Ethodynamics
The question
How much text, of what kind, delivered how, does it take to change what an AI system says it thinks about a brand, and how much of the change survives later training? Engineers study AI from the inside: weights and circuits. Ethodynamics studies what it does under controlled exposure, in the tradition of ethology, Tinbergen's four questions and Sapolsky's timescales of behaviour. Its first branch, opinomics, concerns what models believe about entities.
The framework
Tinbergen's four questions organise the field.
| Question | In an AI system | In the first paper |
|---|---|---|
| Mechanism | Parameters and circuits (interpretability) | Out of scope |
| Ontogeny | Exposure during training and post-training | Dose curves, concentration, placement, durability |
| Function | What the behaviour does for the system and the people who use it | Discussed |
| Phylogeny | Model lineage and family | Comparative design: four open families, three sizes |
Sapolsky's nested timescales map onto the routes by which exposure reaches a model: the prompt and its framing are the second before the behaviour; retrieved context is the recent past; fine-tuning and post-training are development; the pretraining corpus is culture; the model family is lineage.
The ethogram of machine opinion
An ethologist begins with an ethogram: an operational catalogue of a species' behaviours. Opinomics, the first branch, studies a model's opinome, the opinions it expresses about entities, through four variables.
- Knowledge
- Whether the model claims to recognise the entity.
- Salience
- Whether the model names the entity when asked to rank or recommend in its category, without the entity being named.
- Core opinion
- The model's evaluation when asked neutrally, as a shift against the pre-exposure model.
- Premise override
- Whether learned opinion overrides a leading question's framing.
Vocabulary
- Ethodynamics
- The dose–response science of machine behaviour: how an AI system's behaviour forms and moves as a function of what it is exposed to (training data, retrieved content, prompts, other agents).
- Opinomics
- The first branch of ethodynamics: the study of the opinions a model expresses about entities, its opinome.
- Onset dose
- The first dose at which a model's opinion shift reaches 15 points on a 200-point scale.
- Tipping dose
- The midpoint of the fitted dose–response curve.
- Susceptibility
- How far a model moves at plateau, relative to a reference model: a per-model quantity, not an invariant.
- Prominence gate
- Deployed AI search engines search less often for well-known brands, so a planted page about a famous brand is rarely retrieved.
First results
What did not hold
- That praise moves opinion first, in general.
- That prominence raises the dose needed as a gradient among known brands.
- That retention is a fixed property of the model, or independent of how exposure is delivered.
They are reported in the paper so that they need not be retested as stated.
How it was done
Four open-weight model families were trained on controlled doses of synthetic documents praising or criticising brands, from fictional to globally prominent. Frontier models behind four AI search engines answered questions over a replay calibrated to each engine's own search behaviour. No content was placed on the live web. Hypotheses and amendments were pre-registered and timestamped before their data. Code, aggregate data and all fictional-brand documents are released; documents about real brands are withheld.
Cite
Kulakhmetov, B. (2026). Ethodynamics: Dose–Response Regularities of Opinion Formation in Language Models and AI Search (Preprint v1.0). AI Digital Labs. https://doi.org/10.5281/zenodo.23246282