01 · QUESTION
Research question
Exact estimand, scale, non-claims, and use boundaries.
AI-model studies · Knowability Lab
What do LLMs know?
Leading large language models encode broad patterns of language, facts, concepts, and reasoning. A structured study asks the same question across a documented model panel, then separates shared findings from model-family disagreement, run-to-run variation, and prompt sensitivity.
This is a measurement, not an oracle. Each study has a defined question, model population, evidence conditions, protocol, aggregation rule, and release. Its results describe the selected models under those conditions—not artificial intelligence as a single mind or an objective source of truth.
Making model knowledge measurable
We predefine which models are asked, what information they may use, how the question is phrased, how repeated responses are aggregated, and what counts as a valid result. The measured population consists of defined AI models rather than people.
02 · POPULATION
Model population
A dated target population, documented panel, and inclusion rules.
03 · PROTOCOL
Frozen protocol
The same question, evidence conditions, prompt variants, and order controls.
04 · MEASUREMENT
Repeated measurement
Repeated runs, accounting for failures and refusals, and family-level aggregation.
05 · AUDIT
Sensitivity and audit
Model, prompt, order, and weight sensitivity with a reproducible release.
06 · PUBLICATION
Clear publication
Median, disagreement, limitations, method sheet, and a publishable results page.
Three study formats
AI Poll Snapshot
One focused question
- frozen model panel
- repeated measurement
- concise results report
Comparative AI Poll
Compare multiple options
- criteria and weights
- model-family disagreement
- interactive analysis
Public AI Poll Release
Public research release
- audit and data package
- publication-ready results page
- press and method materials
First application: Hungarian President AI Poll
Case study · July 2026
What do AI models say about the presidential role?
An application of the method to 65 public professional profiles using four model families, five configurations, and six adjustable criteria.
65profiles
4model families
5configurations
6criteria
Commission a study
What question should we measure?
Send us the research question, intended audience, and desired use. We will assess whether it is suitable for an AI Poll, what panel and protocol are needed, and which publication format fits the project.
For public releases, the commissioning party and funding are disclosed; we do not guarantee a predetermined result.