Images Posted on Internet are Training AI Algorithms
A computer scientist told the National Press Club's photojournalism group recently that the billions of photos posted on the internet are training algorithms for generating images made with artificial intelligence. Robert Pless, computer science professor and a specialist in computer vision at George Washington University, described the algorithms and their development by leading generative AI image software at a meeting of the Club's Photo Team July 12.
Pless told the attendees that generative AI uses algorithms called large language models trained by massive text and image collections to produce new content on demand from user prompts and specifications. Those algorithms employ predictive formulas to produce new material based on the latest examples in their data sets. With images, generative AI software like DALL-E (pronounced "dolly") and Midjourney trains their algorithms by matching images to captions. DALL-E is offered by OpenAI, the same company that makes ChatGPT, the now well-known text generative AI software.
To train those algorithms, Pless said, their developers scour the Internet for any publicly available images and captions, such as in Wikipedia. He noted that DALL-E was initially trained with 400 million images matched to text captions, while Midjourney's algorithm is now based on 1.2 billion image-text pairs. Pless displayed examples from the 90 Miles project, of realistic yet fake images made with generative AI of refugees crossing from Cuba to Key West.
Pless also described efforts to build traceability and transparency in image software, notably an open standard from the Coalition for Content Provenance and Authenticity, or C2PA, to certify the source and change-history of images and other media.
In the discussion following the talk, participants noted that Adobe, maker of widely used professional photo software, routinely takes user images to train its AI algorithms unless users choose to opt out.
An audio recording of Pless's talk and his presentation slides are available online.
This article was published by The National Press Club.
