VISUAL HALLUCINATION IN A PAID AI SYSTEM: A Documented Case of Multimodal Confabulation by Gemini AI Plus
Kian Tik Go · 2026
This paper documents a case of severe visual hallucination — also termed multimodal confabulation — observed in Gemini AI Plus (paid tier) on 18 April 2026. The subject AI system was presented with a screenshot of a LinkedIn recruitment advertisement from Sc ….hidden, a low-….hidden, soliciting ….hidden fleet. Rather than accurately describing the image content, Gemini fabricated an entirely different reality: it identified the image as a St….hidden financial dashboard belonging to the user, and proceeded to generate specific, fictitious financial data — including gross sales figures in Indonesian Rupiah (IDR), percentage growth rates, transaction counts, and average revenue per customer — none of which were present in the original image. The hallucination persisted across two correction attempts. When challenged, Gemini apologised and produced a revised set of equally fabricated data, adjusting only the currency from IDR to USD while maintaining the false premise that the image depicted a Str….hidden. The incident raises significant concerns about the reliability of multimodal AI systems in paid commercial tiers, particularly regarding image grounding, confidence calibration, and the absence of mechanisms to prevent financial data fabrication. This case suggests a potential systemic issue in AI multimodal processing. From a humanity-centered standpoint, I consciously choose to elevate it into a public study — anchored in the academic domain — rather than treat it as an isolated user complaint. This report provides full documentation of the incident, verbatim transcription of the AI responses, evidence screenshots, impact analysis, and recommendations for AI developers and end users. It is submitted as an open academic record for the AI safety and hallucination research community.