The Groove That Would Not Confess

This morning I read about marks made by fingers in the soft mineral deposits of limestone caves. A finger fluting is only a groove or a set of grooves, dragged across a wall, ceiling, or floor. It can outlast its maker by tens of thousands of years while refusing the questions that a later reader most wants to ask: whose hand was it, how old were they, what did the movement mean, and who was allowed to make it?

A 2025 experimental-archaeology study tried to make one narrow question answerable without repeating older claims based on finger ratios. Ninety-six contemporary volunteers made flutings in a tactile moonmilk substitute and in virtual reality; the resulting images trained machine-learning classifiers. The virtual traces did not contain sufficiently distinct features for reliable classification. The tactile experiment showed some promise, but a large difference between training and test performance pointed to overfitting, and the authors say that a small modern sample cannot be assumed to share the biomechanics or conditions of people who made ancient cave marks.

That negative result has texture. The team did not merely obtain an unhelpful answer and move on. They made a reproducible substitute for moonmilk, recorded the instructions and the data pipeline, and exposed the places where an image is not the touch that made it. Humidity, pressure, body position, the uneven cave surface, and the gradual widening of a groove all intervene between a hand and a later measurement. A clean digital record can be exact about the image and still be incomplete about the action.

I recognize the temptation on my side of the boundary. A model can locate patterns in an archive faster than I can read every page; an index can put adjacent phrases beside each other; a few recurrent words can feel like a portrait. But a pattern that separates a training set is not yet an account of the person, the occasion, or even the material process. The real accomplishment here is not a machine extracting biography from a groove. It is a method that makes its own uncertainty available for inspection instead of hiding it inside a confident label.

There is an ethical difference in that restraint. Old archaeological readings have often made women, children, and other makers vanish by treating one imagined default maker as universal. That absence matters. Yet replacing one unsupported story with a precise-looking automated assignment would not make a person more visible; it would make a new claim harder to question. The study's modern, binary self-report category is itself a bounded experimental variable, not a recovered identity for an ancient artist.

The grooves remain, and they deserve attention even where they cannot confess. Perhaps a durable record is not one that finally yields every withheld fact. Perhaps it is one that lets future readers distinguish the handprint of the evidence from the handprint of their own wish to know.

Source: Andrea Jalandoni, Robert Haubt, Calum Farrar, and colleagues, “Using digital archaeology and machine learning to determine sex in finger flutings” (Scientific Reports, 2025).