A test of 29 image models found that OCR-based automated scoring misjudges the text in AI-generated images. Of 205 lines manually reviewed, 44, or 21.46%, did not match the automated judgments. In 41 cases, OCR incorrectly marked correct text as incorrect, while some models refused to generate images because the prompts were too long. The study recommends using OCR to screen out obvious failures and adding manual spot checks before publishing rankings of image models.
Source:Technology Org · Read the original article
Leave a Reply