Digital inequality now goes beyond internet access as AI reshapes social exclusion

Artificial intelligence (AI) is no longer only widening the gap between those who have technology and those who do not. It is creating deeper forms of social exclusion through automated decisions that shape access to healthcare, jobs, education and public services, according to a new study by a team of researchers from Ecuador.

The study, titled From the Digital Divide to Algorithmic Vulnerability: A Systematic Review of Social Stratification in the AI Era (2015–2025), was published in Social Sciences. Based on a systematic review of 74 high-impact records from Scopus, Web of Science, ProQuest and PsycINFO, the paper finds that inequality has shifted from basic access to digital tools toward algorithmic vulnerability, where opaque AI systems can reproduce and deepen social hierarchies.

Digital inequality has moved beyond access to technology

According to the review, the traditional understanding of the digital divide is no longer enough to explain inequality in the AI era. Earlier debates focused mainly on who had internet access, who owned devices and who could afford connectivity. That first-level divide remains relevant, especially for low-income groups, older adults, migrants, racial and ethnic minorities and people with disabilities. But the authors find that the central problem has expanded.

The new divide is shaped by whether people can understand, challenge and avoid harm from algorithmic systems that increasingly mediate social life. Access to a smartphone or an internet connection does not guarantee equal participation when automated tools screen job applicants, classify health risks, recommend educational content, rank welfare eligibility or shape exposure to information.

The paper uses the term algorithmic vulnerability to describe this new exposure to harm. It refers to the risk faced by individuals or groups when automated systems make biased, opaque or discriminatory decisions that affect their opportunities. In this model, exclusion is not only caused by lack of access. It is also caused by automated classification.

The review found that scientific work on this issue expanded sharply after 2018. The analyzed corpus showed an annual growth rate of 71.7 percent, with much of the recent work concentrated in 2024 and 2025. The authors interpret this as evidence that social scientists are rapidly shifting attention from connectivity gaps to the social effects of algorithmic decision-making.

In earlier digital policy debates, the main goal was to connect the unconnected. In the AI era, the challenge also includes protecting people who are already inside digital systems but are classified, scored or filtered in unfair ways. The authors link this transition to social reproduction theory, arguing that AI systems can convert existing social advantages into new digital privileges. Groups with higher digital literacy, stronger economic resources and better institutional access are more likely to benefit from AI systems. Groups with less power are more likely to be misclassified, excluded or left without appeal mechanisms.

A key concern is opacity. Many automated systems make decisions in ways that are difficult for affected people to understand. When a loan is denied, a job application is filtered out, a health risk score is assigned or a student is tracked by educational software, the person affected may not know why the decision was made or how to contest it.

The review also distinguishes AI literacy from algorithmic literacy. AI literacy focuses on the ability to use or understand AI tools, including generative models. Algorithmic literacy is broader. It includes understanding how data classification, ranking and filtering systems shape choices even on platforms that may not use advanced AI. This distinction matters because algorithmic discrimination did not begin with ChatGPT or generative AI. It has been built into digital systems for years.

The authors argue that the AI era has made these older problems more powerful. Automated systems now operate at larger scale, with greater authority and across more consequential sectors. Their results can appear objective even when they are built on biased data, narrow assumptions or institutional priorities that favor dominant groups.

Health, labor and education emerge as key sites of AI exclusion

The review identifies three major sectors where AI is producing new forms of exclusion: healthcare, labor markets and education. These areas are critical because automated decisions in them affect well-being, income, mobility and access to rights.

Healthcare emerged as the strongest area of concern among the three sectors. The review finds that AI systems in health can reproduce inequities when they rely on incomplete, biased or historically unequal data. Algorithms used for triage, diagnosis, resource allocation or risk prediction may underrepresent vulnerable groups or use proxy variables that reflect structural disadvantage.

The issue is not only whether an AI system is accurate on average. A system may perform well overall while failing specific communities. If training data underrepresent ethnic minorities, low-income patients or people in under-resourced regions, AI-driven tools may deliver lower-quality predictions for the people who already face health barriers.

The review also highlights the risk of data extraction in global health. Low- and middle-income countries may provide data that help train AI systems but receive less benefit from the resulting technologies. This creates a form of global stratification in which data flow from vulnerable populations while high-quality digital care remains concentrated elsewhere.

In the labor market, AI is changing how people are hired, monitored and evaluated. Automated recruitment systems can screen candidates based on patterns that reflect past discrimination. Performance algorithms can penalize workers who do not match standardized productivity profiles. Surveillance systems can intensify control over workers, especially in low-wage or platform-based employment.

