The researchers created the “Faces in Things” dataset, a human-labeled collection of over 5,000 images where people commonly perceive illusory faces. They trained face-detection algorithms on this dataset and found that AI systems initially struggled to recognize these faces. However, performance improved significantly when the models were first trained to detect animal faces, suggesting a possible evolutionary link between recognizing animal features and pareidolic perception.

The team also identified a “Goldilocks Zone of Pareidolia”—a specific range of visual complexity where both humans and machines are most likely to detect faces. This work not only deepens our understanding of human and machine vision but also opens new avenues for improving AI perception by mimicking evolutionary pathways in visual recognition

The MIT SuperCloud and Lincoln Laboratory Supercomputing Center provided HPC resources for the researchers’ results.