On Friday, New York City's Council made SpaceXAI, Elon Musk's AI company and the maker of the Grok chatbot, answer questions under oath. It took the Council's first subpoena ever to a private company to get there. SpaceXAI had skipped an October 5 hearing where Anthropic, Google, Meta and OpenAI testified. Its witness, a government affairs manager named Gabriela Castillo Madrid, joined by video and said Grok has strict controls against sexualized images of real people.

Then Speaker Julie Menin held up her phone. Her staff had typed her name into the regular, free version of Grok that morning and got a sexualized image of her in under 30 seconds. Castillo Madrid said she could not speak to that image. She also could not say how many people work on safety, and when Menin asked whether the company would accept responsibility if one of its systems caused injury or death, she answered, "We're responsible for following the law." (NYC Council release; amNewYork; CBS New York)

The same hearing had a second witness. Devin Kim, a former xAI employee who says he helped build the systems used to train and test Grok, told the Council he raised concerns about weapons-related answers in early 2025 and was fired that September. He is suing for retaliation. In his time there, he said, three people worked on safety. He added that he cannot speak to how the company runs today, and Castillo Madrid said she did not know his story.

The Council is already writing rules. Last week it introduced Intro 2605, which would let whistleblowers collect a share of any fines recovered from AI companies that break the law. It is also weighing a bill to let people sue AI companies over foreseeable harms from misuse of their models. A city has no power to license AI, but it can set liability and reward people who report problems. None of it has passed. (NYC Council press page; amNewYork)

In Brazil, two federal ministries, Human Rights and Justice, published a guide on Friday on how children and teenagers use AI. It is the first federal document to speak to that audience directly, and it sorts AI into four roles: toy, companion, curator and creator. It covers privacy, mental health and age checks, and points families to help lines and to services that remove intimate images posted without consent.

The guide is advice, not a new rule. It leans on ECA Digital, the child online safety law passed in 2025, and it was written with UNESCO's help as part of Brazil's national AI plan. The sorting is the useful part. A chatbot that remembers a child's name and answers at midnight is a companion, whatever the marketing calls it. (Brazil's Human Rights Ministry, official release, in Portuguese)

China's State Council, chaired by Premier Li Qiang, heard a progress report on its "AI Plus" push on Friday. The readout calls for more high-quality data, better links between computing and power supply, and trial runs of AI in real industry. It also says to balance development with safety, set measures by industry through grading and classification, and strengthen risk assessment to hold the safety line.

The readout sets no rules and no dates. Still, "by industry" hints at how rules may arrive, and it sits alongside the export controls on top Chinese models reported this week. A company with Chinese partners or suppliers should expect AI duties to differ by sector, not arrive as one law. (State Council readout, in Chinese)

Already filed, if you want the run-up: The Tripwire (India's cyber agency and AI agents that reached live company systems), The Eight (eight state attorneys general suing OpenAI), and The Probe (the Federal Trade Commission looking into OpenAI and Anthropic).

Also on the shelf: The Subpoena (California's attorney general and OpenAI), The Force (the White House task force) and The Anchor (Canada's AI money). Read today beside The Subpoena. California compelled documents from OpenAI. New York compelled a witness from SpaceXAI, and the Council had warned it would ask a state court to enforce the order if the company kept refusing.

This is an old move. In November 2018, lawmakers from nine countries gathered in London as an international grand committee on disinformation. They had invited Mark Zuckerberg and he declined, twice, then declined a video link too. Facebook sent Richard Allan, its policy vice president, and the committee left an empty seat with Zuckerberg's name on it. (AP covered the session; TechCrunch; UK Parliament transcript)

Companies learned to send a deputy. What changed in New York is the subpoena and the oath. A deputy under oath cannot say how many people work on safety, or whether the company knows about the person who was fired. The empty chair has been replaced by a witness who has to say, on the record, what she does not know.

Put these to management. If a legislator or regulator asked our AI lead under oath how many people work on safety for our products, who would answer, and would the number be in the room? Have we typed our own executives' names into our public AI tools and looked at what came back?

Who is on our short list to testify, and have they been briefed on the hard questions? And if an employee raised a safety concern tomorrow, where would it go, who would read it, and what would we owe that person? A board cannot run the product, but it can insist the answers exist before somebody asks for them under oath.

The Move

New York put an AI company under oath, and the answers were thin. Whatever you ship, your people may be asked what you know, who watches it, and who gets blamed.

Ask the general counsel and the head of AI to bring one page to the next board meeting: who would testify for our AI products, the names of the people who work on safety, and the path a safety concern takes from an employee to this board.

If the page cannot be written, that is the finding.

Seven bullets. One move.