“A Cost Analysis of Generative Language Models and Influence Operations,” posted to arXiv in August 2023 by Micah Musser of Georgetown’s Center for Security and Emerging Technology, tackles a practical question left open by earlier work: not whether language models can produce propaganda, but whether doing so is cheaper than hiring people.
The paper builds an economic model of an influence campaign and estimates that a language model needs only roughly 25 percent reliability, meaning about a quarter of its outputs are usable without heavy editing, to be cost-competitive with human writers, and that a highly reliable model could cut content-generation costs by as much as 70 percent. It also examines mitigations from a cost angle, arguing that monitoring API-accessible models has limited deterrent value once capable open-weight models exist, and that nation-states would generally not save money by training bespoke models purely for propaganda.
By quantifying the economics, the paper sharpened the debate beyond “AI could write fake content” toward when and for whom automation actually pays, which matters for predicting who adopts it and where defenses are worth deploying.
For a business reader, it is a clear example of treating an AI risk as an economic decision: the threat scales with cost savings and reliability, not just with raw capability.