AINEWS
At this year's World AI Conference in Shanghai, a Fudan philosopher told model builders that their systems lack a body, a world, and others; the same week, Fudan published a policy blue book codifying the fix.
Fudan philosopher Sun Ning opened WAIC 2026's AI-and-mind forum with a thesis that would have made a 1920s phenomenologist nod: an intelligence without a body, a world, and others is not yet intelligent, and the AI industry is running out of ways to disagree.
The talk, titled《智能之前,先有世界》(Before intelligence, there is a world), sat at the center of a week when philosophy stopped being a footnote in Chinese AI and became the meeting agenda. According to QbitAI's on-site WAIC reporting by Jin Lei, the same window saw the release of the Fudan National Development & Intelligent Governance Comprehensive Lab's 2026 人文社会科学智能发展蓝皮书 (Blue Book on Intelligent Development of the Humanities and Social Sciences). The 2025 first edition, released 2025-03-01, was led by 53 scholars across roughly 10 institutions under editor Wu Libo, with the Fudan Development Institute coordinating.
What changed in 2026, as QbitAI reports, is the specificity. The new blue book introduces a framework called STRIDES (theory, method, data, execution, and review checkpoints) and a China Universities AI4SSH Index. It draws a line between agent-type and assistive-type AI embedding in social-science workflows, and insists that 人在回路 (human-in-the-loop) must include intervention, correction, and explanation rights, not just oversight.
The mechanism Sun Ning proposed is grounding: the cognitive-science argument that symbols do not mean anything until they are tied to a body, a physical environment, and a social world. He traces the line to Stevan Harnad's symbol-grounding problem (1990) and Rodney Brooks's physical-grounding maxim that the world is its own best model. Translated to 2026 model-building, the claim is that scaling data and parameters will not, by itself, give a system intuition, common sense, or the kind of context-sensitivity a Fudan philosopher expects from an interlocutor.
Four engineers in the same WAIC week produced empirical exhibits that mapped directly to the philosophers' categories: Apex Search on peer-review generation, the Epitome social simulation, the post-AlphaFold shift to dynamic conformational generation, and a cross-national trust study.
Chen Yongchao of Apex Search reported that his team submitted 34 papers to ACL/ARR, with about 10 reaching the Findings or main track and two scoring 3.67, a mark that sits above roughly 95% of human submissions. Fudan big-data professor Wei Zhongyu, on the same panel, flagged the result as a peer-review gaming risk, not a model capability benchmark. The disagreement is the point: high-volume generation has moved from output volume to evaluation games, and the philosophical question of what counts as understanding is now a reviewer-side problem.
Wang Tiandong of the Fudan Shanghai Math Center drew a division of labor: AI expands the search, humans define the value, the machine verifies. Yang Zixiong of the Shanghai Institute for Science-AI described the post-AlphaFold shift as one from static structure prediction to dynamic conformational generation, a form of grounding-by-simulation that does not need a robot body to qualify. Qu Jingjing of the Shanghai AI Lab demoed Epitome, a social-simulation platform that ran roughly 1,000 virtual samples through ten years of social evolution compressed into 14 hours. Each is a different way of saying what Sun Ning said in the first talk: anchor the system in something other than text.
Jiang Zhuoren of Zhejiang University's public management school contributed the cautionary case. A cross-national trust study reported a macro correlation above 90% on country-level means, but significance coverage dropped to 34-37% in regression, and country-type recovery was near random. The pattern fits a grounding critique of large-scale social AI: country-level averages are not the same construct as within-country behavior, and a system that scores well on the first can still misfire on the second.
The blue book and the forum also pushed the governance thread. Ma Lipeng of Qingpu Fudan argued for treating AI meta-cognition as a design requirement. Yang Min of Fudan CS named self-replication as a red line. Gao Qiqi of Fudan IR made the institutional case against 文科无用论 (the uselessness of the humanities). The combined position: a high-capability system that cannot explain its decisions, intervene in its own loop, or refuse a self-replication order is not a controllable system, regardless of its benchmark score.
Luo Xiaozhou of SIAT and Senris Biotech offered process evidence: a Shenzhen synthetic-biology mega-facility, funded above 2 billion RMB (about $280M at mid-2026 reference rates), runs as an AI-driven automated wet-lab, with squalene output already in Merck's supply chain and squalane in cosmetics supply chains. Meger Zhang Yan, in the same WAIC track, described the broader shift as one from process-delivery to outcome-delivery as the unit of payment and allocation. The exhibit is not philosophy as window dressing.
The skepticism still applies. Apex Search's 34 papers and the blue book's STRIDES checkpoints are both new artifacts, and a single conference week is not a track record. A useful test for the next six months: does the 2026 蓝皮书's distinction between agent-type and assistive-type AI embedding show up in a published evaluation that catches the Jiang-style 34-37% regression gap, or does it stay as a vocabulary upgrade. If grounding becomes a peer-review gate, the philosophers have done real work. If it becomes a slide-deck noun, the borrowing is aesthetic.