AI’s Democratic Paradox

A citizen may have the right to scrutinise a public budget and still lack the time, training, or money to understand it. A community may be entitled to challenge a policy and still struggle to make its case heard. AI could narrow the distance between those formal rights and the practical ability to exercise them. It can make explanation, translation, analysis, and organisation available to people who have never had a research department at their disposal.

That possibility deserves to sit at the centre of the debate about AI and democracy. So does its counterpart: the same capabilities can help a fraudster impersonate a trusted voice, a political operator manufacture apparent support, or an authoritarian institution monitor its critics. Broad access can increase both civic capacity and the capacity for abuse. Restricting access to a few organisations may contain some misuse, while giving those organisations greater power over everyone else.

The democratic paradox is therefore a problem of distributing capability and controlling power at the same time. It cannot be resolved by assuming either that the public is too dangerous to be trusted with AI or that public availability automatically produces freedom. Model design, access, economic incentives, and political institutions all shape the result.

The democratic case for AI is a case for broad public capability under accountable institutions. Embrace the technology, equip people to use and question it, and govern the harms it enables. Apply that scrutiny to private abuse, corporate concentration, and state power alike.

What it means to empower a citizen

Democratic empowerment has at least three dimensions: the ability to understand and act; the ability to contribute knowledge that others lack; and the ability to challenge decisions made by more powerful actors. A system that answers millions of questions might advance the first while weakening the other two. Its democratic value depends on whether people can inspect its evidence, introduce their own experience, disagree with its conclusions, and obtain a meaningful response when it helps determine their lives.

Better information can improve a debate without settling it. People can agree on the likely effects of a policy and still disagree about fairness, obligations, or acceptable trade-offs. A democratic assistant should help make those disagreements intelligible. Treating political conflict as a calculation with one correct answer would give the model’s designers authority over questions citizens must decide together.

Friedrich Hayek’s account of dispersed knowledge provides one reason to favour broad access. Useful knowledge is distributed across people and circumstances; no central organisation possesses it all.¹ Applied to AI, this suggests that teachers, local journalists, researchers, small businesses, and community groups can discover applications and failures that a handful of developers will miss. This is an argument about the value of distributed experimentation. It does not establish that every release decision is beneficial, or that a market left alone will distribute the benefits fairly.

Philip Pettit’s account of freedom adds another test: whether people live subject to someone else’s arbitrary power.² Applied here, a useful AI service can still create dependence if its provider can silently change what users may investigate, monitor their inquiries, or withdraw access without effective challenge. Public access becomes more meaningful when people have alternatives, privacy, and enforceable ways to contest decisions. Access to an answer and power over the conditions under which answers are produced are different achievements.

AI as a force for democratic capability

Consider what a well-designed civic assistant could enable. A resident could compare a council’s spending claims with the underlying figures. A local newsroom could search procurement records and identify inconsistencies for reporters to investigate. A disability organisation could turn a lengthy consultation into accessible formats, gather members’ objections, and trace each proposed change to the original text. These are plausible applications, whose accuracy and practical value require evaluation. Their democratic promise lies in making participation and scrutiny less expensive.

Educational research offers evidence for part of that promise. Kestin and colleagues found that a carefully designed AI tutor improved immediate learning relative to an active-learning classroom in a study involving 194 university physics students.³ In a separate field experiment involving nearly 1,000 mathematics students, Bastani and colleagues found that a general chat interface improved assisted practice but reduced subsequent unassisted performance; a tutor designed around teacher-provided hints largely avoided that harm.⁴ Together, these findings support attention to pedagogy and independent performance. They leave open how far the benefits extend across subjects, populations, and time.

There is also evidence closer to democratic life. In a preprint revised in June 2026, Luettgau and colleagues report randomised trials involving 2,858 participants: using conversational AI to research specified political topics improved factual knowledge about those topics to a similar extent as self-directed Google search.⁵ The result supports AI’s potential as a political research aid within that setting. It does not establish superiority to search, lasting resistance to manipulation, or better electoral decisions. The paper remains a preprint, and its structured tasks leave many everyday uses untested.

AI may also support people reasoning together. Tessler and colleagues studied an AI mediator that drafted group statements from participants’ views and revised them in response to criticism. Across experiments involving more than 5,000 UK participants, people preferred its statements to those produced by human mediators, and the process helped groups find common ground.⁶ This is promising evidence for assisted deliberation. Agreement, however, is only one democratic good. Minority rights, the freedom to dissent, and public control over the questions being asked must remain visible even when a system can produce an attractive consensus.

