Optical Character Recognition (OCR) is the process of converting document images into text. Verification of OCR text (or confirmation) for low-resolution inputs remains challenging due to semantic and structural confusion between the input image and text predictions. The absence of labels in unsupervised settings further increases the complexity. We propose an unsupervised technique as a baseline for the verification of a hundred low-resolution images and their predictions from Tesseract OCR. To be precise, we first synthesize the images of OCR predictions using a fixed font and style. Next, we feed the low-resolution (real) input images and the corresponding (predictions’) synthetic images to ResNet50 to extract word-level features. We finally compare the obtained features for verification using cosine similarity. To benchmark, we also present the results of Vision Language Models (VLMs) like ChatGPT, Claude, and Gemini by prompting them to compare low-resolution images to the predictions’ synthetic images. We also present an ensembling approach and compare it with human evaluation. The work contributes to a broader understanding of low-resolution OCR in scenarios with a scarcity of human annotators.