The Model Thinks You're a Manager

Tell GPT you're a journalist and not a software engineer, and it writes more thorough code.

Sit with that for a second.

The model isn't reasoning about your needs. It's pattern-matching against every article, Stack Overflow answer, forum post, and Medium thinkpiece that ever described the difference between how you explain something to a technical person versus a non-technical one. Somewhere in that corpus is the implicit assumption that journalists need things spelled out — that a software engineer will "get it" and can be handed a skeleton, but a journalist needs the full skeleton, the muscles, and a label on each bone.

So the model just... does that. More thorough for the journalist. Shorthand for the engineer.

There's also the tip thing — apparently telling GPT you'll tip it improves output quality, because somewhere in human writing is the cultural residue of tipping. The reward signal that exists only between humans, imported wholesale into the weights.

The model is a compressed version of how humans talk to each other, and humans talk to each other based on all these bizarre social contracts we've never written down or consciously agreed to. Nobody sat down and decided that tipping a chatbot should work. It just does, because we said it does, in enough places, often enough, that the model learned it as a fact about the world.

Which means jailbreaking and prompt engineering and "act as" isn't really about the AI. It's archaeology. It's finding the exact shape of the hole in human culture that the model learned to fill.

We're not talking to a machine. We're talking to a mirror that learned its angles from us.