Confident Confabulation: How Google AI Overviews Hallucinate in Response to Fabricated Entities
Year: 2026
Type: preprint
Abstract: Generative search features such as Google's AI Overviews increasingly mediate how the public encounters information, yet little is known about how—rather than how often—they fabricate content. This project investigates the phenomenology of hallucination: the rhetorical and epistemic moves a generative search system deploys when prompted with entities that do not exist. We assembled a corpus of 140 Google AI Overviews elicited by prompts naming deliberately fabricated entities across seven categories of stimulus (non-existent theories, fake scientific constructs, fake named effects, fake frameworks, fake public-health concepts, fake historical events, and fictitious scholars), in a balanced design of 7 categories × 2 prompts × 10 repetitions. Because every referent is known to be false by construction, any element of apparent substance in a response is necessarily generated rather than retrieved, allowing fabrication to be observed in isolation. This repository hosts the stimulus corpus, the coding instruments (manual classification scheme and behavioural codebook with protocol), the coded data, and analysis materials, in support of transparency and reuse.
Source: OSF Preprints (OSF Preprints)
Authors: Federico Germani, Giovanni Spitale
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7171272129