Researchers develop method to control verifiability of AI hallucinations in multimodal models, distinguishing between obvious and elusive false claims.

Study analyzes 4,470 human responses to AI-generated hallucinations from multimodal large language models (MLLMs)

Hallucinations categorized into two types: obvious (easily detected by humans) and elusive (difficult to verify or miss)

Proposes activation-space intervention method with separate probes to control verifiability of different hallucination types

Addresses gap in research on controlling hallucination properties based on application-specific security and usability requirements