Unravelling Hallucinations

A Research Assistant’s Experience of LLM-Generated Historical Data

Following on from Jakob’s post, ‘Persuasively Wrong’, about the processes of working with LLMs to generate historical data, I give some insight here as to the results for our Instability Events project - for which we used LLMs to gather data on crisis event variables (such as assassinations, civil wars, or rural uprisings) from 571 polities of varying complexities, spanning from the 31st century BCE to the 21st century CE, and across broad geographical regions — and the reality of the follow-up work needed by humans to check the accuracy of AI output.