Corporations are so eager to adopt generative AI (or so terrified of becoming irrelevant if they do not) that consumers may feel as though it’s being shoved down their throats. In fact, a new word has entered the everyday lexicon to serve as a label for dishonest or unwanted AI content: slop. This collectively negative sentiment toward generative AI is easy to justify upon taking a closer look at its societal byproducts.
As a computer science student who has conducted AI research at various scales, I recognize the potential that responsible AI holds. The way generative AI is being deployed today, however, is far from responsible. Models based on Transformer (the algorithm powering today’s generative text, image, and video models) become powerful only at massive scales, both in terms of model size and training data. But scaling up is also where the biggest problems emerge. The largest of large language models (LLMs) guzzle electricity and water at alarming rates; for example, training Meta’s Llama 3 LLMs consumed 22 million liters of water, as much as 164 Americans use in a year. Its training corpus contained tens of trillions of words, many of which were taken without permission from millions of webpages and thousands of books. Yet its immensity doesn’t prevent it from making mistakes. The flagship LLMs of OpenAI, Anthropic and Google averaged a score of 53% on a simple fill-in-the-blank quiz of questions that can be answered with a couple of Internet searches.
LLMs are the ultimate product of convenience culture: they’re designed to spit out whatever sentences you request, and truthfulness is relegated to an afterthought. Their massive energy usage means they rely almost exclusively on venture capital to cover their operating costs. This situation presents a rather dystopian argument that unfathomably wealthy people and corporations want to make libraries obsolete. Somehow, the reality is even worse: they want to make libraries illegal.
In recent years, two federal court cases have rescued AI at the expense of libraries and authors by manipulating the legal doctrine of “fair use.” In Hachette v. Internet Archive, book publishers successfully argued that digitizing books couldn’t be fair use because it counted as unauthorized copying, effectively banning libraries from lending books over the Internet and reducing online research options for people with limited accessibility. In Bartz v. Anthropic, the defendant received a summary judgment allowing them to use copyrighted books to train AI without the authors’ consent because it was a more “transformative” use of the copyrighted material. The plaintiffs won a $1.5 billion settlement from Anthropic, yet the company is still using the infringing AI models today.
For a technology that’s heralded as the future, generative AI seems unable to escape the dishonesty of its past. It’s created by corporations that argue in court that the only way to make their product is by stealing as much data as possible. It’s used by the same techbros who used to believe that the future would be cryptocurrency or NFTs or Web3, all technologies that became widely known for the scams that utilized them. And it’s sold by people who make apps that use existing AI models for their entire functionality, with sales pitches such as “We want to cheat on everything.”
In the third century BCE, the library of Alexandria was the greatest collection of information in the world. By the time it famously burned, it had already fallen into disrepair from centuries of abuse. In 1989, the World Wide Web was born with the explicit purpose of sharing research among scientists. After 35 years, it’s grown into our modern library of Alexandria. And by allowing AI companies to cripple our libraries and market their slop as a replacement for human knowledge, we’re laying the foundation for the Internet, and all libraries, to burn just the same.