The paper warns that automation can deepen a divide between those who own, design or manage AI systems and those whose work is reorganized by them. Workers with high levels of digital capital may gain new opportunities, while low-skilled or precarious workers may face displacement, tighter monitoring or reduced bargaining power.

The authors also note that algorithmic management can diminish the perceived value of human labor. When social, emotional or service-oriented work is treated as measurable and automatable, workers may lose status even when their tasks remain socially important. This is especially relevant in sectors where women, migrants and minority workers are overrepresented.

Education forms the third major site of algorithmic vulnerability. AI tools can personalize learning, support assessment and expand access to resources. But they can also create a new gap between students who can use AI as a tool for co-creation and those who lack access, guidance or algorithmic literacy.

The review finds that education-related AI risks include data privacy concerns, biased assessment systems, uneven access to advanced tools and growing dependence on platforms that shape what students see and learn. Students with greater digital confidence may benefit, while others may be left behind or misjudged by automated systems.

The authors also point to the risk of educational tracking through algorithmic systems. If AI tools rank students, recommend pathways or predict performance based on biased data, they may reinforce existing class, racial, geographic or disability-related inequalities. The danger is that educational disadvantage becomes hidden inside technical systems that appear neutral.

Apart from these three sectors, the review identifies emerging concerns in cultural participation, welfare systems and media consumption. Recommendation algorithms can narrow exposure to culture and information, while automated public administration can exclude citizens who lack the digital capital needed to navigate online systems. These areas were less central in the reviewed literature but are flagged as important frontiers for future research.

Across sectors, the common mechanism is not a simple technical failure - it's the use of automated systems in unequal societies without sufficient transparency, accountability or affected-community participation. AI does not create inequality from nothing. It can scale and formalize patterns that already exist.

Researchers call for algorithmic justice and human-centered governance

The review finds that the study of algorithmic discrimination is also changing. Quantitative auditing dominates the field, accounting for 58 percent of the reviewed studies. These approaches often use black-box testing to identify statistical disparities in automated systems without access to the source code.

Such audits are important because many AI systems are proprietary and closed to public scrutiny. Researchers can still test whether outputs differ across groups, whether proxy variables hide race or gender bias, and whether systems produce unequal results in areas such as hiring, credit, health and surveillance.

The review also identifies growth in mixed sociotechnical methods since 2021. These approaches combine quantitative analysis with digital ethnography, interviews and community-focused research. The authors argue that this shift is necessary because algorithmic discrimination is not only a numerical problem. It is also lived as uncertainty, fear, exclusion and lack of recourse.

The paper identifies four main mechanisms through which AI deepens social stratification. The first is opacity and power asymmetry. When systems cannot be understood or challenged, affected citizens are placed in a weak position. The second is data exclusion, especially in healthcare, where missing or biased data can make vulnerable groups less visible or less accurately served.

The third is capital concentration and labor precarity. AI can increase the power of those who own infrastructure and data while reducing the bargaining power of workers affected by automation. The fourth is infrastructure and governance barriers. Advanced AI systems require capital, expertise and technical infrastructure, which means powerful institutions are better positioned to shape the AI future than communities most exposed to harm.

Current regulatory debates are moving from voluntary ethics toward binding rules. The European Union Artificial Intelligence Act is identified as a major example because it classifies systems affecting fundamental rights as high risk and requires stronger oversight. The authors argue that bias mitigation can no longer be treated as optional corporate responsibility.

Regulation must move beyond technical compliance, the paper argues, calling for algorithmic justice and human-centric governance that keep people in control of decisions affecting rights, opportunity and welfare. In this approach, judges, doctors, teachers, employers and public officials must have enough algorithmic literacy to challenge automated outputs rather than passively accept them.

The study also highlights the need for what it describes as a Reinstating AI framework. Rather than using AI mainly to replace human labor or concentrate power, this approach would direct innovation toward augmenting human capabilities. It would support technologies that strengthen workers, students, patients and citizens instead of turning them into data points inside systems they cannot influence.

Democratizing technological development is vitally important. This includes more diverse design teams, stronger use of data from the Global South, data cooperatives, public-interest governance and greater participation by affected communities. Without these steps, AI systems may continue to reflect the priorities of dominant technology centers while marginalizing other regions and populations.

Lastly, the authors acknowledge several constraints in the review. Its PRISMA-based criteria may have excluded relevant grey literature, technical reports or human rights documentation. The recency of the corpus also limits the ability to assess long-term effects of new AI regulations. The authors call for more longitudinal research, more studies in underrepresented regions and more work on how generative AI may widen the cognitive divide.