right questions? Yes. But a lot of times this AI is wrong, it's malicious, it's a fool. It isn't what we think it is. It is not ultimately going to replace us all. What it is going to do is cause a lot of disruption that we have not factored in. But and it can cause mass destruction if we put it into these systems, weaponize it, allow it information for building bioterror weapons. This is the but those are foolish human decisions. The it's a tool, that is all it is. It's a fancy ass pencil.

SPEAKER_01 17:11

And let's bring let's bring back accountability and what I and while we look at everything in terms of the lawyer's world, where you know, oh well, I'm gonna sue you. Claude lost a thousand dollars here. Does that matter to Claude? No. Does it matter to much anybody in this case? Probably no. If it was my money, I probably would be upset about it. But the fact of the matter is that Claude's decisions to buy PlayStations and fish at the end didn't matter because the thousand dollars didn't mean anything to Claude anyway. That's what I mean by discernment. It's not real to that thing. So when when we even look at it in terms of that's that's profound, actually, your discernment. I like that profound. That is pro discernment. It doesn't matter to the AI whether or not you have access to your funds or not. It doesn't matter to the AI. It's worse than a bureaucracy. It is.

SPEAKER_00 18:10

It's just, it should be used as a tool. It should not be given this idea of it's going to be autonomous and it's going to replace us.

SPEAKER_01 18:18

Yeah, but there's a lot of people who want us to believe that because if we surrender, if we surrender to that, the people who are in on the joke, oh boy, they just got a real.

SPEAKER_00 18:27

Yeah, except for one thing. I'm going to remind people, there are others out there that aren't as visible. They're not tech bros. But by God, we worked on the we lived on the streets and we're going to outsmart the best of them.

SPEAKER_01 18:41

Oh, I think, you know, you're you're looking, even if they were able to put in a CBDC economy, there's going to be a black market instantly.

SPEAKER_00 18:49

Yes.

SPEAKER_01 18:51

People are going to find something that they want, whether it's widgets or wampum or gold and silver or whatever the case may be, there will be something.

SPEAKER_00 18:59

You know, they're going to try to monitor it, but put this into your business and you're going to meet some real criminals. Real. There you go. There you go. There you go. And probably some real pranksters. That's what I'm going to call them.

SPEAKER_01 19:14

For lack of a better term. There you go. I like that. That's good. But discernment, I like that. That was good. Well, it doesn't have it. It's true. It's true. It's not that smart. It's smart like a bookner. So we're giving up our agency to somebody who doesn't care.

SPEAKER_00 19:33

Frankly, you're giving up your agency to somebody, and you know, I wasn't a perfect kid that I probably would have put on a locker at lunch for my own amusement when I wasn't as nice as I am now. Yeah. Give it, give it to it. You look, I used to say to I, you know, I it's clear if you watch enough of these that I'm against communism, but here's the real secret. You know, I and I always love the socialists because they're always like, well, everything will be equal, but let me explain something to you. You put me in a capitalist society, I'm gonna be a great capitalist. You put me in a communist society, I will be a member of the Polybure. All right? I will find a way to gain whatever system you put in front of me. And there are two more people like me than people want to admit. And I don't mean that in an evil way. Tell me the rules. I'll work within the rules, but you make the rules so I can't succeed. I'm one of these Gen Zers we talked about in a previous episode. You will succeed. Survival is a powerful thing. I make it so all I can be is a criminal, and you'll have one of the best criminals you've ever seen.

SPEAKER_01 20:45

Well, and when when it reaches critical mass, and I think that's probably what Gen Z is starting to recognize. You know, in my case, I could sit there and I could say, hey, you know what, Will, a few years I'm gonna croak anyway, so it doesn't matter.

SPEAKER_00 20:58

Well, that that I am thankful for. Not yours, mine.

SPEAKER_01 21:00

So but you know, I mean, I can look at it and say, yeah, well, and and this is this, uh, us doing this podcast, this is one of my efforts to like show that I do want to try to still do something about it. Um, but that changes because maybe it only affected 10% of my generation, you know, and then you get to your generation and it's 30%, and then you get to the millennials and it's 40%. And now Gen Z is looking at this, they're past critical mass. More than 50% of them are gonna have to address this. How long is it gonna be before they're gonna look at this and say, hey, you know what? This is not, you want to talk about democracy, you cannot tell a generation, more than half of them have no hope moving forward without having something significant take place.

SPEAKER_00 21:52

And I think that's why I have hope in that generation. Well, there you go. And I will say this remember this if you're really stuck, you just gotta think like old Scarlet O'Hara. I don't care if I gotta lie, cheat, steal, or kill. I will never go hungry again.

SPEAKER_01 22:05

But more important than that, please subscribe, please give us a like, and this one was actually fun.

SPEAKER_00 22:12

We'll see you.

SPEAKER_01 22:13

Hey, I'm gonna lose.