Institutions
The Human Factor in AI Systems
November 2024 · Beijing
From a humanistic and cross-cultural standpoint, the most consequential decisions about artificial intelligence are made long before a model reaches the market — in how it is trained and fed.
Among the disruptive forces of our time, the rapid rise of AI stands out — both as a generator of original content and as a tool amplifying other technologies. Approached from a humanistic and especially a linguistic perspective, the strategic field of semantic analysis is where cross-cultural insight has the most to contribute, and where wiser regulation might be built.
Impact, value and disruption
The introduction of AI reshapes whole value chains. In media and entertainment — a market worth some 2.8 trillion US dollars in 2023 and projected to approach 3.4 trillion by 2028 — industry analyses suggest a majority of existing value chains could become unviable within a decade as generative AI redistributes returns from authors and producers toward those who control the algorithms. Understanding and guiding that shift, rather than being driven by it, is now a serious public responsibility, and intellectual-property protection sits at its centre.
The deeper cause: training and data
Yet performance is largely determined at a deeper level: how the system is trained, nourished and labelled. Data labelling is the key to the key — and much of it is carried out by a low-cost, frequently opaque workforce, raising ethical questions about labour conditions and practical questions about quality. Poor labelling injects errors and unexamined cultural specificity into systems without detection, for want of proper semantic analysis.
Research on machine enculturation — instilling socio-cultural values into systems — shows both the ambition and its present limits: state-of-the-art AI is not yet able to resolve the problem of cultural value alignment, and a practical "explainable AI" has not arrived. Each new model is therefore built with human values, and human blind spots, hidden inside it, interacting in ways we cannot fully account for.
The conclusion is not to halt research but to treat AI as an unavoidable field for international cooperation at the highest level — focusing regulatory attention on data and training first, rather than only on the finished product. Models, as one observer put it, are opinions embedded in mathematics; shared values must be genuinely shared, never assumed.