Scribd uses Gemini Enterprise batch inference to classify over 400 million documents

Scribd Classifies Content with Gemini Enterprise

Scribd, Inc. used Gemini Enterprise's batch prediction feature to perform trust and safety classification on its entire user-generated content corpus. This initiative involved over 400 million documents and 12 billion pages from Scribd and Slideshare platforms.

The classification process was completed within months, with Google Cloud scaling batch throughput to meet the project timeline. This allowed Scribd to efficiently review a large volume of content for compliance with community guidelines.

Technical Implementation and Benefits

Gemini's native PDF understanding was a key factor, enabling more than 99% of the corpus to be processed directly without requiring optical character recognition (OCR), rendering, or screenshotting pipelines. This eliminated the need for Scribd to build additional processing infrastructure.

The use of Gemini Enterprise's batch prediction also offered a 50% discount compared to interactive pricing, making large-scale LLM classification economically viable for Scribd's extensive content library.

Addressing Trust and Safety at Scale

Scribd, which includes products like Scribd, Slideshare, Everand, and Fable, manages hundreds of millions of user-uploaded documents. Maintaining trust and safety across this vast collection requires balancing content access with community protection.

Traditional methods for content moderation often involve specialized detection models for each policy area, which can be resource-intensive. Scribd evaluated various off-the-shelf tools and open models but found they did not meet the required quality or scale for their 400-million-document backfill.