These studies measure different outcomes: immediate learning, unassisted performance, political knowledge, and endorsement of a group statement. They cannot be combined into a general ranking of AI’s ability to educate versus its ability to persuade. The defensible conclusion is that several useful civic functions are possible, and that their design and evaluation matter. A democracy should be interested in whether AI helps people ask better questions, recognise uncertainty, and act together with greater understanding.

The same capabilities can work against democracy

AI expands the resources available to bad actors. Language generation can make deceptive campaigns cheaper to adapt; synthetic media can facilitate impersonation; automated systems can help institutions classify and monitor people. AI can also produce harm without a malicious user, through fabricated information, biased outputs, or misplaced confidence. NIST’s generative AI risk profile identifies these overlapping risks, including information integrity, privacy, and abusive content.⁷ Calling AI a “bad actor” captures its capacity to participate in harmful conduct, but responsibility still has to be assigned to the people and institutions that design, deploy, and authorise it.

Persuasion is a legitimate part of democratic debate; deception and coercion are what make its use abusive. Experiments establish a real persuasive capability, with important boundaries. Salvi and colleagues found that GPT-4 supplied with personal information could be more persuasive than human opponents in structured online debates involving 900 participants.⁸ In a larger study published in Science, Hackenburg and colleagues tested 19 models across 707 political issues with 76,977 participants. Post-training and prompting had larger effects on persuasion than personalisation or model scale, and methods that increased persuasion also reduced factual accuracy.⁹ Design choices therefore belong inside the explanation of harm. Technical systems arrive with objectives and defaults that can favour informed understanding or mere agreement.

Neither study measures control over an electorate. A shift in reported agreement during an experiment differs from a durable belief, a vote, or a change in public policy. Real influence also depends on reach, trust, repeated exposure, and competing messages. Yet low reach can coexist with severe harm: an impersonation aimed at an employer or a fabricated intimate image circulated within a community can make public participation costly for the person targeted. Democratic damage includes who feels able to remain visible and speak.

Fabrication also gives dishonest actors a way to dispute authentic evidence. Experiments by Schiff and colleagues found that false claims of misinformation could preserve support for politicians facing scandal, principally against text reports; those denials were largely ineffective against video evidence in their studies.¹⁰ The “liar’s dividend” is a documented, context-dependent threat. It strengthens the case for preserving evidence and credible investigation without implying that every denial succeeds or that the public has lost all capacity to distinguish truth.

Concentration is a democratic risk of its own

Restricting AI to a small number of organisations has a serious argument in its favour. A service operated centrally can limit requests, monitor abuse, revoke access, and update safeguards. Concentrated resources can fund expensive research and security. These advantages should be weighed against what concentration allows the gatekeepers themselves to do.

A provider that controls a widely used route to information can influence which sources are visible, which uses are affordable, and whose applications survive a policy change. A government with privileged access to advanced analysis may gain further advantage over the people trying to scrutinise it. These are risks of dependence and unequal power even before anyone deliberately manipulates an answer. In its 2024 foundation-model review, the UK Competition and Markets Authority identified risks arising from control of essential inputs, routes to market, and partnerships among powerful firms.¹¹

Confining development to a few organisations would not necessarily stop technical progress. Large institutions can innovate rapidly. The stronger objection is that exclusion can narrow who experiments, whose needs are served, and who can independently test the resulting systems. It may also make safety knowledge dependent on the organisations being assessed. Public research, independent evaluation, and a diversity of providers help make claims about performance and safety contestable.

Public availability alone does not dissolve concentration. An open model still requires computing resources, relevant data, skills, and a route to users. A well-funded organisation may gain more from the same model than an under-resourced community does. Millions of people consulting systems with shared weaknesses can also reproduce the same error: multiple answers do not necessarily constitute independent confirmation. The democratic objective should be effective access to a diverse information environment, supported by institutions capable of challenging it.

Openness has several forms, and different consequences

A public chatbot, access for independent researchers, the ability to adapt a model, and downloadable model weights are distinct forms of access. Weights are the learned parameters that allow others to run a model themselves. Distributing them can support local adaptation, independent experimentation, and use without sending every inquiry to a central provider. Once widely distributed, however, they are difficult to recall, and a provider cannot reliably enforce its original safeguards on modified copies.

That distinction makes the choice more demanding than “open” or “closed.” The relevant comparison is the additional benefit and harm of a particular form of access relative to realistic alternatives. In July 2024, NTIA recommended against immediate restrictions on widely available model weights while calling for evidence collection and the capacity to respond if risks justified intervention.¹² That was a judgement about the evidence at the time. Future systems and release decisions require their own assessment.

