Blogs›Microsoft›AI-powered Historical Document Analysis and Interaction
This post introduces PRISM2, a pathology foundation model trained on both tissue images and language from pathology reports. This approach aims to create a single model that can support a broader range of tasks by linking visual patterns in tissue with diagnostic language, moving beyond single-purpose AI systems. The model can work with images only or images and text, allowing interaction through prompts. This builds upon the concept of AI-powered analysis and interaction with specialized data, extending it to the domain of pathology.
8/4/2026
Introduces PRISM2, a pathology foundation model trained on paired pathology images and text from pathology reports. This model is designed to link visual findings with diagnostic language, enabling it to respond to prompts and support multiple pathology tasks without requiring task-specific model retraining. The model weights are publicly available for research.
7/22/2026
Introduces the Community Library Creator, a platform enabling communities to build curated datasets for AI image generation. Details the process of image selection, annotation, and AI model evaluation by community members. Highlights the technical implementation of community data ownership and consent management within the AI data pipeline.
7/1/2026
This post details the technical implementation of the 'Box 1' system, which uses AI to process and organize hundreds of thousands of historical documents for the Theodore Roosevelt Presidential Library. It describes how AI is used to reconstruct fragmented materials into searchable records and powers an AI avatar for interactive user experiences. The post also mentions the intention to release a paper documenting the technology and open-source the software.