A Back-of-the-Envelope Model for "Confident Hallucinations" in ChatGPT Usage

This short note turns a qualitative concern—that a widely used generative AI system can produce persuasive but incorrect statements—into a simple scaling model. We distinguish between (i) generated confident hallucinations and (ii) those that are propagated into durable artifacts (reports, code, policies). We avoid hard-coding uncertain usage totals (e.g., weekly active users) and instead express results in terms of parameters that can be measured or scenario-assumed. The result is not a measurement of error rates; it is a framework meant to clarify how volumes can scale.