Did I see it or think it? Probing reality monitoring in vision-language models with concept injection

The researchers tested whether vision-language models can distinguish internally injected concepts from content actually seen in images. Among six models, only Gemma-3-27B could relatively stably identify injected concepts (identification rate 35%, versus 3% with steering alone), and it almost never produced false positives. This corresponds to the reality monitoring problem in cognitive psychology. The test paradigm varies images and injected concepts independently; in each round the model is asked to report what it saw, whether any thought was injected and its content, and is scored simultaneously on concept identification and source attribution. Even so, Gemma-3-27B's reported sources were unstable: strong injections or injections via the visual pathway were treated as seen content, and when a visible object appeared in the image, correct thought reports almost disappeared; in ordinary image description tasks, injected objects appeared in up to 87% of real photo descriptions. The paper accordingly argues that evaluating multimodal self-reports requires checking both content and the reported source.