3 min read
The ethics and values of AI: Challenges with alignment in a divided tech world
AI ethics and responsible AI have dominated headlines over the past year. From conferences to policy debates, the conversation often circles around diversity, inclusion, and fairness. These are critical issues, but they hide a deeper and harder question: how do we make sure AI is aligned with our values?
Alignment, at its core, is about ensuring that AI systems behave in ways consistent with human intentions. In theory, that sounds straightforward. In practice, it’s one of the hardest problems in technology today, and it becomes even more complex when we ask which values AI should embody.
What alignment really means
When researchers talk about AI alignment, they are describing methods to keep powerful AI models working toward human goals rather than drifting into unwanted or harmful behaviours. Think of it as teaching a system not just to complete a task, but to complete it in the way we consider acceptable.
For example, an aligned AI should refuse to generate harmful content, respect privacy, and provide answers that are accurate and useful. But notice the problem: deciding what counts as “harmful,” “respectful,” or even “useful” requires value judgements. And values are not universal.
Whose internet, whose values?
Generative AI models are trained on vast swathes of publicly available internet content. That content is overwhelmingly in English, and within English, disproportionately American. This creates a quiet but powerful bias.
Even if we addressed the technical problem of alignment perfectly, what we would end up with is an AI aligned primarily with American cultural norms, assumptions, and values. Politeness, humour, politics, history, gender roles, and more, all are expressed differently across cultures. Yet the training data does not treat these differences equally.
Independent tests, such as those reported by FusionChat, have even shown that large models like GPT-4 and its variants map onto different points of the political spectrum, reinforcing the fact that bias is already visible in practice (FusionChat, 2023).
The result is that when we ask for “inclusive AI,” we may only be getting inclusivity through a Western, English-speaking lens. For a global technology, that is a serious limitation.
The time factor
Bias is not just geographic, it is temporal. Training a model involves capturing a snapshot of the internet at a given moment. What was trending, debated, or dominant at that point will shape the system’s world view.
Consider how polarised online opinion has become. Social media amplifies extremes, while moderate perspectives often struggle to be heard. If this content is then used to train generative AI, the model risks amplifying the loudest voices rather than the most balanced ones.
That means AI alignment is not only about values in the abstract, but also about which moment in cultural history those values were frozen. An AI trained in 2021 will capture a different world view than one trained in 2025.
Possible futures
Different regions and communities might develop their own large-scale AI models to better serve their specific populations. For example, models tailored to local languages, dialects, cultural norms, and legal frameworks. Some models might be designed to reflect the linguistic and historical context of a specific country, while others could be developed by a community to process and understand their unique data and religious or cultural heritage.
These examples are not predictions of what should or should not happen. This isn’t a moral stance on whether such outcomes would be positive or negative. The intention is merely to highlight that multiple pathways exist, and each will reflect the context in which the models are built.
Philosophical traditions make this even more complex. Some cultures see values as absolute, rooted in fixed principles or laws. Others see them as contextual, where outcomes justify the means. This suggests that “alignment” itself may mean very different things across the world.
The key point is that we may not end up with a single global standard. In some cases, reconciliation of values may be possible, but in others the differences may be too deep. For those, parallel approaches or cultural silos might be the most realistic outcome, with models reflecting the traditions of the communities they serve.
A more divisive AI?
The internet was originally seen as something that would unite people. In practice, it has amplified division. Social media bubbles, misinformation, and culture wars have shown how technology can fragment societies rather than bring them together.
AI could accelerate this trend. Imagine a future where one model explains history from a Western liberal perspective, another from a nationalist lens, another from a religious worldview. Each insists it is aligned with human values, but “human” now means “our group, not yours.”
Instead of bridging divides, AI could deepen them, leading to different versions of reality, each reinforced by the tools people use. Alignment, in this scenario, becomes less about universal human values and more about enforcing tribal ones.
Pluralistic training
Hybrid models
Educated use
Transparent alignment
Regulation and standards
None of these paths are easy. All require international cooperation, technical innovation, and above all, humility about the limits of any single worldview.
A hard but necessary conversation
The debate about ethical and responsible AI cannot stop at generic principles. Diversity and inclusion matter. But without grappling with the core question, whose values are we embedding? We risk aligning AI to a narrow slice of humanity.
Alignment is not just a technical puzzle for engineers. It is a cultural, political, and philosophical challenge that cuts across borders. In some areas reconciliation may be possible, in others distinct cultural models may need to coexist. If AI is to serve the world, it must reflect the world. Sometimes through shared principles, and sometimes through respectful parallel approaches.
The stakes are high. Get it right, and AI could help bridge divides or at least allow different systems to coexist productively. Get it wrong, and we may find ourselves in a future where technology divides us more deeply than ever before.
Filippo Sassi is Head of the AI Labs at Version 1.
For more insights on how our AI expertise is driving innovation and helping businesses around the world transform, click here.
Related blogs
-
Blog
Prompting GPT-5
