The AI digital asset management assistant: Testing GPT-4’s description and keyword tagging abilities on product imagery

Abstract

The process of tagging descriptive metadata to digital assets remains a time-consuming and largely manual process for digital asset management (DAM) system administrators. This paper explores the potential of OpenAI’s GPT-4, a large language model, to automate product descriptions, keyword tagging and alt-text. The research team developed six generative AI prompts that instruct GPT-4 to draft one-sentence descriptions and ten keywords for sample product images from six categories of household brands, namely: bicycles, food & beverage, home goods, office furniture, footwear and tools. Using an assessment framework that measures accuracy and precision, the team evaluated GPT-4’s performance by prompt and product category. GPT-4 demonstrated the highest accuracy when describing food & beverage images and the highest keyword precision when tagging footwear images. GPT-4 struggled to be accurate when images displayed low colour contrast or partially obstructed text. It also struggled when attempting to correctly identify gender, relationships and settings. However, GPT-4 showed surprising aptitude at identifying product materials like carbon fibre and species of wood. An analysis of prompts revealed that changes in persona, task description and specifications significantly influence accuracy and precision. The highest average accuracy score and highest average precision score among the prompts suggest that GPT-4 requires careful human oversight when generating keywords, product descriptions and alt-text for accessibility. Even so, it likely saves time for DAM administrators and professionals in marketing and e-commerce.