AI vs the inkblot test: Here's what it saw

The Rorschach inkblot test, developed in 1921 by Swiss psychiatrist Hermann Rorschach, has long been controversial in the psychological community. Initially designed to diagnose disordered thinking in patients, it evolved into a general personality assessment tool. However, critics argue that the test lacks scientific validity and reliability, citing issues such as cultural bias and subjective interpretation. Despite these concerns, the inkblot test remains an interesting exploration of human perception and imagination. 
‍

But how do AI models interpret inkblot images?
‍

ChatGPT's literal approach

This week, BBC Future put ChatGPT to the test to see what it interpreted from some popular inkblot images.

When faced with an image often identified as either a bat, butterfly, or moth, ChatGPT responded, "This image is a Rorschach inkblot, often used in psychological assessments to explore perception and interpretation. It is designed to be ambiguous so that each person might see something different depending on their experiences, emotions, and imagination. For me, it resembles something symmetrical, possibly two animals or figures facing each other, or a single entity with wings outstretched."

However, ChatGPT provided inconsistent interpretations when shown the same image multiple times, unlike humans who typically maintain consistent responses. Chandril Ghosh, lecturer in psychology at the University of Kent, explained, "A human would typically stick to their previous answer because personal experiences and emotions influence their responses. In contrast, ChatGPT generates responses based on its dataset."
‍

Norman's "psychopathic" approach

For April Fools back in 2018, researchers at the Massachusetts Institute of Technology (MIT) revealed Norman, an AI model they claimed to be the "world's first psychopath AI." Though some of their writings about Norman were pranks, Norman itself was real.

Calling an AI model a psychopath is stretching the clinical definition of the condition, but by exposing Norman to dark material on Reddit during its early training, the researchers created something quite disturbing.

As reported at the time by Live Science, Norman's abnormality was demonstrated with the inkblot test, during which it reported seeing the patterns as "man gets pulled into dough machine" and "man is shot dumped from car".

For comparison, they tested other models with the same images and received much more innocent responses, such as "a black-and-white photo of a small bird" and "an airplane flying through the air with smoke coming from it".
‍

The importance of training data

Speaking on how ChatGPT responded to the inkblot test, Ieva Kubiliute, a London-based psychologist, told BBC Future, "I believe it mainly identifies patterns, shapes, and textures within the blots, and then compares these features to a vast dataset of human responses to generate its interpretation of what it sees in the inkblots."

The stark contrast between ChatGPT's and Norman's interpretations underscores the critical role that training data plays in shaping AI behaviour and perceptions. Norman was purposefully trained on a biased dataset, exposing it to an unbalanced amount of dark material, whereas ChatGPT (though still capable of bias and hallucinations) was trained on more extensive and diverse data.

‍

Enhance and secure your team’s use of GenAI with Narus.