UBTECH's embodied-AI model clears China's generative-AI filing
Two stories from the AI front today: China's embodied-robotics leaders are learning to navigate the state's content-safety regime, while a shadow library warns that the race for training data is quietly erasing physical books.
UBTECH says its "Xingzhe" embodied-intelligence model has passed the Cyberspace Administration of China's review and completed its generative-AI service filing — making the robotics maker one of the first in the embodied-intelligence industry to clear that compliance bar. According to QbitAI, in a company-authorized report, the filing landed on August 18 and is presented as a milestone for the model's content-safety engineering, covering training-data production, red-team evaluation sets, safety guardrails, and agent design. The model was specifically fine-tuned for emotional interaction, so a humanoid can read expressions, gaze, and posture and respond with more natural companionship — the capability UBTECH wants for home, reception, and customer-facing robots.
The regulatory clearance matters because in China a generative-AI filing is the gate that lets a model actually ship to consumers, and embodied robots are now squarely inside that regime. Xingzhe shares its embodied base with Thinker, the visual-language model UBTECH open-sourced in February 2026 and claims took nine global number-one results on embodied-intelligence benchmarks. Watch the moat this builds: as Beijing turns model registration into a routine requirement for robots that talk and touch people, the companies that clear it early get a documented compliance lead over labs that ship first and ask later.
Anna's Archive says AI companies are buying, scanning, and destroying millions of physical books to hoard pre-2022 training data no machine has touched. The shadow library's blog post argues that, by pulping the originals after digitizing them, labs become the only holders of those copies — locking human knowledge onto private servers. It cites Anthropic's "Project Panama," exposed through a $1.5 billion copyright settlement with authors in 2025, as the headline example of a confidential book-buying-and-destruction operation; Wikipedia's entry on Anthropic independently confirms both the settlement amount and the project name.
It's an advocacy piece from an organization with a vested interest in free book access, so treat the "volunteer and scan everything" call as a pitch, not neutral reporting. But the underlying dynamic is real and under-covered: training-data sourcing now reaches into the physical book market, and destruction removes the public-domain copy that competitors or archives could otherwise preserve. With AI-generated text already making up more than half of new internet content since 2025, the post asks a fair question — if the last human-written sentence on paper gets absorbed and the original is gone, what's left for future models to learn from that isn't their own output?
Should AI labs be allowed to destroy the physical books they train on once the scan is done? Tell us in the comments.
Sources: QbitAI · Anna's Archive · Wikipedia — Anthropic
