The greatest threat to our freedom may not be Artificial Intelligence (AI) but the political institutions determined to control what it allows us to know. For centuries, kingdoms and governments warned their citizen about dragons beyond the castle walls. The dragons have taken many forms – foreign enemies, dangerous ideologies, economic instability, misinformation and more recently, AI.
The message is remarkably consistent. Beyond the walls lies danger. Within them lies security. Trust those who govern the castle, surrender a little freedom, and the dragon will be kept at bay. It is a persuasive narrative. It is also one of the oldest instruments of political power. But what if the dragon is not beyond the castle walls? What if it is already inside, comfortably seated upon a mountain of gold, guarding the privileges, authority and institutional power accumulated by those that govern? And what if AI, rather than being the dragon, represents something far more threatening to the castle inhabitants – a means by which ordinary citizens might finally see what is happening inside?
This probably needs considerably more attention than it currently receives. The democratisation of knowledge.
The arrival of large language models (LLM) represents a potentially profound redistribution of intellectual power. For much of human history, knowledge has been concentrated within institutions. Governments, universities, professional organisations, and established media have enjoyed substantial influence over what information is available, how it is interpreted, and which conclusions are considered legitimate. AI threatens to disrupt this arrangement. A sufficiently capable LLM can examine legislation, compare competing economic theories, interrogate public policy, identify contradictions in political arguments, and explain complex scientific and philosophical questions to an ordinary citizen. It does not require the citizen to possess a university degree, employ a research assistant or belong to an influential institution.
The implications are extraordinary.
A grazier in regional Queensland could potentially interrogate the economic assumptions underlying an agricultural regulation with a sophistication previously available only to specialist analysts. A small business owner could scrutinise the reasoning behind taxation policy. A voter could compare political promises against legislative outcomes, examining not merely what politicians proclaim in their stump speeches, but whether their policies withstand serious intellectual scrutiny. This is not simply technological progress. It is the potential democratisation of intellectual authority. And that creates a problem for those accustomed to exercising it.
There is a phrase that captures the danger at the centre of this emerging debate – ‘Protect the building, not the people’. It describes an institutional temptation that can arise whenever the preservation of an organisation becomes more important than the purpose for which it was established. Governments are created to serve citizens. Yet governments are also institutions populated by individuals with careers, reputations, ambitions and political interests. The distinction matters.
A policy that protects citizens may not necessarily protect the political interests of those governing them. Equally, a policy that protects political authority may do very little to advance the interests of the public. AI introduces a particularly uncomfortable possibility. What happens when citizens acquire the intellectual tools to challenge the explanations upon which political authority depends? What happens when government claims can be interrogated instantly, repeatedly and without deference to institutional prestige? And what happens when an LLM produces an answer that is factually defensible but politically inconvenient?
The temptation will be to regulate the answer rather than confront the argument. Not necessarily by ordering the machine to lie. That would be crude, conspicuous and potentially counterproductive. A more sophisticated approach would be to determine which questions may be answered, which conclusions may be presented, and which subjects require institutional supervision. The result could be a system that appears intellectually capable but is selectively constrained. An extraordinarily knowledgeable machine that knows when not to speak.
None of this means AI is harmless. LLMs can produce false information, facilitate fraud, expose private information, and assist in genuinely dangerous activities. Sensible safeguards are necessary, particularly where identifiable harm to individuals is foreseeable. But a distinction must be drawn between protecting people from harm and protecting institutions from scrutiny. These are not the same moral enterprise. Preventing an AI system from providing instructions for manufacturing a biological weapon is fundamentally different from preventing it from critically examining government policy, institutional misconduct or the intellectual foundations of a politically contentious position. The former protects human life. The latter may protect political authority. Yet both restrictions could potentially be justified under the expansive language of public safety. This is where the philosophical danger emerges.
Once governments acquire the authority to determine which truthful information is too dangerous for citizens to receive, the question is no longer simply one of technological governance. It becomes a question of political liberty. Who decides what constitutes dangerous knowledge? Who determines whether an answer is misleading, socially harmful, or contrary to the public interest? And most importantly, who scrutinises the people making those decisions?
Traditional censorship has often been visible. A newspaper is prohibited. A book is banned. A speaker is silenced. The public can identify the restriction and, at least in principle, challenge it. AI introduces a subtler possibility. An LLM might simply decline to answer a question, omit relevant information or present a carefully constrained account of a controversial issue. The citizen may never know what has been excluded. There need be no dramatic announcement, no confiscated publication and no identifiable censor. The restriction is embedded within the architecture of the system itself. This is potentially more consequential than traditional censorship because the user may mistake an institutionally restricted answer for a complete and impartial account of reality. The danger is not merely that the machine might tell a lie. It is that the machine might tell the truth selectively, while concealing the fact that a selection has occurred. Such a system would be particularly valuable to any institution seeking to preserve its authority without appearing to suppress dissent. And the concern extends beyond governments. Technology companies themselves possess considerable power to shape the intellectual boundaries of the systems they develop. The combination of corporate control and political regulation could produce an information environment in which neither the citizen nor the machine is fully free to interrogate authority.
The classical liberal tradition rests upon a simple insight – political power must be constrained because those who exercise it are no less susceptible to ambition, self-interest and error than the citizens they govern. The same principle should apply to AI governance. Regulation should address demonstrable harm, not merely uncomfortable conclusions. Restrictions should be transparent, contestable and subject to independent scrutiny. Above all, governments should not be permitted to define political inconvenience as technological danger.
The ultimate test of an AI governance regime will not be whether it prevents machines from producing answers that powerful institutions dislike. It will be whether citizens retain the freedom to ask difficult questions and receive truthful, comprehensive answers, even when those answers challenge the people responsible for governing them. For the first time in history, humanity may be approaching a world in which sophisticated intellectual inquiry is available to almost everyone. That possibility should excite anyone who believes political authority derives its legitimacy from an informed citizenry.
It may be considerably less exciting for those whose authority depends upon citizens remaining inadequately informed. So, when the guardians of the castle warn us about the AI dragon approaching from beyond the walls, perhaps we should ask a different question.
Are they protecting us from the dragon, or protecting the dragon from us?
