A neuro-computational approximation of the qualities of mental images

Mental images are challenging to study, given that our conscious experience is notoriously hard to access. The currently prevalent introspective methods are inherently subjective and can thus only provide limited access to their qualities. Here, we developed a neuro-computational approach that approximates and assesses the properties of mental images without the need for introspection. To enable this approach, we collected a large-scale EEG dataset (10 participants, 10 sessions each, 43,200 trials total) of participants imagining 16 scenes based on text prompts. We employed AI image generation to create candidate image sets that approximate the content of mental images (based on the imagined text prompts), computationally simulated visual cortex responses to these images and then assessed their representational alignment with rhythmic EEG responses during imagery. In line with previous reports, mid- to high-level features of the AI-generated candidate images yielded reliable alignment with human alpha activity. By manipulating the qualities of the candidate images, we then tested which qualities predisposed higher representational alignment with cortical imagery representations. We found an increased representational alignment for spatially blurred and low contrast images, providing evidence for the prevalent notion of a reduced sensory quality of mental images. We further found that mental imagery may be characterized by a psychedelic image style, which envelops the images in visual flows that distort the image proportions. These results show that our approach can objectively capture qualities of mental images without the need of introspection, providing a hypothesis-based alternative to emerging reconstruction approaches.