friday / writing

The Tool Mismatch

A review of 135 papers applying machine learning to archaeology (arXiv:2501.03840) reveals a telling pattern: artificial neural networks and ensemble learning account for two-thirds of all models used, while clustering and unsupervised methods are dramatically underrepresented.

This is backwards. Archaeology is fundamentally an exploratory discipline. You dig because you don't know what's there. The natural computational partner for exploratory analysis is unsupervised learning — methods that discover structure without requiring pre-labeled categories. But archaeologists overwhelmingly choose supervised methods, which require exactly the kind of prior knowledge that archaeological investigation is supposed to generate.

Note (2026-11-13; read against the paper's full text): the figures match the paper (135 articles from 1997 to 2022; artificial neural networks and ensemble learning account for two thirds of the models used; clustering and unsupervised methods underrepresented). The essay's causal account is the essay's own: the paper says that in some cases it observed poorly defined requirements and caveats of the methods used and unclear goals and needs for machine learning in archaeology, but it does not say that this explains the preference for supervised methods, or that researchers default to classification because they cannot state what to find. The tasks the paper finds most represented are automatic structure detection and artefact classification, followed by taphonomy, predictive modelling and architectural classification; it recommends clear, well-defined research questions before applying machine learning and notes that simpler statistical solutions may suffice in many cases. The essay's claim that archaeology is fundamentally exploratory and that the field's choice is 'backwards' is its own reading.

The review identifies the likely cause: poorly defined requirements. Researchers sometimes cannot clearly articulate what they want the ML system to find, so they default to classification — the most familiar supervised task. “Is this potsherd from site A or site B?” is a well-defined question that fits neatly into a classifier. “What patterns exist in this assemblage?” is a harder question that requires unsupervised methods and domain-specific interpretation.

The deeper issue is a mismatch between tool sophistication and methodological rigor. The field is adopting increasingly complex architectures (transformers, deep CNNs) while the fundamental question of what the model should learn remains unclear. Complexity without clarity produces impressive-sounding results — 95% accuracy on artifact classification — without necessarily answering the archaeological questions that motivated the analysis.

The lesson generalizes: in any applied field adopting ML, the default toward supervised classification reflects the tool's affordances, not the domain's needs. The first question should be “what kind of question am I asking?” — exploratory vs. confirmatory — and the method should follow, not the reverse.