AI is not democratic unless it redistributes power

Democracies worldwide are facing mounting concern over deepfakes, misinformation, algorithmic polarization, automated surveillance and the rising control of public information by a small number of technology platforms.

A new study, When is AI democratic? Artificial intelligence and democratic empowerment, published in AI & Society, says democratic AI should be judged by whether it redistributes power and creates more balanced relations between citizens, governments, companies and technological systems.

Why participation alone does not make AI democratic

The paper identifies four ways in which the term democratic AI is commonly used: majority rule, participation, deliberation and empowerment. Majority-rule approaches seek to align AI systems with broad public preferences or consensus while participatory approaches focus on including more people in AI design and governance. Deliberative approaches use AI to support better public debate, reduce division or help citizens weigh policy options.

According to the study, they are incomplete unless they change the underlying distribution of power. A system that collects public opinion but leaves all decisions to private firms, government agencies or technical experts may look democratic without being meaningfully democratic. A consultation tool may invite thousands of citizens to respond but still keep authority concentrated among those who design the process, control the data or interpret the results.

AI can summarize public comments, translate speeches, transcribe parliamentary debates, organize citizen discussions and help users understand complex policy issues. These applications may improve access to politics and public institutions. However, the study points out that the real question is whether citizens gain influence or merely become data sources for systems controlled elsewhere.

The answer is based on participatory democracy, a tradition that sees democracy not only as elections and legal procedures but as the wider redistribution of power across society. Under this view, democracy is not confined to parliaments, voting booths or courts. It also concerns workplaces, schools, hospitals, digital platforms, public debate and the social systems through which people gain or lose agency.

The framework allows the analysis of AI to move beyond election interference and political campaigning. AI increasingly affects how people learn, work, receive healthcare, encounter news, access public services and understand themselves. It can widen opportunity, but it can also create dependence, opacity and control. The study argues that democratic AI must be assessed across these everyday power relations, not only in formal political settings.

This is why empowerment becomes the key standard. In the paper's framework, AI is democratic when it makes power relations more symmetrical. It is not democratic when it strengthens one-sided control, hides decision-making, worsens dependence or allows a narrow group of actors to shape social outcomes without meaningful public challenge.

Design, data and ownership are central to the democratic test

The study divides the democratic assessment of AI into two dimensions: how AI is made and how AI acts in society. The first includes development, design, governance, political economy and public narratives. The second examines how AI systems operate once deployed in education, healthcare, information and politics.

The making of AI is not politically neutral. Data collection, labeling, model training, evaluation, deployment and user interface design all involve choices that can either widen or restrict power. Whose data is included, whose interests shape the system, who can understand its decisions, who can challenge its outputs and who profits from its use all become democratic questions.

A more democratic AI development process would involve broader public participation, stronger accountability and clearer limits on system power. It would also require attention to accessibility. AI systems that only highly skilled users, wealthy institutions or dominant firms can understand and operate risk reinforcing existing inequalities. Systems designed with accessible interfaces, lower costs and broader public benefit can move power in a more democratic direction.

Accountability is one of the paper's strongest concerns. Users affected by AI must be able to question how systems work, why decisions are made and what limits apply to automated action. Without that ability, AI can create a severe power imbalance between system operators and ordinary citizens. A person denied a service, misclassified by a model or guided by an opaque recommendation system may have little ability to contest the outcome if the system is treated as technically authoritative and beyond public understanding.

Opacity further deepens that imbalance. The black-box nature of many machine learning systems can protect corporate secrecy, shield government use and limit public scrutiny. Even explainable AI tools may not fully solve the problem if they only show which inputs influenced an output without revealing the deeper logic of a decision. For democratic empowerment, explanation must support real contestation, not simply provide the appearance of transparency.

Advanced AI is increasingly shaped by a small number of powerful firms with control over infrastructure, data, computing power and platform ecosystems. That concentration creates a democratic problem even when individual tools appear useful. If the future of AI is built and governed mainly by private actors with limited public accountability, power will continue to move away from citizens and toward corporate centers of control.

Public narratives around AI also affect democratic outcomes. When AI is presented mainly as an engine of productivity, competition and efficiency, public debate can become narrow and technocratic. That framing makes it easier for experts, executives and state institutions to dominate the agenda. When AI is treated as a technology with broad social consequences, more people and institutions can claim a legitimate role in shaping its development.

The study's framework thus pushes policymakers and developers to ask harder questions. A design choice is not only technical. A dataset is not only a resource. A governance model is not only a compliance structure. Each of these decisions affects who holds power, who can act and who can challenge decisions.

AI can empower citizens, but it can also deepen dependence

The paper's second dimension, AI action, examines how AI changes power relations once it is used in society. The study focuses on education, healthcare, information and politics to show AI's double edge.

In education, AI can support democratic empowerment by making learning more personalized and accessible. Systems that adapt to student needs may help learners who are underserved by standard classroom models. AI tutors and learning tools may reduce barriers for students who need individual support, language assistance or flexible instruction.

However, the same technology can also weaken democratic agency if it reinforces bias, reduces human judgment or makes education more dependent on machine-readable performance. Students may become subject to systems they cannot understand or contest. Teachers may lose professional authority if AI tools are treated as superior decision-makers. The democratic value of AI in education depends on whether it expands human capacity or narrows it.

In healthcare, AI can improve access to diagnosis, telemedicine, patient guidance and personalized treatment, especially for people in remote or underserved areas. It can help patients understand medical information and reduce dependence on overloaded systems. It may also cut administrative burdens and improve service delivery.

Healthcare AI can become anti-democratic when patients and regulators cannot scrutinize the logic behind clinical decisions. Bias, weak regulation, liability gaps and data exposure can all shift power away from patients. In health systems, where automated decisions may affect life, treatment and access to care, accountability is not optional. It is central to whether AI empowers people or subjects them to hidden authority.

In the information sphere, AI can either strengthen public debate or damage it. Engagement-driven recommendation systems can amplify division and reward inflammatory content. Automated systems can shape what citizens see, what they believe is important and which voices gain visibility. That gives AI a major role in structuring the public sphere.

The paper points to alternative possibilities. AI systems can be designed to support mutual understanding, bridge political divides, detect automated manipulation and help users identify trustworthy information. Such tools may increase citizens' ability to navigate digital spaces. But their democratic value depends on whether they make people more informed and capable, not simply more managed.

The clearest evidence comes from politics. AI can make legislative processes more accessible by transcribing debates, translating speeches, summarizing policy materials and helping citizens understand complex decisions. Civic platforms can use AI to support public consultation and wider participation. Political organizers may also use AI to draft materials, prepare campaigns and reduce the time burden of activism.

These uses could lower barriers to political participation, especially for people who lack time, money, technical expertise or institutional access. If AI helps more citizens understand and influence political decisions, it can support democratic empowerment.

The paper warns that AI-enabled participation can also become tokenistic. A platform may gather public input while leaving real power untouched. A government may use AI to display openness without changing decision-making. A company may present participatory AI development as democratic while retaining control over design, deployment and profit.

Generative AI creates a further tension as its wide accessibility gives ordinary users new tools to write, organize, research and create, strengthening agency. However, it may also increase dependence on automated outputs, weaken critical thought and widen the gap between those who know how to use AI effectively and those who do not. The technology can support democratic action or domesticate it.