ryanmio/kenya-nigeria-prep-hallucination-audit

# AI Hallucination Audit: Citation Mismatches in HIV Prevention Research 00158-4/abstract) This repository contains reproducible experiments demonstrating AI hallucination patterns that explain citation mismatches in a published HIV prevention research Comment. We test two disputed claims that appear to originate from the Audere sponsor white paper but are instead attributed to other academic sources. > **Terminology Note**: Throughout this repository we use "AI hallucination" as a shorthand, even though the behavior we investigate is more nuanced than inventing facts out of thin air. In this case the models conflate two genuine statements of the Audere white paper— for example, AI in Kenya and tele-PrEP in Nigeria—and merge them into a single, misleading claim. The output is therefore better described as *source-blend hallucination* or *AI conflation*, but we keep the familiar term for searchability and to align with the broader literature on LLM errors. > **Transparency Note**: This investigation follows an open, evidence-based approach. Our investigation log documents our evolving understanding, uncertainties, and commitment to following evidence wherever it leads. ## Background We identified apparent citation mismatches in a published Lancet HIV Comment where key sentences do not match their cited sources. Two claims stood out: 1. **Kenya + Nigeria tele-PrEP claim**: *"In Kenya and Nigeria, AI powers telemedicine-based PrEP services that eliminate the need for frequent clinic visits."* 2. **Supply forecasting claim**: *"Predictive models could further enhance these programmes by helping providers identify individuals at greatest risk of HIV acquisition, forecast supply needs, and target outreach efforts more effectively."* Both sentences cite peer-reviewed academic papers that do not contain these claims. However, the Audere sponsor white paper (which is cited elsewhere in the Comment) contains both phrases. Our experiments test whether AI summarization …