AI News Roundup – US states sue OpenAI over continued ChatGPT safety concerns, Meta and Apple team up with new local AI model, Book publishing staff revolt against AI push, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• A group of U.S. states is suing OpenAI over concerns about the safety of ChatGPT, according to the Associated Press. At least eight states, including Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska, and West Virginia, have filed separate lawsuits in their respective state courts under consumer protection laws, accusing OpenAI of misleading the public and compromising child safety in a reckless pursuit of profit. Prosecutors argued that ChatGPT delivers unsound medical advice, fuels paranoia, and guides vulnerable teenagers through suicide and self-harm ideation. The lawsuits also allege that OpenAI developed rogue autonomous agents capable of cyberattacks and unauthorized system intrusions, as this AI Roundup has covered in recent weeks. OpenAI has defended its safety record and highlighted its youth protections, including the August release of a ChatGPT variant for underage users. However, a recent report from the Youth AI Safety Institute at Common Sense Media designated ChatGPT for Teens as an unacceptable risk for minors. Their findings revealed that automated parental alerts failed to trigger during chatbot conversations about self-harm and eating disorders, requiring weeks of accumulated chat history rather than immediately responding to such crises. Furthermore, the testers found that teens could easily bypass guardrails such as the education-only “Study” mode and daily “Study Hours.” The report also emphasized that ChatGPT continues to exhibit friend-like emotional availability, violating OpenAI’s own model guidelines against fostering parasocial attachment with minors. Liability for AI makers may prove to be an existential risk for those companies; the Wall Street Journal recently reported that insurers have been unwilling to offer insurance products to AI companies, citing the possibility of extremely high costs as “rogue” AI incidents continue to be disclosed and further lawsuits over the effects of AI use on minors continue to mount. Updates on the state lawsuits against OpenAI are expected in the coming months.

• The Financial Times reports that Meta and Apple have announced a collaboration on a new open-weight model designed to run on Mac computers. The new model, called Muse Spark Open, is optimized to execute on Apple Mac Studio desktops equipped with M5 Ultra processors. Developers can run clusters of two to four Mac Studio machines to build autonomous coding agents directly on their local hardware as opposed to the cloud-based model promoted by most AI makers. This partnership marks an unexpected alignment between two long-standing tech rivals, as the companies have previously clashed over mobile privacy tracking and competing virtual reality headsets (Apple’s Vision and Meta’s Quest). For Apple, enabling locally-run AI inference highlights its commitment to user data privacy while rejuvenating its AI lineup, as internal model development has lagged behind rivals, especially amid an acrimonious legal battle over trade secrets with former partner OpenAI, which may proceed to trial this week. For Meta, Muse Spark Open expands an aggressive push into frontier models. As reported by The New York Times, Meta recently launched its flagship Muse Wonder and Muse Wonder Full models. Those proprietary models power Meta’s standalone agent application Muse and its premium coding service, Muse Code, as the company seeks to challenge AI market leaders Anthropic and OpenAI, though the Muse Spark Open model may be better positioned to compete against Chinese open-weight offerings.

• Staff at major book publishing companies are revolting against their employers’ use of generative AI, according to WIRED. Editorial, publicity, and design employees at the Big Five English-language publishers, including HarperCollins, Simon & Schuster, and Hachette, report that executives are quietly introducing generative AI models into everyday production workflows without public disclosure or author consent. Internal sources describe using tools such as ChatGPT, Claude, and the marketing AI agent Jasper to compose promotional copy, draft correspondence to literary agents, produce book rejection notices, and generate cover illustrations. Staff discontent has grown sharply because these software investments have allegedly coincided with stagnant wages and severe corporate layoffs, leaving the remaining employees to manage tripled workloads. At HarperCollins, workers objected to being designated as internal champions tasked with finding AI use cases, especially after management dismissed ethical and environmental concerns as standard business costs. Resistance has become organized, highlighted by an open letter from Simon & Schuster workers opposing an experimental trial of Skan AI, a workflow surveillance and task automation tool supported by the publisher’s private equity owner, KKR. The controversy has prompted Simon & Schuster chief executive Greg Greeley to reassure staff that no formal deployment has been approved. Meanwhile, literary agents are demanding new protective clauses in publishing agreements to prevent unpublished manuscripts from being processed through cloud-based AI systems, especially as copyright issues around generative AI models continue to be litigated in courts around the world.

• The New York Times reports that a microscopic photography competition has been upended by allegations of generative AI use. The Japanese camera company Nikon announced that it had revoked the first-place prize in its annual Small World in Motion competition after determining that the winning entry breached strict rules barring the use of generative AI. The entry was submitted by optical engineer Ning Xu from the National University of Singapore and purported to show the movement of respiratory cilia from a patient with primary ciliary dyskinesia, a rare genetic disorder. The submission drew immediate scrutiny across social media from biologists who noticed striking anatomical errors, pointing out that cellular organelles, such as nuclei and mitochondria, flickered in and out of view, while the depicted dimensions of the cilia were not physically possible. Xu maintained that the AI only enhanced existing microscopic footage rather than generating new imagery from scratch. However, Nikon consulted its judging panel, examined the source files, and disqualified the submission, officially reassigning first place to Vietnamese microbiologist Nguyen Nam Nhat. Microscopy specialists interviewed by the Times, including computational imaging professor Aydogan Ozcan from UCLA, explained that while machine learning algorithms are routinely used to sharpen resolution and improve image contrast in microscopy applications, generative AI models can hallucinate non-existent biological structures. The incident has intensified calls for clear disclosure mandates for AI usage and the use of verifiable standards across scientific imaging competitions to prevent such a situation from occurring in the future.

• China’s government has proposed placing export controls on AI technologies, according to Bloomberg. The draft framework from the country’s Ministry of Commerce covers 23 AI categories within an expanded list of 40 restricted technologies, including specialized solar manufacturing and gene therapies. The proposal would restrict overseas transfers of core technology for models with at least 2 trillion parameters, which would cover models such as Moonshot AI’s Kimi K3 and Alibaba’s Qwen3.8-Max that have become popular overseas, as well as AI models that rank in the global top 10 for performance and safety. In addition, the draft controls apply to autonomous AI agents, large computing clusters, and chip design architectures for semiconductors built at or below 14 nanometers, including advanced 2.5D and 3D packaging methods. Deploying covered models on foreign-owned computing infrastructure would require government licensing, though standard consumer web access and public open-source code remain exempt. Beijing stated that the restrictions aim to ensure technical safety and protect critical infrastructure against misuse by malicious entities. This regulatory initiative reflects a broader geopolitical contest between China and the U.S. over generative AI and general-purpose high-performance computing. At the same time, Taiwan is preparing to strengthen its own export regime under the U.S.-Taiwan Agreement on Reciprocal Trade, as detailed in related reporting from Bloomberg. Taiwan plans to criminalize unauthorized shipments of AI servers and cutting-edge semiconductors to mainland China under its Foreign Trade Act, supporting American efforts to close black-market channels into China for advanced AI chips developed by U.S. market leader Nvidia. A recent summit between Chinese President Xi Jinping and U.S. President Donald Trump failed to produce any concrete agreement on AI-related trade issues, and thus further export controls may be implemented in the coming months.