Whether a synthetic persona is reliable doesn't depend on the language model.

The Big Five taxonomy — openness, conscientiousness, extraversion, agreeableness, and neuroticism — is the dominant framework of modern personality psychology. It goes back to lexical studies of the early 1980s and was turned into a test-based inventory (NEO-PI) by McCrae and Costa¹. Its decisive advantage: it is not culture-specific — it replicates across languages and cultures.

For synthetic personas, that means we can express a Persona's psychological disposition as a five-dimensional vector — one that can be measured consistently in surveys, in behavioral experiments, and in language-pattern analyses. That is not an approximation of a “type.” It is an operationalizable quantity with known reliability and validity.

In 1979, Kahneman and Tversky³ showed with Prospect Theory that human decisions under risk deviate systematically from normative expected-utility theory. People overweight certain gains, underweight small probabilities, and react more strongly to losses than to equivalent gains. These deviations are not errors — they are predictable patterns.

On top of that come more than 180 documented cognitive biases, many of them replicable in real-world settings: status quo bias, anchoring, the availability heuristic, in-group bias. For each of our Personas we select a subset of these biases — deterministically, not at random — and make it a fixed parameter of their decision-making. That is the core of what separates a Persona from a stereotype.

Hofstede⁴ laid the groundwork with his 1980 IBM study and kept developing his cultural dimensions over the following decades — the most influential edition of “Cultures and Organizations” appeared in 1991 and was revised in 2010. His dimensions show that national cultures differ systematically along a small number of measurable dimensions: power distance, individualism, uncertainty avoidance, long-term orientation, masculinity, and later indulgence. The dimensions are contested — they reduce complex societies to a handful of numbers — but they are replicable and predictively useful in organizational and consumer contexts.

For the German-speaking market, we complement Hofstede with our own value-based Life-Worlds⁵ — derived from established values and lifestyle research. They map the social structure of the DACH region into roughly ten Life-Worlds (e.g. The Preservers, The Balanced, The Pioneers, The Stewards). The two frameworks interlock: Hofstede provides the national macrostructure, the Life-Worlds the fine-grained structure within a nation.

The last building block is new — and it is the most fragile one. In 2024, Park et al.⁶ showed in a study of 1,052 participants at the Stanford Human-Centered AI Institute (arXiv preprint, peer review still pending) that interview-based AI agents replicated their human counterparts' answers to General Social Survey items with an accuracy of about 85% of the participants' own two-week test-retest reliability. On top of that: less demographic bias than description-based agents.

What the paper shows: language models can — when conditioned correctly — plausibly generate survey answers for fictional people. What the paper does not show: everyday behavior, purchase decisions, emotional surprises, edge cases. The 85% figure is a replication measure, not a universal correctness rate. This paper supplies the central causal bridge our product stands on: from “the theories are valid” (pillars 1–3) to “a language model can operationalize them reliably.” Without that evidence, our methodology would be a hope; with it, it is a testable architecture.