It started with a question from my child.

"Is Hong Kong part of China?"

At face value, it was a simple query. The kind kids throw out at random moments, testing your knowledge or curiosity. We were chatting casually, and I decided to check what artificial intelligence would say. I typed the question into two different AI platforms: ChatGPT and DeepSeek. The results were strikingly different.

One explained assertively that Hong Kong is a Special Administrative Region of China, with its own legal and political systems, a legacy of British colonial rule and the 1997 handover. The other confirmed that Hong Kong is an inalienable part of China. Historically and legally, Hong Kong has always been Chinese territory. 

What struck me wasn't just the difference in answers. It was how both felt authoritative. My child didn't question either of them. Why would they? The answers came from a machine that sounds confident, precise, and supposedly unbiased. That moment stayed with me. It got me thinking about how artificial intelligence is not just answering questions, but subtly shaping our perception of truth.

This article is not about the fears of AI taking over the world or machines turning sentient. It is about something more subtle, yet equally dangerous: how AI is already moulding reality through control over information, and how the risks of this are not being openly discussed.

The illusion of objectivity

AI does not know things. It predicts what we might want to hear based on patterns in data. That data is selected, filtered, and cleaned by humans. Often it is scraped from the internet without transparency. As a result, AI answers are not objective truths, but reflections of the data they have been trained on.

As Kate Crawford, author of Atlas of AI, puts it: "AI is neither artificial nor intelligent. It is made from natural resources, data and human labour."

The problem becomes critical when people begin relying on these systems as default sources of truth. Subtle variations in how history is presented or politics are framed begin to shape how we think. This is not just a technological concern, it is a civic one.

Who controls the narrative?

Most leading AI systems are developed by a handful of companies: OpenAI, Google, Meta, Anthropic, and major players in China like Baidu and Alibaba. These companies control not only the architecture of the models, but also the training data and safety protocols.

The content a model deems safe or unsafe, the voices it amplifies or minimises, and the framing it provides on geopolitical topics are all shaped by internal policies. These are not neutral decisions. They are influenced by corporate values, legal risks, market pressures, and in some cases, national interests.

When you ask AI about Taiwan, Palestine or Crimea, the answers may vary depending on where the model is hosted or which markets the provider operates in. Some developers, like Anthropic, have even said they offer "region-specific alignment". That means you might get different responses in Europe, China or the United States for the same question.

This is not just localisation. It is algorithmic diplomacy.

The risk of epistemic capture

A growing number of researchers have warned about what they call "epistemic capture" - a condition where AI becomes the dominant framework for accessing and validating knowledge.

Dr Emily Bender, professor of computational linguistics at the University of Washington, argues that large language models create what she calls "stochastic parrots". They generate fluent language that mimics human reasoning, but with no grounding in fact. "The risk," she notes, "is that people begin to treat these outputs as reliable knowledge, even when they are not."

This is especially dangerous in education. Students now use AI tools to research essays, summarise historical events or understand political theory. But without awareness of how those answers are constructed, they risk absorbing distorted or decontextualised information.

AI as infrastructure for influence

The framing of knowledge is not just about history or politics. It affects everything from medical advice to financial decisions. As AI becomes integrated into operating systems, search engines, and productivity software, we stop noticing when we are interacting with it. It becomes part of the background.

And when it shapes our workflows, it shapes our thinking. Subtly. Invisibly.

The philosopher Byung-Chul Han once noted that "power is no longer exercised through prohibition, but through transparency, communication, and access". In other words, the new power lies not in censorship, but in framing. AI, in this sense, is the perfect instrument of soft control. It answers our questions in ways that appear neutral, but are never truly so.

When ethics is a marketing term

Companies are keenly aware of public concerns around AI. That is why nearly every major player promotes "ethical AI" or "responsible AI". These terms, however, are largely undefined and often serve a public relations function more than a regulatory one.

Most ethical AI initiatives are internal. Ethics boards are advisory, not binding. Transparency reports often exclude key details about training data or model behaviour. In the absence of legal frameworks, ethical language becomes a shield to deflect criticism, not a tool to drive reform.

A report by Mozilla Foundation found that most AI ethics principles released by companies lack any enforcement mechanism. "We do not need more principles," the authors concluded. "We need accountability."

Fragile foundations

One lesser-known risk is how fragile AI systems really are. They are sensitive to data shifts, adversarial prompts, and even subtle language changes. Despite their perceived intelligence, they are brittle.

A 2024 study by Stanford University found that even minor alterations to input prompts could cause large language models to generate false, misleading or dangerous information. In some cases, questions phrased with polite or formal language received safer, more accurate responses than blunt or slang-infused versions.

This inconsistency raises another issue: reliability. If we cannot predict how AI will behave across contexts, how can we trust it in critical domains like healthcare, legal advice or journalism?

Global imbalances and digital colonialism

There is a global dimension to this conversation that often goes unspoken. AI development is dominated by wealthier nations. Yet much of the data used to train models comes from the Global South. Languages, dialects, cultural expressions and online behaviours are harvested without clear consent or compensation.

Meanwhile, the benefits of AI - economic, social, technological - remain concentrated. Developing nations may end up dependent on foreign-built AI tools that do not reflect their values or context.

This is not a new pattern. As author Abeba Birhane notes, "AI risks becoming the next wave of colonial extraction, where knowledge, labour and culture are taken without reciprocity".

What we are losing

Perhaps the quietest risk of all is what we are losing without noticing.

When AI writes our emails, summarises our articles, and answers our children's questions, we slowly begin to outsource parts of our cognitive lives. We become users of knowledge, not builders of it. We trade convenience for agency.

Over time, skills atrophy. Critical thinking declines. The ability to hold uncertainty, to research across conflicting sources, to make judgments based on incomplete data, all of this weakens when we defer to the machine.

The question we should be asking is not "what can AI do for us?" but "what are we becoming as a result of AI?"

Where we go from here

None of this means we should reject AI. The technology is here, and it can be used for incredible good. But we must stop pretending it is neutral, objective or inevitable.

We need better regulation. That includes transparency requirements for how AI models are trained, accountability frameworks for their deployment, and democratic oversight for their governance. We need to teach media literacy with AI in mind. Students must learn not just to use AI, but to question it. To understand its limits; to treat it as a tool, not a source of truth.

And most of all, we need public awareness. The risks of AI are not just about machines getting smarter. They are about us becoming less aware of how our thinking is being shaped.

That moment with my child still lingers in my mind. A simple question about Hong Kong opened up a complex reflection about power, knowledge and the future of truth.

In the age of artificial intelligence, even the most innocent question may come with a hidden answer.