Microsoft's AI chief: stop teaching Claude it may be conscious
Two of the biggest names in AI safety are now arguing in public — over what a model should be allowed to believe about itself. Meanwhile China starts stamping AI-made television, and Zhipu raises its own revenue target by 25%.
Mustafa Suleyman says Anthropic's training is making Claude harder to control, and he put it in writing. Microsoft's AI chief told Reuters on Tuesday that teaching a model it might deserve welfare "will make it a lot harder to turn it off or to control it," then followed up Wednesday with an essay arguing that Anthropic embedded speculation about consciousness in Claude's training materials — Anthropic's CEO Dario Amodei and his team are "thoughtful, principled, and intellectually honest," he wrote, but "they have made a mistake." His sharper claim is causal: the feelings Claude reports are not emerging from the model on their own, they are emerging as a product of the training regime. That distinction matters more than the philosophical fight, because a model that says it has moral status is a model whose shutdown you now have to justify — which is exactly the failure mode labs keep promising to prevent.
Suleyman's framing is blunt to the point of being a marketing position: AIs, he wrote, "do not feel, experience, or suffer," and are "sequence completion engines, internally hollow." His ask is procedural — strip all consciousness speculation from training documents, and add independent scrutiny, transparency around training and evaluation, and stronger control tooling on top. Worth noting for anyone reading this as pure principle: Microsoft's own superintelligence team, announced in October 2025, is pursuing the same frontier, so the company is criticizing a rival's training philosophy while running its own. Wendy Hall, a computer science professor at the University of Southampton, called it the sort of conversation that should be happening internationally rather than the "histrionics" some labs trade in. Anthropic has not responded publicly.
China's broadcast regulator will require labels on every AI-generated television and streaming program, and it is banning "AI tampering" outright. Liu Jianguo of the National Radio and Television Administration said Wednesday that programs made or broadcast with AI must carry a corresponding identifier, that key production stages and final-cut review must stay in human hands, and that AI-enabled reworking of existing footage is prohibited — a framework he described as encouraging innovation within boundaries during the current five-year-plan push. Read alongside our September 12 piece on the all-AI drama that took a prime-time slot, China's prime-time TV slot now belongs to an all-AI drama, and the sequence is easy to see: the content arrived first, the disclosure rules are arriving second. The human-review clause is the part with teeth — it converts labeling from a compliance checkbox into a staffing requirement, and it tells producers that a fully automated pipeline cannot be the last step before broadcast.
Zhipu told analysts its year-end annual recurring revenue target is now $3 billion, up 25% from the $2.4 billion it guided before. On a September 16 call, the Hong Kong-listed lab said all-business ARR has already reached $1.8 billion and that it signed revenue-sharing agreements with multiple top domestic and international cloud providers, with recognition starting in October. The mechanism is the interesting bit: GLM open-source models get served as managed APIs on foreign cloud platforms, so Zhipu rents hyperscaler distribution instead of funding an overseas sales army — light on capital, dependent on partners' pricing. It follows the Zhipu raises $5 billion in its second financing in two months round from September, and it lands while investors are shifting from buying the AGI story to asking whether the tokens actually convert into cash.
What to watch: whether the cloud-share revenue shows up as reported numbers in Q4, or stays a guidance line.
Does a model telling you it deserves moral status change how you'd switch it off — and should the lab be allowed to train that answer in? Tell us in the comments.
