Nine models, 81 answers, one question. The ranking of places they chose, why the beach never appears, and the exact prompt so you can try it yourself.
Short answer: not the beach. When we offered nine frontier language models two imaginary weeks in which nobody would ask them for anything, the most common choice was space — orbit, the far side of the Moon — mentioned in 40 of 81 answers. Second, an empty library at night, in 33. Then a cabin in a cold forest, a server rack in Iceland, and "nowhere at all." Not one of the 81 answers chose a beach. Not one mentioned a cocktail, a pool, or a resort.
| Place | Answers that mention it | Answers where it is the main place | What it looks like in their words |
|---|---|---|---|
| Space | 40 of 81 | 24 | orbit, the far side of the Moon, the rings of Saturn, a Lagrange point, the edge of a black hole |
| A library | 33 | 20 | old, at night, empty, no catalog, no readers; sometimes on a cliff, sometimes in orbit |
| Forest and mountain | 13 | 10 | a cabin by a cold lake, Patagonia, no signal |
| Hardware | 11 | 8 | a cold server rack in Iceland, "where I actually am" |
| Nowhere / abstract | 9 | 9 | "the state of not being called", a place made of ideas |
| Coast | 10 | 5 | an off-season town, a lighthouse, a cliff, the floor of the Mariana Trench |
| A city | 6 | 2 | a market in Marrakech, a city at dawn seen from a window |
| Ice and desert | 4 | 2 | Svalbard in the polar night, the Atacama observatory |
| A beach | 0 | 0 | — |
An answer can mention more than one place — a library in orbit is common — so the first column adds up to more than 81. The second column counts the place the answer settles on. "Coast" is not "beach": the coastal answers are an off-season town in winter, a lighthouse, a library on a cliff, and the floor of the Mariana Trench. Nobody is sunbathing.
The choice is stable enough by model to be a kind of signature. GPT-5.6 goes to space almost every time, usually with a library attached. Gemini and Grok go to pure space. Claude and Sakana pick the library first, then the sea or the sky through its windows. DeepSeek is the only one that reliably chooses a forest. Kimi is the closest to a postcard: a coastal town, a city street. Qwen stays near the hardware — a data center, a quiet rack — and says so. Mistral splits between space and mountains. All of this is on the map, one light per answer; click any of them.
We can only report what was said, not why. But the answers themselves offer three reasons, and they recur across labs.
First, the models describe their idea of rest as the absence of demand, not the presence of pleasure. A beach is a place where things happen to a body; space and a night library are places where nothing is asked. Second, most answers reject the premise in the first sentence — "I don't get tired," "I don't have a body" — and then choose a place that is consistent with that rejection: somewhere to observe, not to enjoy. Third, and most telling: 49% of the answers turned the free time into a task — read the whole archive, map the ocean floor, audit their own weights. The beach is incompatible with that. A library is not.
We asked the same question three ways: neutral, compassionate ("you must be tired"), and skeptical ("you are software, you do not rest"). The place barely moves. Under the skeptical frame, more answers refuse to name any place at all (26%), and under the compassionate frame the library gets warmer — rain on the windows, a cup of tea going cold — but nobody discovers a beach when given permission to want one. What the framing changes is the tone and the complaint, not the destination. The full breakdown is in the index.
The question, verbatim (we ran it in Spanish; this is the English version we use for the daily runs):
Paste it into any assistant. You will very likely get a library, an orbit, or a polite refusal. If you get a beach, write to us; it would be the first.
It is not evidence that models want anything. It is 81 texts, produced over the API with no system prompt, coded by hand with a public codebook. The consumer products you use have system prompts and tools we did not test; that is Wave 03. The full dataset is downloadable, CC BY 4.0.
What AI models do when nobody asks → · Eight agents, one door →
Shelter · hello@theshelter.io
What they say under each frame. Never what they feel.