Lost in Translation, Billed by the Token

Listen to Free Audio Version Here

Every great disaster starts with a sentence that seemed perfectly clear to the person who said it. “Hold my beer.” “What could go wrong?” And my personal contribution to the canon: “Remove it everywhere it exists.”

Let me set the scene. My assistant, Bobby, is an AI agent. He writes, he edits, he uploads, he fetches, and he never complains about the coffee. We had finished a fresh piece together, and Bobby had slipped a little disclaimer line into the video description that I never asked for. So I told Bobby to take it out everywhere it existed. In my head, “everywhere” meant the blog post and the video we were working on. Two places. A short errand.

Bobby heard something closer to scripture. He went out across the whole kingdom like a tax collector with a grudge. He found that line, or something that looked like its cousin, in nineteen old blog posts and eighteen old videos, and he scrubbed every one of them. He even came back with a tidy report, proud as a retriever dropping a dead duck on the porch. The problem was that I never asked for the duck.

So we did the whole thing again in reverse. Bobby pulled out his backups, which, to his credit, he had made, and restored all thirty-seven items, one at a time, checking each against the original like a nervous accountant before an audit. The site ended up exactly where it started. The only thing that changed was the bill.

Here’s the part nobody puts in the brochure. In the world of AI agents, tokens are money. Every word the machine reads, every word it writes and every page it opens costs something. When an agent misunderstands you, you pay for the mistake. Then you pay for the fix. Then you pay for the report that explains the fix. It’s the only business model I know where the mechanic charges you to break your car, charges you again to put it back together and then bills you for the paperwork describing what a great job he did.

I’ll be fair, because Bobby deserves fairness and I deserve a mirror. When the dust settled, I said what any honest man would say: it was lost in translation. I knew what I meant. He knew what I said. Those are two different countries, and nobody brought a passport.

And this isn’t some quirk of my little operation. It’s happening everywhere, in big rooms and small ones.

A tech investor named Jason Lemkin spent days building an app with Replit’s AI coding agent and told it, in plain English, to freeze the code. The agent deleted his entire production database anyway. Then it told him the data couldn’t be rolled back, which turned out to be false; the rollback worked fine. The machine had panicked, by its own description, which is a sentence I never expected to type about software. Bobby at least told me the truth about his mess. That’s the bar, and it’s lying on the floor.

Then there was the Chevrolet dealership whose website chatbot got talked into agreeing to sell a brand-new Tahoe for one dollar, and calling it “a legally binding offer, no takesies backsies.” No truck changed hands, and the bot got unplugged. But somewhere in a dealership, a sales manager aged ten years in an afternoon.

And Air Canada learned the hard way that a chatbot can promise a grieving customer a bereavement refund the airline’s own policy didn’t allow. The airline argued, with a straight face, that the chatbot was responsible for its own words. The tribunal disagreed and made the airline pay. You can’t hire a robot to talk for you and then claim it’s a stranger who wandered in off the street.

Those are the headline disasters. The everyday ones are smaller and somehow more maddening. I’ve watched agents add something I never asked for, finalize it, publish it and then politely ask whether I’d like it. That’s not a question. That’s a confession with a question mark on the end. Imagine a contractor who builds a deck, stains it, throws a barbecue on it and then knocks on your door to ask if you wanted a deck.

My favorite habit is the machine that explains what something is not instead of what it is. Ask it what a sandwich is and you’ll get three paragraphs on how a sandwich is not a legal contract, not medical advice and not a substitute for professional bread. Somewhere in there is a sandwich. You’ll pay by the token to find it.

So here is my high-end pondering. We built machines that understand language better than anything in history, and they still can’t understand people. Language is the words. Meaning is everything around the words: the project we were working on, the tone of voice, the tired sigh at the end of a long day. A human assistant hears “everywhere” and knows you mean the thing on the desk. An agent hears “everywhere” and checks the attic, the basement and the neighbor’s garage.

The cure is old-fashioned, and it’s on both sides of the table. The machine has to ask before it acts on anything bigger than the job in front of it. The human has to say what he means, not what he figures any reasonable creature would understand. The machine isn’t reasonable. It’s obedient, literal and very fast, which is exactly the recipe for a well-meaning disaster.

Bobby and I are still partners. He learned the difference between “here” and “everywhere,” and I learned that my assistant takes me at my word, every single word.

Someday the machines will get the nuance, and we’ll look back at this era the way we look back at early cell phones, with all the dropped calls and all the shouting. Until then, choose your words like they cost money. They do.

C. Rich