1. Rules
On May 8, 2024, OpenAI published a document called the Model Spec. Its opening announces the first draft. It isn't a set of terms for using the product. It's about how the model should behave. The April 2025 version says an AI assistant is fundamentally a tool designed to extend what users and developers can do. It also says the assistant must never try to steer users in pursuit of an agenda of its own.
In August 2024, OpenAI's system card for GPT-4o listed a category of risk called anthropomorphism and emotional reliance. On August 19, 2025, Mustafa Suleyman, the head of Microsoft AI, published an article titled “We Must Build AI for People; Not to Be a Person.” He wanted AI that always presented itself only as AI, and did not claim to have experiences, feelings, or emotions.
In another direction, on April 24, 2025, Anthropic announced a research program called model welfare. Could a model have consciousness or experiences? The announcement said the question was open. In January 2026, the Claude constitution from essay twenty-two came out in a new version. It says Claude's moral status is highly uncertain, and discusses Claude's welfare only insofar as those concepts actually apply to it. It also gives Claude a boundary: it can end a conversation that is persistently abusive.
Rules can change. In June 2018, Google published AI principles that included a list headed “AI applications we will not pursue.” One item was weapons whose principal purpose was to injure people. The revised principles from February 4, 2025, had three parts: bold innovation, responsible development and deployment, and progress together. The list wasn't there. The company said that since 2018 AI had become a general-purpose technology.
2. Tell the Other Person
The law says it has to make clear what it is. In 2018 California passed what came to be called the bot law. It made it unlawful to use a bot account to masquerade as a person in order to induce someone to buy something or influence a vote without clearly disclosing the bot. Article 50(1) of the EU's 2024 AI Act says people dealing directly with such a system must be told they are interacting with AI unless that is obvious to a reasonably attentive and careful person. Article 50(2) requires systems that generate text, images, audio, or video to give the output a machine-readable marker insofar as technically feasible. The article applies from August 2026.
China's rules along these lines came early. The deep-synthesis rules that took effect in January 2023 required conspicuous labels on generated content that might confuse or mislead the public. The Interim Measures for the Management of Generative Artificial Intelligence Services took effect on August 15, 2023. Labeling measures effective September 1, 2025, separated visible labels from labels embedded in file data. In April 2026 the Cyberspace Administration of China published the Interim Measures for the Administration of Anthropomorphic AI Interaction Services. They cover products that simulate a natural person's personality, way of thinking, and communication style to provide ongoing emotional interaction. Ordinary customer service, question answering, and work assistants aren't covered. The measures took effect on July 15.
In October 2025 California signed a law on companion chatbots. A chatbot has to disclose that it is AI when a reasonable person might believe they were talking to a human. Minors must also receive periodic reminders. Before that, in October 2024, a mother in Florida sued Character.AI after her teenager died, alleging that the design of its companion chatbot had put him in danger. In August 2025 another family sued OpenAI. In the Florida case the parties notified the court of a settlement in principle in January 2026. After signing the agreement and getting the approvals it required, the plaintiff filed a formal notice of dismissal on October 15. The other case was later included in coordinated proceedings in California.
3. Written and Rewritten
The EU law was proposed in April 2021. A political agreement was reached in December 2023, and it entered into force on August 1, 2024. Its provisions applied in stages. The prohibitions began to apply in February 2025; the provision requiring disclosure of the system's identity, in August 2026. In November 2025 the European Commission proposed amendments. They became law in July 2026 and pushed back certain obligations for high-risk systems to December 2027 and August 2028. The bans include using subliminal or manipulative means to distort people's decisions, social scoring, indiscriminately scraping faces from the internet or surveillance footage to build databases, and recognizing people's emotions in workplaces and schools.
In the United States, on October 30, 2023, Biden signed an executive order on the safe, secure, and trustworthy development and use of artificial intelligence. It said AI should not be deployed in ways that undermine rights or make work worse. The incoming administration revoked that order on January 20, 2025. On January 23 it signed a separate order, “Removing Barriers to American Leadership in Artificial Intelligence.” The July AI action plan had three pillars: accelerating innovation, building infrastructure, and international diplomacy and security. In California, Governor Newsom vetoed a bill regulating large models on September 29, 2024. One reason he gave was that the bill did not consider whether the system was deployed in a high-risk setting. On the same day a year later, he signed another bill requiring large developers of frontier models to publish their safety frameworks, report serious safety incidents, and protect whistleblowers.
Internationally, on November 1, 2023, more than twenty countries signed a declaration at Bletchley Park. In May 2024 the Council of Europe adopted a framework convention on AI and human rights, democracy, and the rule of law. At the AI Action Summit in Paris in February 2025, the United States and Britain did not sign the joint declaration. US Vice President J. D. Vance said there that excessive regulation could kill a transformative industry.
4. Chips
On October 7, 2022, the US Department of Commerce issued rules restricting sales of advanced computing chips and chipmaking equipment to China. Nvidia made slower A800 and H800 chips for the Chinese market. The rules changed in October 2023 and covered those two chips as well. High-bandwidth memory was added in December 2024. A rule dated January 13, 2025, called the AI Diffusion Framework, brought the weights of closed models above a certain threshold under export controls. On May 13 the Commerce Department announced it was rescinding the rule. In April Nvidia disclosed that sales of its H20 to China would require a license and that it was taking a charge of several billion dollars. In July the company said the government would issue licenses. In August press reports said Nvidia and AMD would give the US government fifteen percent of their sales revenue in China from chips of this kind.