The presumption in favour of broad civic access can coexist with targeted constraints on capabilities that create credible risks of severe harm. Their justification should identify the capability, the plausible damage, the effectiveness of the proposed restriction, and its cost to legitimate users. Public-interest access should also be supported through education, accessible design, local-language resources, and affordable infrastructure. Otherwise, nominal openness can coexist with practical exclusion.

There is no established law that AI helps defenders more than attackers, or that additional public analysis will cancel additional deception. Attackers can concentrate effort on a vulnerable target; citizens have limited attention; powerful institutions possess advantages in data and distribution. Broad access gives the public a better opportunity to investigate and respond. Collective capacity in journalism, libraries, universities, and civil society helps turn that opportunity into something durable.

What history can teach us

The historical intuition behind embracing technology is persuasive: greater access to knowledge and productive tools can expand what people are able to do. But the claim that civilisation has consistently advanced by adopting technology is too broad to serve as evidence for a particular AI policy. Adoption, distribution, political rights, and institutional reform are separate processes. New capabilities can strengthen domination as well as emancipation.

Research on radio in prewar Germany makes the danger concrete. Adena and colleagues found that radio’s political effects changed as control and content shifted: broadcasts initially slowed the growth of Nazi support, then supported it after the Nazis gained control.¹³ This is evidence from a particular historical setting, not a prediction about AI. It shows why access to a powerful medium and control over what it carries must be examined together.

A defensible lesson from history is therefore a political commitment: expand people’s capabilities while building institutions that prevent those capabilities from becoming instruments of domination. For AI, that means pursuing public empowerment deliberately. Neither technical progress nor regulation guarantees it. Regulation can protect citizens, but it can also entrench incumbents or arm governments with new powers over speech.

Governance must strengthen public agency

Elinor Ostrom’s work on polycentric governance offers a useful starting point: complex problems can be governed through several centres of responsibility, with rules suited to their circumstances and mechanisms for coordination.¹⁴ Applying that idea to AI is a proposal, rather than a result established by her research. It suggests that developers, public authorities, independent researchers, courts, and civic organisations should have distinct responsibilities and the ability to check one another. Fragmented responsibility still needs enforceable obligations and a clear place for an affected person to seek remedy.

For civic uses, the design objective should be informed agency: explanations linked to inspectable evidence, uncertainty made visible, and opportunities to challenge an answer. Evaluation should test whether people understand more, detect errors, and make better-supported decisions, including after assistance is withdrawn when the objective is learning. For public institutions using AI in consequential decisions, citizens need understandable reasons, meaningful human review, and an effective route to appeal. The authority of a decision must remain open to challenge.

For abuse, governance should address deceptive impersonation, covert sponsorship, discriminatory surveillance, and targeted harassment while protecting lawful persuasion, satire, dissent, and anonymous speech. Responses need to reflect both scale and severity. Preserving original evidence, making reporting channels usable, and providing timely remedies can matter as much as reducing the number of synthetic posts.

Technical provenance can assist that work. C2PA Content Credentials provide evidence about an asset’s recorded origin and editing history. They do not establish that its claims are true; missing credentials do not establish that it is false, and credentials can be removed.¹⁵ Verification still requires judgement about sources and corroboration. Asking every citizen to authenticate every image unaided would leave too much of the burden with the least-resourced participant.

Governance must also address its own concentration of power. Competition policy, support for independent research, proportionate obligations for smaller providers, and scrutiny of state procurement can preserve alternatives. Restrictions should have public reasons, independent review, and opportunities for revision as evidence changes. A policy that prevents some misuse while disabling legitimate scrutiny may weaken the democracy it claims to protect.

Success would mean more than widespread adoption or fewer reported incidents. It would mean that people across languages, incomes, and abilities can use AI to participate; that independent organisations can investigate its failures; that victims can obtain remedy; and that citizens can challenge institutions deploying it. These are outcomes to measure over time. The evidence will sometimes require changing a design, a deployment, or the conditions of access.

AI’s democratic promise is substantial: more people could gain the analytical and organisational capacity to shape decisions that affect them. Realising that promise requires confidence in people alongside realistic expectations of the tools and institutions around them. Broad access and accountable governance belong in the same project.

A democratic AI future should give people greater power to understand, to create, to organise, and to hold power accountable.

Disclosure: This Insight is based entirely on public information and does not represent the views of the Government of Canada. Maple Quanta Inc. provides AI governance and assurance services; readers should consider that commercial context. This article is for general informational purposes and does not constitute legal, electoral, cybersecurity, or regulatory advice.