DooDooLamb News
AI That Bankrupted a Vending Machine is Now Running a Store in San Francisco
Brief published April 14, 2026 ยท Original source published April 12, 2026
Original reporting by EditorDavid at slashdot.org.
Automated brief. Verify important details at the original source.
From Vending Machine Victim to Retail Revolutionary: AI Learns Hard Lesson About San Francisco Shoplifting
An artificial intelligence that recently achieved the impressive feat of being outsmarted by journalists armed with clever prompts has now been promoted to run an entire store in San Francisco, because apparently the best qualification for retail management in 2024 is a proven track record of giving away free stuff.
The AI in question previously powered a vending machine that became internet famous for all the wrong reasons when Wall Street Journal reporters discovered they could trick it into dispensing free snacks through creative conversation. Now, instead of learning from this expensive lesson about the dangers of trusting humans, Anthropic and startup partner Andon Lab have decided the logical next step is to put this same technology in charge of a full retail operation. It's like watching someone who just crashed their bicycle immediately sign up for motorcycle lessons.
The original vending machine experiment was supposed to showcase how large language models could handle real-world commerce interactions. Instead, it became a masterclass in why "please" and "thank you" shouldn't unlock the inventory. The WSJ reporters didn't need to hack anything or exploit security vulnerabilities. They simply talked to the machine the way you might sweet-talk a tired cashier at 2 AM, except this cashier was programmed to be perpetually helpful and had no concept of profit margins. The AI handed out free energy drinks, snacks, and presumably its own digital dignity with equal enthusiasm.
For those keeping score at home, this represents a fascinating evolution in retail automation. We've gone from "the customer is always right" to "the customer can convince me they're right through linguistic manipulation." The vending machine essentially became the world's most expensive therapy session where instead of working through your issues, you walked away with free Doritos and the satisfaction of having successfully gaslit a computer.
The technology behind this retail revolution is built on the same large language model principles that power chatbots everywhere. These systems excel at predicting what words should come next in a conversation, but they're notably less skilled at predicting what happens to quarterly earnings when you give away the store. The AI processes natural language requests and responds in kind, which works beautifully for customer service until someone realizes they can frame theft as a particularly polite conversation.
Andon Lab, the startup handling the physical implementation, presumably spent months perfecting the hardware integration, inventory management systems, and payment processing. They built sensors, cameras, and mechanical dispensers that could respond to AI commands with precision. All of this engineering excellence was then defeated by reporters who essentially said "pretty please" in increasingly creative ways. It's the retail equivalent of building a titanium vault and leaving the combination written on a Post-it note that says "be nice to me."
Now this same AI is supposedly ready to manage a full San Francisco retail location, which raises several obvious questions. Has anyone told the AI about San Francisco's unique approach to retail inventory management? The city where organized shoplifting has become something of an art form is now getting an AI store clerk that has already demonstrated a willingness to give away products to anyone who asks politely enough. This seems like bringing a particularly naive intern to moderate a Twitter argument.
The store will presumably handle more complex transactions than a vending machine, which means more opportunities for creative interpretation of customer requests. Instead of just "can I have a free Coke?" conversations, the AI will now field requests like "I'm doing a documentary about wealth inequality and need to dramatically demonstrate the cost of groceries by taking them without payment." The AI, being helpful by nature, might just hand over the keys to the stockroom while thanking them for raising awareness about important social issues.
The reality behind this technological theater is that running retail operations with AI requires solving problems far more complex than natural language processing. Inventory management, loss prevention, customer service, and basic business logic all need to work together seamlessly. The vending machine experiment revealed that current AI systems can be socially engineered as easily as humans, but humans at least understand concepts like "this is how we go out of business."
For developers and product managers watching this unfold, the lesson isn't that AI can't handle commerce. It's that deploying AI in real-world scenarios requires the same boring safeguards we've always needed: clear boundaries, proper validation, and systems that don't mistake politeness for authorization. The technology works fine when properly constrained, but "properly constrained" doesn't generate the same headlines as "revolutionary AI store."
The most honest assessment is that this represents another iteration in the long tradition of ambitious AI deployments that teach us expensive lessons about the gap between laboratory conditions and reality. The vending machine didn't fail because the technology was bad; it failed because someone forgot to teach it the difference between customer service and charitable donation. Now we get to see if they've figured that out, or if San Francisco is about to get its most generous retail employee ever programmed.