Researchers are increasingly using language models to build synthetic survey samples, sometimes called silicon samples. A new arXiv preprint focuses on a newer class of these models: decision models, which return a probability for each point of a rating scale rather than generating text. The paper is described as a psychometric audit of such models when used as synthetic survey respondents.

The title of the preprint, "Probable but Not Valid," captures the central concern: decision models may produce outputs that look probable or reasonable, yet fail standard psychometric validity checks. The abstract introduces the practice of silicon samples and the decision-model architecture, but the full audit details are not included in the announcement.

Because this is a single preprint, its conclusions should be treated as preliminary rather than settled findings. The key takeaway is that plausibility of a model's probability outputs does not by itself establish that the model is a valid stand-in for human survey respondents.