Researchers from Google DeepMind and Stanford University found that a red heart attached to a programming prompt could consistently push two models toward recursive code. With that image, Claude Opus 4.7 and GPT-5.4 produced recursive solutions in 100% of cases, compared with 0% without an image, while a yellow checkmark raised Gemini 3.1 Pro’s recursion rate to 90%. The team evaluated 288 shape-and-color combinations and found a distinct steering profile for each model. These preliminary results show that visual elements unrelated to the task can alter outputs without the models’ explanations acknowledging the influence, creating audit challenges and potential consequences for applications that use screenshots or diagrams.
Source: Read the original article
