Text-to-Image model vulnerable to grammatical phenomena

A recent study from the Universitat Rovira i Virgili, the University of Texas, and NYU aimed to evaluate DALL-E 2 by seeking for any deep vulnerabilities rather than just standard random error. OpenAI, the creators of the system, state: “DALL-E 2 is a new AI system that can create realistic images and art from a description in natural language”. DALL-E2 is a generative text-to-image model. People often seem to believe that AI systems can imitate the fundamentals of cognitive psychology, but this study shows otherwise. The researchers tested the model onto a series of grammatical phenomena such as binding principles, passives, word order, and structural ambiguity. The study concludes that given the failures they have observed on DALL-E 2 from the grammatical phenomena, the struggle of an AI system to perceive human language is significant.

https://arxiv.org/abs/2210.12889