The Evens Foundation has published a new essay exploring a fundamental flaw in the use of existing large language models (LLMs) to simulate human populations.

Written in collaboration with researchers from the Université Libre de Bruxelles (ULB) and the Vrije Universiteit Brussel (VUB), One Hand in a Furnace, One Leg in a Freezer: The Average Body Temperature Fallacy in AI argues that the way LLMs are trained causes them to flatten human opinion rather than reflect it. The result is what the authors call a "statistical ghost" that has significant implications for the development of a "synthetic citizenry" model.

The essay looks at the mechanisms behind this homogenisation, examines why common fixes fall short, and explores potential directions for building AI systems that maximise representation rather than minimise disagreement.

The findings are one output from an ongoing project exploring the concept of synthetic citizenry – the potential use of AI as a digital stand-in for human populations in polling and public insight.

"This is one of the biggest challenges in developing a reliable sandboxing tool. We must be honest about these limitations if we want to properly interrogate the models being developed to compete with or supplement our current methods for taking the temperature of the population," write the authors.

"We don’t need models that minimise disagreement, but ones that maximise representation. Because a “mean” AI is not just meaningless, it is potentially dangerous."

Read the full essay on the Evens Foundation's LinkedIn –>