Ernad Mostaque is the founder of Stable Diffusion, the image generator. He was previously a fund manager, and later turned into a tech disruptor. Stability AI is the start up behind this art generator tool.
The tool has thrown open the question of copying and originality. AI companies have the cover of the ‘fair use’ doctrine for content generation. In fact, they reap without sowing.
AI generators are built upon the backgrounds of copyrighted work. They produce copies with varying degrees of variations and sophistication, taking advantage of a legal loophole.
First, artists feed this ever hungry engine new art feed, which is more and more in their own style. However, the origination is de novo, copyright is not applicable. It is not a copy — it is stylistic inspiration.
The data set for Stable Diffusion is called LAION 5b. It was a collection of 6 billion images from internet using a practice called data scraping. LAION could scour the whole internet as it deems itself a not-profit organisation devoted to academic research. It was partly funded by Stability AI, the creator of Stable Diffusion, a separate entity. Stability AI first used its non-profit research arm to create AI generators through Stable Diffusion. It then commercialised it in a new model called DreamStudio.
AI-generated images from Stable Diffusion cannot be distinguished stylistically from those of the original artist. AI works out patterns, styles and relationships by examining billions of images on the internet.
Stable Diffusion have 10 million plus daily users. Other text-to-image generators which have emerged on the scene are Midjourney and DALL-E.
These AI products are built on collection of images known as datasets. They make a detailed mapping of their contents. They discover the connections among images and between images and text.
The tool does not do anything on its own. It takes into account the intention of the user. It is an algorithm that learns just like a human brain, only it is far more faster, powerful and untiring.
There are AI-generated images of historical or mythological characters which may not be realistic. At most, one can say these are realistic-looking fake images.
OpenAI launched Dall-E in 2021, and Dall-E 2.0 in the second half of 2022.
Ai-generated images can have plasticky appearance and semantic consistencies. However, as technology advances, the quality of the images will improve too.
In February 2023, the US Copyright office ruled that Midjourney-generated images of Zarya of the Dawn, a comic book of 18 pages by an AI expert Kristina Kashtanova are not protectable under current copyright laws as they ‘are not product of human authorship’. The ruling has great implications for the artists.
Getty accused Stability AI of using 12 million plus Getty Photos along with captions and metadata to train the algorithm behind Stable Diffusion tool. It is illegal. The newly generated images could be mistaken as images created by the artist.
There could be face-swapping software, and it could be used to supplement human models.
The line between true and fake has already blurred. The technology can also create trust deficit.
The US cCommerce Department is looking for feedback from the public on how to create AI accountability standards. These will guide the policy makers.