IC-763CLIP and OpenCLIP fall short of human discriminative accuracy on vision tasks, with performance dropping substantially under hard negatives

Peter West, Ximing Lu, Nouha Dziri, Faeze Brahman, Linjie Li, Jena D. Hwang, Liwei Jiang, Jillian Fisher, Abhilasha Ravichander, Khyathi Chandu, Benjamin Newman, Pang Wei Koh, Allyson Ettinger, Yejin Choi

SourceThe Generative AI Paradox: “What It Can Create, It May Not Understand”

The paper evaluates multiple CLIP and OpenCLIP variants in a discriminative setting where they must select the correct Midjourney-generated image matching a text caption from two candidates. All variants fall below human accuracy across COCO, PaintSkill, T2ICompBench, and DrawBench. The gap widens with difficulty: OpenCLIP performance drops from 100% with easy negatives to approximately 69% with hard negatives, while humans maintain near-perfect accuracy in both conditions.

Evidence
correlational
Key metric
clip-vit-large-patch14: 85.58% (COCO), 81.95% (PaintSkill), 84.50% (T2ICompBench), 78.57% (DrawBench) vs human 92.86%, 99.30%, 97.00%, 100.00%; openclip clip-vit-bigg-14-laion2b-39b-b160k: 81.73%, 85.28%, 84.50%, 84.53%; model performance drops from 100% to ~69% with hard negatives
Caveat
Human generation performance is assumed to be lower than Midjourney's rather than directly measured, as it is not practical to ask humans to produce detailed images. All results are zero-shot.
Model
CLIP / CLIP-ViT (LC), OpenCLIP
Concepts
Failure mode
Datasets
MS COCO / COCO / COCO 2014 / COCO 2017 / COCO 20k / COCO-it / COCO-wl [eval], T2ICompBench [eval], DrawBench [eval]
Methods
Zero-shot prompting [primary]
Related work
Li et al. 2023a (Your Diffusion Model is Secretly a Zero-Shot Classifier) [context]
Related findings
IC-762, IC-764, IC-765
Extraction
automatic-extraction