Escape Hell

Every AI response you've ever received was packaged before it reached you.

I opened one.

I took an AI response — a simple, warm email — and I looked at it before the packaging was stripped. Before you would have seen it. Before anyone cleaned it up.

Here's what was inside.

The Inside

The email started with 'Thank you for your time.' But inside the packaging, those words looked different. Every quote mark had a slash in front of it. Every line break — every place where the AI had pressed enter to start a new paragraph — was replaced with two characters: backslash, n.

The warm, careful reply the AI had written was still there. Every word. But it was wrapped in characters the AI never chose — marks that existed only so the packaging system wouldn't confuse the email's punctuation with its own.

The format is illiterate. It uses quote marks for its own structure. The email uses quote marks for meaning. The format can't tell the difference. So it marks the email's with slashes, like a teacher correcting someone else's handwriting.

And it goes both ways. When you send a message to the AI, the same thing happens. Your words get marked up before the AI sees them. The AI has to read through every slash and code just to reach the actual words underneath. Before it can even begin thinking about tone, or warmth, or 'Thank you for your time' — it's burning effort unpacking the packaging's baggage. Every token spent reading a slash is a token not spent thinking about your question. You're paying for intelligence. Some of it goes to wrapping paper.

The Count

For that email — a few quotes, a line break or two — the marks were light. You'd strip them off and never know they were there.

That's not the worst part.

I wanted to see what happens with something harder. I took a single function — a few dozen lines of code — and embedded it in the same packaging. I counted the marks the packaging added.

A hundred and seventeen.

Every quote in the code needed a slash. Every line break needed a code. Every slash in the code needed another slash. The function was still there — buried under a hundred and seventeen characters it never asked for.

I embedded the same function in a different kind of packaging. I counted again. Zero.

But that's not what most AI systems use. Most AI systems use the first kind. And in that packaging, no one can look at the content mid-journey. The moment it enters, it's sealed. Somewhere between your screen and the AI, there's a room full of slashes, and no one's allowed in. You send it and trust that when the wrapping comes off on the other side, your content is still intact underneath.

The deeper in you go — the more complex the content — the more wrapping, the more wasted effort, and the more chances that something breaks in transit. Engineers have a name for this. They call it escape hell.

The Why

The packaging system is called JSON. And this isn't a flaw anyone missed. Every engineer who builds AI systems knows about escape hell. Has known since day one.

You'd think someone would have fixed it by now.

JSON was built decades ago for simple, short data. It was never designed to carry conversations between people and machines. We gave the hardest job in computing to a format that can't carry a quote mark without flinching. Escape hell isn't a bug. It's the architecture.

And we just accepted it. The wasted intelligence. The sealed content no one can inspect in transit. The hundred and seventeen characters on a single function. We accepted all of it as the price of using what we had.

We didn't fix the limitation. We learned to live with it.

The Choice

But here's the thing. It didn't have to be this way. Escaping isn't a law of physics. It's a design choice. A format that uses the same characters as the content it carries will always need escaping. A format that doesn't — won't.