Image and Alt Text Dataset for "Identity-Based Description Bias in AI-Generated Alternative Text"

This repository contains the image and corresponding alt text data for the paper "Identity-Based Description Bias in AI-Generated Alternative Text," which is being reviewed during the Jan. 2026 ARR cycle. In the .zip file are the 2,000 images, generated by GPT-image-1. The three .csv files contain alt text generated by three respective models to describe the images. alttext_gpt5.csv has alt texts generated by OpenAI's GPT-5. The alt texts in alttext_claude.csv and alttext_gemini.csv were generated by Anthropic's Claude Sonnet 4.5 and Google's Gemini-2.5-Pro, respectively. The leftmost column in each .csv file shows the name of the corresponding image plus the version (a or b). 

The code to accompany this data can be found at the project's Anonymous GitHub.

本仓库包含论文《AI生成替代文本（alt text）中的基于身份的描述偏差》（Identity-Based Description Bias in AI-Generated Alternative Text）的图像及对应替代文本数据，该论文目前处于2026年1月ARR评审周期中。压缩包内含2000张由GPT-image-1生成的图像。三个CSV文件分别对应三款模型为上述图像生成的替代文本：其中alttext_gpt5.csv包含OpenAI的GPT-5生成的替代文本；alttext_claude.csv与alttext_gemini.csv分别包含Anthropic的Claude Sonnet 4.5与Google的Gemini-2.5-Pro生成的替代文本。每个CSV文件的最左列标注了对应图像的名称及其版本（a或b）。本数据集的配套代码可在本项目的匿名GitHub仓库中获取。