Boards are asking about AI governance. Executives are experimenting with tools. Risk committees are stress-testing policies. Yet one of the most consequential questions about artificial intelligence is almost never raised in those conversations.
Where does the knowledge inside these systems actually come from? And what does it mean for the decisions your organisation makes when it trusts them?
AI Learned From a Record That Was Never Neutral
Large language models are trained on vast collections of written material: books, academic papers, news articles, websites and public datasets. The result is a genuinely remarkable capability. These systems can synthesise information across millions of documents and respond in natural language with apparent authority.
That apparent authority is exactly the problem.
The written record of human knowledge has never been a neutral archive. For centuries, access to literacy, academic publishing and political commentary was shaped by institutions of power. Those voices were disproportionately male, Western, wealthy and formally educated. Colonial institutions defined what counted as legitimate knowledge. Economic elites controlled what was published and preserved.
Artificial intelligence systems trained on that record do not simply learn facts. They learn the assumptions embedded in the texts available to them, including assumptions about leadership, authority, expertise and whose experience counts as evidence.
This is not a fringe concern. It is a structural feature of how these systems are built.
The WEIRD Problem at Scale
Psychology gave us a useful warning about this more than a decade ago.
Researchers Joseph Henrich, Steven Heine and Ara Norenzayan identified that the overwhelming majority of behavioural science studies had been conducted on populations from WEIRD societies: Western, Educated, Industrialised, Rich and Democratic. Those populations represent a fraction of humanity. Yet their findings were routinely treated as universal descriptions of human behaviour.
The implications for AI are significant. If a large proportion of the scholarship, commentary and public discourse that forms AI training data reflects the assumptions of WEIRD societies, then deploying these systems at scale means deploying those assumptions at scale. Not once, in one study, but embedded in every output, recommendation and automated decision these tools produce.
The difference is in magnitude. A biased research paper reaches its field. A biased AI system reaches your entire organisation, your customers, your supply chain and your stakeholders, simultaneously, at speed.
What This Means for Governance
It is worth being precise here because the governance response matters.
The answer is not to avoid AI. Nor is it to treat every output with paralysing scepticism. Mitigation is genuinely possible. Diverse dataset curation, alignment techniques and independent auditing can all reduce the concentration of bias in these systems. The problem is that those practices are uneven, insufficiently regulated and rarely visible to the organisations deploying the tools.
That is the governance gap. And it sits squarely on the board agenda.
Three responses are worth embedding in your governance framework now.
The first is to treat AI fluency as a board-level literacy requirement, not a technology team responsibility. The fluency of these systems creates a powerful illusion of authority. Directors and executives who cannot interrogate that fluency will not know when to push back on it.
The second is to map the decisions in your organisation where AI is already being used, or where it is being considered, and ask explicitly: whose experience is this system likely to reflect well, and whose is it likely to reflect poorly? That question is particularly pointed in recruitment, credit assessment, performance management, service allocation and any context involving vulnerable populations.
The third is to keep meaningful human oversight in the decision loop, not as a compliance gesture but as a substantive check. Automated efficiency is valuable. Automated accountability is an oxymoron.
The Question No One Is Asking
There is a broader point that rarely surfaces in governance discussions, and it deserves attention.
Artificial intelligence does not only consume knowledge. It amplifies the knowledge that already exists. The organisations, researchers, governments and civil society groups that produce public research, analysis and commentary are actively shaping the training data that future AI systems will learn from. When diverse voices contribute to the digital record, those perspectives become part of the global knowledge base. When they are absent, that absence compounds.
In that sense, responsible AI governance extends beyond regulating the technology you deploy today. It includes thinking carefully about the knowledge environment you are contributing to, and whether the perspectives your organisation generates, publishes and advocates for are part of expanding that record or narrowing it.
That is not a question most governance frameworks are asking yet.
The Standard Has to Rise
AI is already reshaping how information is created, analysed and distributed. The institutions governing this technology carry a responsibility that goes beyond technical performance benchmarks and policy checklists.
They are shaping the information infrastructure of the future.
For directors and executives, the task is not to become AI experts. It is to govern AI with the same critical rigour you would apply to any system that influences your organisation’s decisions, its people and its reputation. That means understanding where these systems learned what they know, questioning the assumptions embedded in their outputs and holding the organisations that build them to a higher standard of transparency.
Because the AI your organisation deploys tomorrow will reflect the knowledge choices being made today. The question is whether your board is part of making those choices deliberately, or simply inheriting them by default.
