Synthetic personas—LLM-based models that simulate human respondents for market research—are a bounded forecast technology, akin to weather prediction, that yields reliable insights only within carefully delimited conditions, according to Ishan Anand, chief AI officer at InsightSciences.ai. Anand argues that synthetic personas excel at capturing stated attitudes when richly grounded, but collapse into nonsense when treated as cheap proxies for human subjects. He identifies three crucial failure modes. In a pricing experiment, a simple prompt produced an inverted U‑shaped curve where willingness to purchase increased with price over part of the range, because the LLM used price as a proxy for latent confounders like expiration date or competing products. Anand notes that historical attempts like Simulmatics in the 1950s and 1960s failed, but today’s large language models provide “a new atomic unit of language that we can model against,” making simulation newly tractable if boundaries are respected.