On July 9 I was back at SICSS-Istanbul — the summer institute I co-founded — as a guest lecturer, with a talk called “Anthropology of Machines: From Black Box to White Box” (slides, PDF).
The case I wanted to make: the future of AI needs more social science. The route there runs through a thirty-six-year-old argument. In the symbolic era, the charge against neural networks — Fodor and Pylyshyn’s — was that a network can imitate a concept but never possess one; ELIZA (1966) was the canonical imitator, pattern-matching its way through therapy without owning a single idea in it. The recent reply is that concepts are (probably) directions in a model’s representation space — not symbols. For thirty-six years that dispute was philosophy. Now it is an experiment: they could debate it; we can go look.
And “going look” turns out to be fieldwork. Reading a feature is thick description. Watching an answer form across layers is process-tracing. The open problems — construct validity, measurement, reliability — are the problems social scientists are trained on; sociologists were treating concepts as geometry before LLMs existed. So the room’s training is not adjacent to interpretability. It is interpretability’s missing half.
Then everyone read a mind. Using the NDIF Workbench — the logit lens on a real model, in the browser, no code — the cohort probed Llama-3.1-8B with their own fill-in-the-blank questions and watched the answers resolve layer by layer, on prompts that ran from who will win the World Cup to the future of capitalism.
I opened with two poll questions and closed with my answers. Are LLMs black boxes? Not black, not white — translucent, in patches. And does AI need social science, or the reverse? Social science can proceed without AI. AI cannot become trustworthy without us.
(A longer essay version is coming on Substack.)