Artificial intelligence (AI) is often presented as a universal technology: build it once, deploy it anywhere, and expect similar results. This narrative of portability and neutrality has become central to how AI is marketed, funded, and imagined. It promises efficiency, scalability, and global applicability. However, AI systems do not operate in a vacuum. They are shaped by the environments in which they are designed, trained, and deployed, and those environments are far from universal.
When AI systems travel across borders, they carry with them the assumptions, priorities, and limitations of their place of origin. These embedded characteristics do not disappear; they persist, often clashing with local contexts in subtle but consequential ways. As a result, AI does not simply “fail” in new environments; it systematically misaligns with them. Understanding why this happens requires moving beyond technical explanations and examining the deeper structures that shape how AI is built and distributed globally.
The discussion is inspired by research from Lancaster University published in the paper AI ethical challenges: a perspective of AI, published in the journal Emerald Publishing’s Information Technology & People.
The myth of universal AI
At the core of contemporary AI discourse is a powerful assumption: that intelligence, once encoded into algorithms, can function independently of context. This assumption underpins the idea that AI systems are inherently transferable, that a model trained in one setting can be deployed in another with minimal adjustment. But AI systems are not abstract intelligence engines. They are products of specific social, economic, and technical conditions. They are trained on datasets that reflect particular populations, behaviours, and languages. They rely on infrastructure, cloud computing systems, data pipelines, and evaluation benchmarks that are unevenly distributed worldwide. And they are developed within institutional environments that shape what problems are prioritised and how success is measured.
Most AI systems today are produced in a relatively small number of regions, particularly North America and Western Europe. The data used to train these systems overwhelmingly reflects the linguistic patterns, cultural norms, and social structures of these regions. The benchmarks used to evaluate performance are similarly constructed. This creates a feedback loop: systems are optimised for the contexts they already represent, reinforcing their apparent effectiveness in those environments. When these systems are deployed elsewhere, the assumption of universality begins to break down. What appears to be a “general” model is, in practice, highly localised.
Embedded assumptions and contextual misalignment
The misalignment of AI systems across contexts is not random. It emerges from the assumptions embedded within them. Language provides a clear example. Natural language processing models trained primarily on American or British English often struggle with other linguistic forms—whether Nigerian Pidgin, Indian English, or multilingual code-switching common in many societies. These are not marginal linguistic variations; they are central modes of communication for millions of people. Yet they remain underrepresented in training data, leading to systematic errors in interpretation.
The same pattern appears in other domains. Facial recognition systems trained on limited demographic datasets perform unevenly across populations. Recommendation systems reflect cultural preferences embedded in their training data. Predictive models rely on behavioural patterns that may not hold in different socio-economic contexts. These failures are often framed as issues of “bias” or “data gaps.” While this framing is not incorrect, it is incomplete. It suggests that the problem can be solved by simply adding more diverse data. In reality, the issue runs deeper. The structure of AI development itself privileges certain forms of knowledge and experience while marginalising others.
Dependency and the limits of local innovation
For developers working outside the core centres of AI production, these dynamics create a structural dilemma. Building AI systems from scratch requires access to large datasets, computational resources, and specialised expertise—resources that are unevenly distributed globally. As a result, many developers rely on pre-trained models, APIs, and cloud platforms produced elsewhere. This reliance is not merely a practical choice; it is a structural condition. Development becomes an act of adaptation rather than creation. Developers modify existing systems to fit local needs, but the underlying architecture remains externally defined. This limits the scope for fundamentally different approaches.
Over time, this dynamic produces a form of technological dependency. Constraints imposed by external systems shape local innovation. Even when developers seek to address local problems, they do so using tools that may not fully align with those problems. The result is a cycle in which dependency reinforces itself: the more systems depend on external infrastructure, the harder it becomes to develop alternatives. Thus, who defines the parameters within which innovation occurs? Who sets the standards that determine what counts as “good” AI?
Infrastructure and uneven possibilities
AI development depends on infrastructures, such as stable electricity, reliable internet connectivity, and access to computational resources. These infrastructures are often taken for granted in discussions of AI, yet they are unevenly distributed worldwide. In contexts where these conditions are limited, developers must work within constraints that shape both what can be built and how systems perform. Models may need to be simplified to run on less powerful hardware. Connectivity issues may constrain data collection. Deployment may be affected by inconsistent access to digital services.