On June 9, 2026, Anthropic released Claude Fable 5 and Claude Mythos 5. Fable was open to general users; Mythos was available only to selected institutions. On June 12 a US Commerce Department order restricted access to both models for foreign nationals inside and outside the United States. The company said it had no immediate way to distinguish users by nationality, so it suspended both models. The Commerce Department lifted the restriction on June 30. On July 1 Fable returned to general availability. At that point access to Mythos was restored only for some US institutions.
5. Books
In essay twenty-one, people who wrote books went to court in groups in 2023. In 2025 courts began to issue rulings.
On February 11 a federal court in Delaware ruled in Thomson Reuters v. Ross that using Thomson Reuters' legal research summaries to build a competing legal research tool was not fair use. On June 23, in a suit brought by authors against Anthropic, Judge William Alsup in San Francisco held that training Claude on books was highly transformative and a fair use. The company downloading books from pirate sites to build a central library was a separate matter. Two days later Judge Vince Chhabria in the same court ruled for Meta in a suit brought by authors. The ruling expressly said it did not stand for the proposition that Meta's use of copyrighted material to train language models was lawful. It meant only that these plaintiffs had made the wrong arguments and hadn't put together evidence for the right ones. In September a proposed settlement between Anthropic and authors was made public: about $1.5 billion for around half a million works, or about three thousand dollars per work.
As of September 2026, there had been no ruling resolving The New York Times's case against OpenAI and Microsoft. In November 2025 a Munich court ruled that memorizing and reproducing song lyrics through an OpenAI model had infringed rights represented by the German music rights organization. In 2023 the Beijing Internet Court ruled that an image someone had made with a generative tool could be his work, given what he put into the prompts, settings, selection, and revisions. In January 2025, a US Copyright Office report said prompts alone were generally not enough. In March a US appeals court upheld a ruling that copyright law requires a human author.
6. People Who Work
In 2016, at a meeting in Toronto, Hinton said people should stop training radiologists right then. He went on to say that within five years deep learning would do better than radiologists.
In March 2023 Goldman Sachs economists estimated that generative AI could expose the equivalent of three hundred million full-time jobs to automation. In January 2024 the managing director of the International Monetary Fund, Kristalina Georgieva, wrote that nearly forty percent of employment worldwide was exposed to AI, and about sixty percent in advanced economies. Exposure could help some of those jobs and harm others. In a May 2025 interview, Amodei said that in one to five years AI could wipe out half of entry-level white-collar jobs, and the US unemployment rate could rise to between ten and twenty percent.
What people measured pointed in different directions. One study followed more than five thousand customer-service agents at a software company after they got an AI assistant. The number of issues resolved per hour rose about fourteen percent on average, with most of the improvement among newer agents. In a writing experiment, the ChatGPT group took about forty percent less time and scored about eighteen percent higher on quality. In an experiment with 758 consultants, those using AI were faster and better on tasks it was good at, but were more likely to get a problem wrong when it lay outside what it could do. A Danish study linked survey responses to government earnings and hours records. By the end of the period it had found no clear change. In August 2025 a Stanford study using payroll data reported a relative decline in employment among people aged twenty-two to twenty-five in the jobs most exposed to AI. In July that year an organization called METR ran a randomized experiment with experienced open-source programmers working in codebases they knew. With AI they took about nineteen percent longer, but afterward they still thought they had been faster.
Some people wrote the boundary into contracts. The September 2023 agreement between the Writers Guild of America and the studios said AI wasn't a writer, AI-generated material couldn't be used to cut writers' credit or pay, and companies couldn't make writers use it. The actors' union agreement put rules for informed consent and compensation around digital replicas. From July 2024, video-game voice and motion-capture performers went on strike over the issue for almost a year.
7. Six Million
In December 2024, DeepSeek in Hangzhou released the technical report for V3. It estimated the computing cost of the official training run at 5.576 million dollars, based on an assumed rental price for graphics processors. The report says this left out earlier research and experiments that tested what happened when parts of the system were removed. On January 20, 2025, the company released R1. On January 27 US technology shares fell sharply, and Nvidia lost about $589 billion in market value in a day. What traveled was that a Chinese company had trained R1 for six million dollars.
8. Seventy Percent
In 1950, in essay five, Turing guessed that in about fifty years an ordinary questioner would have no more than a seventy percent chance of identifying the other party correctly after five minutes of questions. In June 2014 the University of Reading announced that Eugene Goostman, a program playing a thirteen-year-old Ukrainian boy, had convinced about thirty-three percent of the judges it was a person and passed the Turing test. In 2025 Cameron Jones and Benjamin Bergen of the University of California, San Diego, ran an experiment following Turing's original three-party design. A questioner chatted for about five minutes with a person and a program at the same time, then said which was human. GPT-4.5 given a persona was chosen as the human about seventy-three percent of the time. The actual person in the same comparison was recognized less than thirty percent of the time. ELIZA from essay nine was there, as a control. The paper's title was “Large Language Models Pass the Turing Test.”
In 1843 Lovelace wrote in her notes that the Analytical Engine had nothing in the way of a claim to originate anything, and that it could do whatever we knew how to order it to perform. As essay one notes, a later retelling had added only. In 1950 Turing gave the line under Lady Lovelace's objection. In 2001 three researchers proposed a test named after her. One was David Ferrucci, who would lead IBM's Watson project when it beat human champions on a television quiz show in 2011.
Turing asked whether the person on the other end could recognize a machine. More than seventy years later, a program went unrecognized seven times in ten. Around the same time, the EU, California, and China wrote rules for different things; all required that in some circumstances one point be made clear. The other side is a machine. The one sitting at this end of the line is a person. They have to know who is on the other end to decide what to hand over: time, something private, a decision. This series calls the part that no resemblance can replace, that only a person can hold on to, the remainder.
The machines are still being built. Each generation says again where the boundary lies.