Was it fair to train it on other people’s work?

Models learn from enormous amounts of text, much of it written by people who were never asked (Chapter 2). In the United States, whether that copying is legal turns largely on fair use, the part of copyright law that weighs, case by case, what was copied, why, and what the use does to the market for the original. The Copyright Office, in a May 2025 report released in draft form, expects that “some uses of copyrighted works for generative AI training will qualify as fair use, and some will not.”¹ So far the courts have decided cases one at a time, and they have not agreed on the reasoning.

Two federal trial judges in California ruled two days apart in June 2025. In Bartz v. Anthropic, Judge William Alsup found that using books to train the company’s models was “exceedingly transformative” and a fair use. He sent the claims about the company’s pirated library copies to trial.² The case then settled. In July 2026 the court gave final approval to a $1.5 billion settlement fund, an estimated $3,000 or so for each book.³ A settlement is not a ruling that training is legal or illegal.

In Kadrey v. Meta, Judge Vince Chhabria ruled for Meta. He wrote that his ruling “does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful,” only that “these plaintiffs made the wrong arguments.”⁴ The two judges openly disagreed on one point. Judge Alsup compared training to “training schoolchildren to write well.”² Judge Chhabria answered that “using books to teach children to write is not remotely like using books to create a product” that one person could use to turn out countless competing works.⁴

An appeals court has ruled in a related case, involving an older kind of AI. In September 2026 the federal appeals court in Philadelphia held that Ross Intelligence’s use of Thomson Reuters’s headnotes, notes that explain the points of law in court decisions, to build a legal-research platform was not fair use. It stressed that “ROSS’s AI was not a generative AI.”⁵ “Courts have ruled that AI training is legal.” As of October 2026, that headline overstates the record. There are two trial rulings for AI companies on training, with different reasoning; one settlement; and one appeals ruling against fair use, about a different kind of tool.

Will it take my job?

Two words help here. Exposure asks whether AI could do some of a job’s tasks. Displacement asks whether people actually lost work. They are different questions, and headlines often blur them. The U.S. Bureau of Labor Statistics, which makes the official job projections, notes that when new technology has displaced jobs in the past, “it tends to take longer than technologists typically expect,” and that many affected occupations still grew.⁶

A 2023 study estimated that about 80 percent of U.S. workers could have at least a tenth of their tasks affected by language models. Its authors said plainly, “We do not make predictions about the development or adoption timeline.”⁷ That is exposure. “A study found AI will replace most jobs” mistakes it for displacement.

What has happened so far is early and mixed. Economists at Stanford, using payroll data through June 2026, found “no evidence of widespread, economy-wide job displacement.” But employment of workers aged 22 to 25 in AI-exposed jobs stood 19 percent below where it would have been had it kept pace with their peers in less-exposed jobs. The gap came mostly from less hiring, not from layoffs. The authors call these “early, descriptive indicators,” not proof of cause.⁸ The Budget Lab at Yale, looking at the whole U.S. job market, found no “discernible disruption” in the 33 months after ChatGPT’s release.⁹ The two don’t strictly conflict: neither finds economy-wide displacement, and they differ over a smaller group.

Forecasts deserve the same care. For 2004 to 2014, as digital cameras spread, the BLS projected that photo-processing jobs would fall 23.6 percent. They fell from 86,300 that year to 28,800 in 2014. For truck drivers it made no such change, despite talk in the 2010s of self-driving trucks.⁶ Chapter 11 and Chapter 14 are the practical side: what the tool does well, and which calls stay yours.