OpenAI pushes for liability limits
In the US debate over AI accountability, OpenAI backed an Illinois bill that would limit when AI firms can be held liable for major harms, even in extreme cases like mass deaths or financial disasters. At the same time, a judge in a recent copyright lawsuit criticised an OpenAI witness for failing to explain how the company prevents chatbots from reproducing copyrighted material, intensifying legal scrutiny as the firm seeks lighter liability ( ).
OpenAI is backing an Illinois bill that could make it much harder to sue the company if one of its most powerful artificial intelligence systems is tied to a catastrophe, including deaths, serious injuries, or billion-dollar property losses. Wired reported that OpenAI testified in favor of the measure this week. (wired.com) The bill is Illinois Senate Bill 3444, introduced on February 4, 2026, and it creates an “Artificial Intelligence Safety Act” built around a liability shield for the companies that make so-called frontier models. The Illinois General Assembly summary says developers would not be liable for “critical harms” if they did not intentionally or recklessly cause them and if they publish safety and transparency reports online. (ilga.gov) In plain English, the bill treats the model maker a bit like an engine manufacturer and the company using the engine like the driver. If the model maker says it published the right paperwork and did not act intentionally or recklessly, the lawsuit target shifts away from the lab and toward whoever deployed the system. (wired.com) The scale in the bill is unusually high. Reporting on the proposal says “critical harm” includes death or serious injury to at least 100 people, or at least $1 billion in property damage, which means the liability fight is centered on extreme worst-case scenarios rather than ordinary product complaints. (wired.com) The protection is not for every chatbot on the internet. Coverage is aimed at “frontier” systems, which bill summaries describe as models trained with computing resources expensive enough that only a small group of top labs can usually build them. (letsdatascience.com) This is a change in tactics for OpenAI. Wired says the company had mostly been trying to block state bills it disliked, but in Illinois it moved from defense to offense by supporting language that writes narrower legal responsibility for model makers into state law. (wired.com) At almost the same moment, OpenAI is being pressed in court on a different question: whether its systems can reproduce copyrighted journalism and books too closely. On April 9, 2026, Magistrate Judge Ona Wang ordered a second deposition for OpenAI witness John Vincent “Vinnie” Monaco after finding he could not answer basic questions about the company’s anti-plagiarism work. (nationaltoday.com) Judge Wang wrote that Monaco relied on “hazy recollections” about Project Giraffe, an internal OpenAI effort described as a way to limit copyright infringement by its chatbots. She also said OpenAI’s lawyer objected at least 200 times during the first deposition, which “impeded and frustrated” questioning by the news organizations suing the company. (nationaltoday.com) That puts the two fights on the same track. In Springfield, OpenAI is arguing that model makers should face less legal exposure unless they acted intentionally or recklessly, while in Manhattan a federal judge is demanding a clearer account of how the company actually prevents one concrete harm: a chatbot spitting back copyrighted text. (wired.com; nationaltoday.com) If Illinois advances Senate Bill 3444, other states will have a ready-made template for separating the lab that builds a model from the business that uses it. If the courts keep forcing more disclosure in copyright cases, judges may end up learning exactly how much control those labs really have over the systems they want shielded. (ilga.gov; wired.com; nationaltoday.com)
