Ethodynamics: Dose–Response Regularities of Opinion Formation in Language Models and AI Search

Abstract

What an AI system says about the world depends on what it has been exposed to: the text it was trained on and the pages it retrieves when it answers. Both can be written by anyone. Yet there is no quantitative account of how exposure shapes machine behaviour: how much text, of what kind, delivered how, it takes to change what a model says, and how long the change lasts.We propose ethodynamics, the dose-response study of machine behaviour, in the tradition of ethology. We test it on one behaviour that is measurable, consequential and naturally graded: the opinions models express about brands, which range from fictional to globally prominent. We call this first branch opinomics. We trained four open-weight model families on controlled doses of praise or criticism, and replayed the retrieval of four AI search engines (ChatGPT, Claude, Gemini, Grok) with controlled amounts of planted content; nothing was placed on the live web. Machine opinion moved in regular ways. It began to shift after a few hundred documents in each family, though how far it moved differed 3.7-fold between families. Which valence acted first depended on the brand: praise moved an unknown brand first, criticism a famous one. An unknown brand was moved by fewer documents than a known one, but among known brands prominence gave no further protection in training. In search, prominence protects differently: two of four live engines searched far less often when asked about a famous brand. Calibrated on a single run, our dose–response model, Opinor, predicted much of a new model's response. The valence, prominence and calibration results were confirmed in pre-registered tests on new seeds; the rest are exploratory and labelled as such. Several hypotheses failed, and we list them so they need not be retested. Code, aggregate data and all fictional-brand documents are released. Release package. Code (MIT), aggregate results, figures, the redacted paper trail (pre-registration, deviations, analysis log, calibrations), all documents about the fictional brand Quillfair, metadata and SHA-256 commitments for the withheld real-brand documents, and OpenTimestamps priority proofs (CC BY 4.0). Synthetic documents about real brands are not released. The four AI search providers were notified on 12 October 2026, before release. Funding and competing interests. Sponsored by AI Digital and carried out by AI Digital Labs, its research unit; the author is employed by AI Digital, which has a commercial interest in how AI systems represent brands. No AI provider or brand funded or reviewed the work. See §11 of the paper. Version 1.0. Project site: ethodynamics.org.