Is it biased?

A model learns patterns from text people wrote, so it can learn unfair patterns too, without anyone meaning to put them there. Chapter 2 explains how training works. The U.S. National Institute of Standards and Technology (NIST) puts the principle plainly: “Bias is neither new nor unique to AI and it is not possible to achieve zero risk of bias in an AI system.”¹ NIST names three sources. Systemic bias comes from the habits of institutions. Statistical bias comes from data that doesn’t represent the people a system serves. Human bias comes from errors in people’s own thinking, in the decisions they make about the system and in how they read what it produces.¹

An early, simple case shows the mechanism. In 2016 researchers found that word patterns learned from Google News articles completed “man is to computer programmer as woman is to x” with “homemaker.”² One might have hoped for little bias, they wrote, because many of the writers were professional journalists.² The data simply held the patterns of the world that wrote it.

“If it says kind things about a group, it isn’t biased.” A 2024 study in Nature tested that belief. Researchers gave language models the same content written in African American English or in Standardized American English, never mentioning race, and asked the models about the writer. What the models said about African Americans when asked directly was more positive than what they implied from dialect alone.³ In invented trial scenarios, they convicted the writer of African American English more often: 68.7 percent against 62.1 percent.³ Training with human feedback, the authors wrote, “obscures the racism on the surface, but the racial stereotypes remain unaffected on a deeper level.”³ The newest model tested was a version of GPT-4, and newer models may score differently.³

People disagree about how far bias can be fixed. The 2016 paper proposed a method to remove the bias it found.² NIST holds that zero risk isn’t possible,¹ and the 2024 study found that one common fix hid bias rather than removing it.³

In hiring, the question has reached court. In Mobley v. Workday, job applicants allege that Workday’s AI-based screening system discriminated by race, age and disability. In May 2025 a federal trial court in California let an age claim go forward as a collective action for applicants aged 40 and over. The judge called that step preliminary. It rests on what the applicants allege, and it is not a finding that anyone discriminated.⁴ In September 2026 the applicants asked the court to certify a class action, with a hearing set for March 2027.⁵ As of that filing, the case was still in the trial court. Chapter 14 explains why decisions about people are the place to keep a person in charge.

Who owns what it makes?

In the United States, copyright protects work that a person creates. What a machine produces on its own isn’t protected, but a person’s own creative choices still can be. In March 2025 a federal appeals court in Washington, D.C., ruled in a case brought by the computer scientist Stephen Thaler, whose AI system, the Creativity Machine, had made a picture. The court held that the Creativity Machine could not be the author of a copyrighted work, because the law “requires all eligible work to be authored in the first instance by a human being.”⁶ The rule, the court added, requires only that the author be a human, including “the person who created, operated, or used artificial intelligence.”⁶ In March 2026 the Supreme Court declined to hear the case.⁷

The U.S. Copyright Office, which registers copyrights, set out its own view in January 2025. With the technology then generally available, it said, “prompts do not alone provide sufficient control” to make the person who typed them an author. But a person’s creative selection, arrangement or changes to what the tool produced can be protected, judged case by case.⁸ That is an agency’s view, not a court’s.

“Anything made with AI can’t be owned.” The record says something narrower. The parts a machine made alone are not protected. The parts a person wrote, chose or arranged may be.

Patents follow the same line. In 2022 the federal appeals court that hears patent cases ruled, in another case brought by Thaler, that “only a natural person can be an inventor, so AI cannot be.”⁹ It left open whether people who invent with AI’s help can get patents.⁹ The U.S. Patent and Trademark Office’s current guidance, from November 2025, calls AI systems “instruments used by human inventors,” like lab equipment or databases. The question is whether a person conceived the invention.¹⁰ The same office warned in 2024 that “simply relying on the accuracy of an AI tool is not a reasonable inquiry,” and in its notes it cited the case of the lawyer in Chapter 1.¹¹ Chapter 13 covers telling readers how AI helped with work you share.