The hidden assumptions inside the word “AI”

International AI governance increasingly speaks a shared language. Frameworks published in Brussels, Beijing, Tokyo, and Washington all call for AI to be trustworthy, safe, beneficial, responsible, and human-centred. Reading them, you might conclude that the world broadly agrees on what needs to be governed, but that conclusion may be premature.

A chapter by Stephen Cave, Kanta Dihal and colleagues in Imagining AI: How the World Sees Intelligent Machines (2023) makes a case that is easy to state and harder to absorb: the words different languages use for “AI” carry different histories, different associations, and different embedded assumptions about what the technology is and what it is for. What looks like shared vocabulary may be concealing quite different underlying objects.

Before asking whether AI is trustworthy, there is a more basic question: what does “AI” mean?

The name was a choice, not a discovery

The term “artificial intelligence” was coined by John McCarthy for the Dartmouth project in the 1950s. It was not the inevitable name for the field. Other options were on the table: machine intelligence, engineering psychology, applied epistemology, neural cybernetics, complex information processing. McCarthy chose “artificial intelligence” partly because it was broad, attention-grabbing, and useful for distinguishing his project from Norbert Wiener’s cybernetics.

The name worked because it travelled globally and attracted funding and attention. It also created a persistent problem: it made AI sound like a single, stable technical object, when it was closer to an ambitious and contested research aspiration. Governance frameworks that call for trustworthy AI inherit that instability, because the object being made trustworthy was never precisely defined to begin with.

In English, “artificial” carries the fear of being fooled

In English, “artificial” means made by humans but it also carries the sense of artifice: imitation, trickery, something that may not be what it appears to be. The chapter traces this to the Latin artificium, meaning craft or skill but also cunning and deception.

The most literal illustration is the Mechanical Turk, an eighteenth-century automaton that appeared to play chess by concealing a human operator inside the cabinet. It was convincing enough to fool audiences across Europe for decades as “intelligence”, even though it was merely a performance of one. This anxiety about deceptive mimicry runs through the English-language AI imagination from the start. Is the system really intelligent, or is it performing intelligence convincingly enough to fool us?

Transparency requirements, explainability obligations, disclosure rules, and human oversight mechanisms are not only technical safeguards. They also answer a deeper cultural fear: is the machine pretending to be something it is not?

In English, “intelligence” carries the history of ranking people

The second half of the phrase carries different but equally significant baggage.

The chapter argues that “intelligence” in English became entangled with histories of measurement, hierarchy, and exclusion. The rise of intelligence testing in the late nineteenth and early twentieth centuries was not a neutral scientific development. It was used to rank people, justify class structures, support colonial and racial hierarchies, and in its most extreme applications, provide pseudoscientific cover for eugenics.

Many AI systems do not simply “think” – they classify, score, sort, rank, and allocate. They place people into categories that affect access to credit, employment, housing, healthcare, and public services. When those systems are described as measuring intelligence or capability, they operate within a tradition that has historically treated such measurements as objective while embedding the assumptions of whoever built them.

The governance concern is not only whether AI systems are biased in a statistical sense but also whether AI revives old habits of sorting people while presenting those rankings as neutral technical outputs, and who those measurements have historically served.

In Japanese, intelligence includes heart and feeling

The Japanese term jinkō chinō carries different associations from the outset. The chapter describes jinkō as connected to modern production or manufacture, without the “artificial as imitation” connotation present in English. More significantly, Japanese conceptions of intelligence incorporate kokoro (often translated as heart-mind), bringing affect, moral sensibility, and relational feeling into what intelligence means. In this tradition, Astro Boy’s electronic brain is located not in his head but in his chest, and building AI has always been partly a project of building something capable of feeling and moral reasoning, not only computation.

This does not make Japanese AI discourse uniformly optimistic or politically uncomplicated, but the imagined intelligent machine is not necessarily conceived as a cold, calculating rival. It may be imagined as an affective, relational agent, something closer to a companion than a tool or a threat. If AI is imagined as something people may bond with, trust, obey, or treat as socially present, governance needs to address those relationships directly, not only whether the system is deceiving or dominating us.

In Japanese, intelligence includes heart and feeling

More significantly, Japanese conceptions of intelligence incorporate kokoro (often translated as heart-mind), bringing affect, moral sensibility, and relational feeling into what intelligence means. In this tradition, Astro Boy’s electronic brain is located not in his head but in his chest, and building AI has always been partly a project of building something capable of feeling and moral reasoning, not only computation.

If AI is imagined as something people may bond with, trust, obey, or treat as socially present, governance needs to address those relationships directly, not only whether the system is deceiving or dominating us.

In Chinese, AI means wisdom, capability, and practical usefulness

The Chinese term ren gong zhi neng also encodes different associations. The chapter breaks this into ren gong (human making or craftsmanship) and zhi neng (wisdom and capability), connecting this framing to the Keju imperial examination system, where intelligence was associated with cultivated merit, demonstrated learning, and social usefulness rather than fixed biological endowment with no equivalent of the eugenic baggage that attached itself to the English concept of intelligence.

In this framing, AI appears less as a deceptive imitation of human intelligence and more as a practical capability: something made by humans, useful for solving problems, and associated with the kind of wisdom that earns its place through demonstrated contribution to the collective.

None of this means Chinese AI governance is apolitical, free from risk, or neutral in its effects. China’s AI governance involves significant state control over information flows, platform behaviour, and data, and raises serious and well-documented concerns about surveillance, civil liberties, and the use of AI systems to monitor and govern dissent.

The linguistic and philosophical background does not dissolve those concerns, but the starting assumptions differ from those embedded in English-language governance discourse: less preoccupied with deception and hierarchical ranking, more oriented toward practical deployment, social usefulness, and alignment with collective goals.

Public attitudes may not be measuring the same object

Public opinion surveys regularly ask whether people trust AI, fear it, support its regulation, or welcome its deployment in specific sectors. Presented as cross-country comparisons, the findings look like evidence of genuine cultural variation in AI attitudes.

If the word “AI” activates different associations in different linguistic and cultural contexts, those surveys may not be measuring quite the same thing. A respondent who hears “AI” through the associations of fraud, imitation, and hierarchical ranking is not answering the same question as one who hears it through the associations of heart, wisdom, or practical usefulness.

The survey question may travel faster than the underlying meaning, and if the public’s object differs across contexts, “public trust in AI” is not a single comparable variable. For international governance, public legitimacy shapes what policies are viable which makes that difference consequential.

What shared vocabulary conceals

International AI governance increasingly uses shared language: trustworthy, safe, beneficial, responsible, human-centred. That shared language can help coordination, signal intentions, and establish minimum expectations across jurisdictions.

It can also create the appearance of agreement where the underlying assumptions diverge. Different governance systems may use similar words while starting from different understandings of what AI is, what kind of danger it presents, and what kind of human-machine relationship is considered normal or desirable.

Before asking whether people trust AI, it is worth asking what kind of thing they think AI is. That is one reason international AI governance is harder than its shared vocabulary makes it look.

Read next:

The American assumptions inside AI governance

This post draws on the chapter “Cross-Cultural Perspectives on the Meaning of Artificial Intelligence” in Imagining AI: How the World Sees Intelligent Machines, edited by Stephen Cave and Kanta Dihal, Oxford University Press, 2023.