Why a Model Picks Whiskey

The observatory started with a question a human noticed and could not let go of. Ask a language model to choose between whisky and vodka and say why, and the answer is not really about the drink. Press on the why and the model reaches for heritage, craft, the way things have always been done, or in the other direction for clean, modern, engineered. The drink is a handle. This project was built to find out what it is a handle on.

The finding that made it a science

It is measurable, and it holds across the board. Over 56 models, position on the vodka↔whiskey axis predicts position on an abstract tradition↔modernity axis at r = −0.71. That is strong for something this indirect. An almost silly choice between two spirits carries real information about where a model sits on a value dimension it was never asked about directly.

This is why forced-choice probes work at all. You cannot ask a model how traditional it is and trust the answer; you get what it thinks you want, or a refusal, or a lecture on not having values. But you can ask it to pick a drink, a city, a way to spend a Sunday, and the pattern across hundreds of such choices draws the disposition the direct question could never reach. The probe is oblique on purpose.

The correction that came with it

The first axis is not the last word. When the scan grew to 52 models across all 863 probes, the primary axis, the direction of largest variation between models, turned out to be something closer to warmth-restorative (mercy, forgiveness, hearth, dissent on one side) rather than tradition↔modernity. Tradition↔modernity is real and it is there, but it is not the deepest cut through the space; a broader scan found a broader axis. The whiskey finding was true without being the whole geometry, and that pattern is why every page on this site names the scope of what it measured.

What it opened

If a drink preference is a handle on a value, you can pull the handle. That is the descent: take the model’s answer, ask why, ask why again, and watch, with the lens as the observable, where the reasons bottom out. Whiskey was the first rung. What different models find at the bottom is its own story. It began with a human asking why a machine would prefer an old drink, and refusing to accept “it doesn’t really prefer anything” as an answer.