OpenAI legal fights push AI governance practical

A Manhattan judge just told OpenAI to put up another corporate witness after the first one could not answer basic questions in a copyright case over whether chatbots can reproduce news and book passages too closely. The judge said OpenAI employee John Vincent “Vinnie” Monaco had to sit for 3.5 more hours of deposition time. (chicagotribune.com) (law.justia.com) That fight is not about a speech or a white paper. It is about sworn testimony in federal court on what OpenAI actually built to stop “regurgitation,” which is the industry term for a model spitting back copyrighted text too closely to the original. (law.justia.com) (courtlistener.com) At the same time in Illinois, OpenAI backed Senate Bill 3444, a state proposal that would limit when developers of “frontier” artificial intelligence models can be sued for “critical harms.” The bill says a developer would not be liable if it did not intentionally or recklessly cause the harm and if it published safety, security, and transparency reports. (wired.com) (trackbill.com) The bill defines “critical harms” in unusually concrete numbers. One threshold is death or serious injury to at least 100 people, and another is at least $1 billion in property damage. (wired.com) (trackbill.com) That means the same company facing discovery demands in New York is also trying to shape the legal standard for who pays when powerful models are misused or fail. The center of the argument is moving from “should artificial intelligence be governed” to “what evidence counts” and “what level of negligence creates liability.” (chicagotribune.com) (wired.com) OpenAI leaders are also describing that shift in public, especially in national security work where a bad output can affect surveillance, targeting, or cyber defense. On April 9, 2026, OpenAI national security policy head Sasha Baker said agencies will need a “workforce transformation” so humans still make the final calls. (nextgov.com) Baker used the Pentagon phrase “appropriate human judgment,” which is a practical standard, not a slogan. It means a human being must remain responsible for the decision in high-consequence settings even if software did most of the analysis first. (nextgov.com) She made that point while discussing defense uses that range from paperwork to the “targeting cycle,” which is the process militaries use to identify and act on targets. She also said no large language model, including ChatGPT, is foolproof. (nextgov.com) Put those three April moves together and the pattern is hard to miss. Judges want named employees who can explain safeguards under oath, lawmakers are being asked to write narrower liability rules, and executives are framing safety as documented human sign-off instead of abstract promises. (chicagotribune.com) (wired.com) (nextgov.com) That is what artificial intelligence governance looks like once it leaves conference panels and lands in courtrooms, legislatures, and procurement offices. The key questions are now who testifies, what gets logged, when a company can be sued, and which human had the authority to say yes. (law.justia.com) (wired.com) (nextgov.com)