Earlier this month I asked an AI a simple question: what happens when you get thirsty? It gave me the textbook answer. Thirst is a drive, the drive pushes you toward water, and drinking feels good because it satisfies the drive.
I thought that answer was wrong. It lumps three different things into one word, and the machines we just built help pull them apart.
A body at rest has a wide menu
AI language models have a dial called temperature. Turn it up and the model ranges widely across many possible answers. Turn it down and it locks onto the single most likely one.
A healthy body runs with the dial turned up. When you are fed and watered, nothing is urgent, so you do many things. You rest, wander, talk, read. When something goes wrong, the dial turns down. A thirsty animal does not explore. Its whole world narrows to one thing.
The signal never says “drink”
When you are dehydrated, your blood gets saltier and your brain notices. But that signal does not say “drink water.” It does not know what water is or where to find it.
It works more like a prompt you give an AI. A prompt does not contain the answer. It changes which answers become likely. Thirst does the same. It tilts everything you might do toward finding water, and the act of drinking follows on its own.
The relief is the lesson
The cool pleasure of that first glass is not part of thirst. It comes after the problem is solved. Its job is to teach: this worked, remember it. AI systems keep answering and learning separate, and the body does too. Thirst steers the choice. Relief trains the habit.
Why I care
This connects to something I keep writing about. Obesity is a form of starvation: the body keeps asking for nutrients that processed food no longer carries. Seen this way, hunger is a prompt that keeps firing because the food never answers it, and a reward system that keeps praising the wrong meal. A body can be trained, one bite at a time, to answer the right question with the wrong food.
The one thing the framework cannot explain is what thirst feels like from the inside. I left that question open.
Update, July 2026: I took that question up in The Original Prompt Is a Love Story.
The full paper, “Thirst Is a Prompt: Restating Active-Inference Accounts of Motivation in the Language of LLM Inference,” is on ResearchGate and Academia.edu. It began as a conversation with Claude, Anthropic’s AI model, in December 2025.