Boston Review: “Questions of AI authorship and ownership can be divided into two broad types. One concerns the vast troves of human-authored material fed into AI models as part of their “training” (the process by which their algorithms “learn” from data). The other concerns ownership of what AIs produce. Call these, respectively, the input and output problems. So far, attention—and lawsuits—have clustered around the input problem. The basic business model for LLMs relies on the mass appropriation of human-written text, and there simply isn’t anywhere near enough in the public domain. OpenAI hasn’t been very forthcoming about its training data, but GPT-4 was reportedly trained on around thirteen trillion “tokens,” roughly the equivalent of ten trillion words. This text is drawn in large part from online repositories known as “crawls,” which scrape the internet for troves of text from news sites, forums, and other sources. Fully aware that vast data scraping is legally untested—to say the least—developers charged ahead anyway, resigning themselves to litigating the issue in retrospect. Lawyer Peter Schoppert has called the training of LLMs without permission the industry’s “original sin”—to be added, we might say, to the technology’s mind-boggling consumption of energy and water in an overheating planet. (In September, Bloomberg reported that plans for new gas-fired power plants have exploded as energy companies are “racing to meet a surge in demand from power-hungry AI data centers.”) The scale of the prize is vast: intellectual property accounts for some 90 percent of recent U.S. economic growth. Indeed, crawls contain enormous amounts of copyrighted information; the Common Crawl alone, a standard repository maintained by a nonprofit and used to train many LLMs, contains most of b-ok.org, a huge repository of pirated ebooks that was shut down by the FBI in 2022. The work of many living human authors was on another crawl, called Books3, which Meta used to train LLaMA. Novelist Richard Flanagan said that this training made him feel “as if my soul had been strip mined and I was powerless to stop it.” A number of authors, including Junot Díaz, Ta-Nehisi Coates, and Sarah Silverman, sued OpenAI in 2023 for the unauthorized use of their work for training, though the suit was partially dismissed early this year. Meanwhile, the New York Times is in ongoing litigation against OpenAI and Microsoft for using its content to train chatbots that, it claims, are now its competitors. As of this writing, AI companies have largely responded to lawsuits with defensiveness and evasion, refusing in most cases even to divulge what exact corpora of text their models are trained on. Some newspapers, less sure they can beat the AI companies, have opted to join them: the Financial Times, for one, minted a “strategic partnership” with OpenAI in April, while in July Perplexity launched a revenue-sharing “publisher’s program” that now counts Time, Fortune, Texas Tribune, and WordPress.com among its partners. At the heart of these disputes, the input problem asks: Is it fair to train the LLMs on all that copyrighted text without remunerating the humans who produced it? The answer you’re likely to give depends on how you think about LLMs…”