These constraints are frequently framed as local deficiencies, problems to be solved through investment or capacity building. While investment is important, this framing obscures a broader reality: global technological infrastructure is itself unevenly developed. The concentration of data centres, cloud services, and computational capacity in certain regions reflects historical and economic patterns that extend beyond the control of individual developers or countries. As a result, the challenges developers face in resource-constrained contexts are not merely local issues; they are manifestations of global inequality
Structural incompleteness
Public discussions of AI often focus on bias, the idea that systems produce unfair outcomes because of skewed data or flawed algorithms. While bias is an important concern, it does not fully capture the nature of the problem. A more accurate description is structural incompleteness. AI systems do not, and perhaps cannot, fully represent the diversity of human experience. They are built on partial datasets, shaped by selective priorities, and constrained by the limits of their design.
When these systems are deployed in contexts that differ from those in which they were developed, their incompleteness becomes more visible. They misinterpret behaviours, overlook relevant factors, and produce outputs that do not align with local realities. This incompleteness has real consequences. AI systems are increasingly used in decision-making processes that affect access to services, employment, healthcare, and education. When these systems misrepresent individuals, they do not merely produce errors—they shape opportunities and outcomes.
The limits of current AI ethics frameworks
In response to these challenges, a range of AI ethics frameworks has emerged. These frameworks emphasise principles such as fairness, transparency, accountability, and human oversight. They represent an important step toward recognising the social impact of AI. However, these frameworks often operate at an abstract level, limiting their effectiveness. They focus on how systems should behave without fully addressing the conditions under which they are produced. As a result, they tend to treat issues such as bias as isolated problems rather than as symptoms of deeper structural dynamics.
For example, improving fairness metrics does not address the dependency on external infrastructures. Increasing transparency does not redistribute control over data or models. Accountability mechanisms may identify harm without altering the systems that produce it. In this sense, current AI ethics approaches risk managing the symptoms of inequality without confronting its underlying causes.
Rethinking AI from the ground up
Addressing these challenges requires a shift in how AI is conceptualised. Rather than treating it as a universal solution, it must be understood as a situated system, one shaped by context, history, and power. This shift has several implications. First, it requires greater involvement of local actors in the design and development of AI systems. This goes beyond participation; it involves recognising local expertise as central to innovation. Developers, researchers, and communities must have the capacity to shape systems from the ground up, rather than adapting externally defined models.
Second, it requires investment in contextually relevant data and infrastructure. Building AI that works across diverse contexts depends on datasets that reflect those contexts and on infrastructures that support local development. Without these foundations, efforts to improve AI will remain limited. Third, it requires rethinking how success is measured.
Current benchmarks prioritise technical performance, often based on standardised datasets that do not capture local realities. Alternative metrics should consider social impact, usability, and alignment with community needs. Fourth, it requires engaging with diverse forms of knowledge. AI development has historically prioritised certain epistemologies, ways of knowing, over others. Expanding the scope of AI requires recognising the value of local languages, cultural practices, and alternative perspectives.
The global stakes
The uneven functioning of AI is not a marginal issue; it has global implications. As AI systems become embedded in critical sectors, their limitations shape how resources and opportunities are distributed. In healthcare, misinterpretation of local symptom expressions can lead to misdiagnosis. In recruitment, algorithms trained on narrow datasets may disadvantage candidates from different educational or cultural backgrounds. In governance, automated systems may fail to account for local social dynamics, producing decisions that appear neutral but reinforce existing inequalities.
These outcomes are not isolated incidents; they reflect broader patterns. When AI systems are misaligned with local contexts, they can amplify existing disparities rather than reduce them. At the same time, the concentration of power in AI development raises questions about sovereignty. When key technologies are controlled by a small number of actors, decisions about how they function and whose interests they serve are made at a distance. This creates a disconnect between those who design systems and those who live with their consequences.
For whom does AI work?
The question is not whether AI can function globally, but under what conditions it does so, and for whom. The assumption of universality obscures the realities of how AI is produced and deployed. It masks how systems reflect the contexts of their origin and the inequalities embedded within global technological infrastructures. Recognising this does not mean rejecting AI. It means rethinking how it is built, who it serves, and what it is designed to do. It means moving from a model of universal deployment to one of situated innovation. Only by addressing the structural conditions that shape AI, dependency, infrastructure, and power can we begin to build systems that reflect the diversity of human life. Until then, AI will continue to work unevenly, reinforcing the very inequalities it is often expected to resolve.
Reference
Abraham, I., Sutanto, J., Zhu, R., & Honary, M. (2026). AI ethical challenges: a perspective of AI developers in postcolonial countries. Information Technology & People, 1-27. https://doi.org/10.1108/ITP-11-2024-1466
