AI Is Scaling Access and Inequality Together

IEEE Spectrum has reframed the AI divide around a sharper conflict: global adoption is accelerating, while control remains concentrated among a few countries and companies. The concern is no longer limited to whether people can open an AI application. It now includes who owns the computing infrastructure, builds the models, represents local languages, and sets the rules.

That distinction matters because AI is entering systems that distribute opportunities and public services. It helps screen job applicants, recommend educational material, support medical decisions, and process government requests. People with strong digital skills can use these systems to extend their capabilities. Others are more likely to encounter AI as an opaque decision-maker they cannot inspect or challenge.

The IEEE Spectrum analysis presents this imbalance as a new stage of the digital divide. The central contest is between the promise of broadly available AI and the reality of narrowly held production power. Access to a chatbot can spread quickly, but meaningful participation requires infrastructure, skills, relevant data, and institutions capable of governing imported systems.

This is not an argument for abandoning cloud platforms or slowing every AI deployment. It is a warning that adoption and agency are different outcomes. A country can have millions of AI users while remaining dependent on foreign computing capacity, external model providers, and standards developed elsewhere.

The next phase of AI inequality will therefore be measured by more than user counts. The more revealing questions concern who can build useful systems, who can adapt them safely, and who has the authority to decide what they optimize.

IEEE Spectrum Moves the Digital Divide Beyond Access

The important change is conceptual: using AI does not necessarily give a person or country influence over AI.

Earlier digital-divide debates often focused on physical access. Policymakers counted internet connections, household computers, mobile subscriptions, and broadband coverage. Those measurements remain important because an unreliable connection still blocks participation before any AI question begins.

AI adds several layers above basic connectivity. Modern systems depend on computing capacity, specialized chips, cloud services, training data, technical expertise, and sustained institutional oversight. Weakness in any layer can turn an apparent deployment into long-term dependence.

A school, for example, might give students access to a generative AI assistant. That does not mean teachers understand its limits, administrators can audit its treatment of student data, or local researchers helped evaluate its behavior. The institution has access, but it may lack the capacity to govern the tool.

The same distinction applies to healthcare. A clinic can subscribe to an AI-supported service without possessing the data infrastructure or technical staff needed to test it locally. If the system performs poorly across regional languages or medical conditions, clinicians may struggle to identify the source of the problem.

The World Bank describes inclusive AI development through four foundations: connectivity, compute, context, and competency. Connectivity covers energy and digital networks. Compute includes chips, data centers, and cloud services. Context refers to relevant data, while competency covers the skills required to develop and use systems effectively.

Its AI foundations report says low- and middle-income countries face steep barriers to deploying AI at scale. It also points to smaller, task-focused systems running on common devices as a practical alternative. That approach does not erase the divide, but it broadens the available strategy.

This framework explains why distributing consumer accounts cannot solve the entire problem. A free interface can reduce the immediate cost of experimentation. It cannot supply dependable electricity, local-language evaluation data, technical training, procurement expertise, or an effective appeals process.

The divide also appears inside wealthy countries. Two employees may use the same AI assistant while receiving very different benefits. One has domain expertise, time for experimentation, and an employer that provides training. The other faces automated monitoring, unclear policies, and little authority over how the tool affects performance reviews.

That gap turns AI literacy into more than prompt-writing ability. Practical literacy includes recognizing uncertainty, checking sources, protecting sensitive information, and knowing when human review is necessary. It also includes understanding where an AI system enters a decision and how someone can contest its output.

Knowledge workers already face this problem when AI tools generate confident but unsupported statements. A personal knowledge management practice can help individuals preserve sources and context. However, individual organization cannot substitute for institutional accountability when AI influences employment, credit, healthcare, or public benefits.

The immediate news is therefore a shift in what counts as inclusion. An adoption metric shows who touched the technology. It does not show who obtained durable economic value, who shaped the system, or who absorbed its mistakes.

Compute Concentration Turns Adoption Into Dependency

AI infrastructure is spreading far more slowly than AI interfaces, creating a structural advantage for the places that already control cloud and chip capacity.

Stanford University’s 2026 AI Index reports that the United States hosts 5,427 data centers, more than ten times the total in any other country. The same report says one company, TSMC, fabricates almost every leading AI chip. These concentrations connect global AI development to a narrow physical and commercial base.

The 2026 AI Index also shows how much frontier development remains concentrated. Industry produced more than 90 percent of notable frontier models during 2025. The United States recorded far more private AI investment than China, although Stanford cautions that private investment figures understate Chinese government-backed activity.

These figures do not mean every country needs a frontier-scale model or a domestic semiconductor industry. Training the largest general-purpose systems is only one form of participation. Countries can create specialized models, local datasets, evaluation methods, public-interest applications, and rules suited to their communities.

However, infrastructure concentration changes the terms under which those alternatives develop. Organizations that rent foreign cloud capacity remain exposed to provider policies, currency movements, network reliability, data-transfer rules, and geopolitical restrictions. Their ability to negotiate depends heavily on market size and technical options.

The consequences extend beyond service availability. Cloud providers influence which model architectures are easy to deploy, which tools integrate smoothly, and which monitoring systems become standard. These choices shape local development before a government writes a formal AI policy.

Concentrated compute also directs research attention. Universities with abundant accelerators can train, test, and reproduce systems that resource-constrained institutions cannot examine. Researchers outside the main compute centers may become evaluators or downstream adapters rather than authors of influential model designs.

That imbalance can become self-reinforcing. Strong research institutions attract engineers and investment. Investment finances more infrastructure, while infrastructure supports more experimentation. Regions entering later must compete for talent while paying external providers for essential capacity.

The resulting dependency resembles earlier cloud and platform concentration, but AI raises the stakes. A conventional software service usually executes rules its developers explicitly wrote. A machine-learning model develops patterns from data, making local evaluation critical when languages, institutions, or social conditions differ.

Data-center counts alone do not capture quality, available accelerators, energy constraints, or actual access. A facility designed for storage or ordinary web hosting does not automatically support demanding AI workloads. The Stanford figure should therefore indicate geographic concentration, not a complete inventory of usable AI capacity.

Even with that qualification, the direction is clear. Consumer-facing AI can reach a new market through a website or mobile application. Building domestic technical capacity requires years of investment in power, networks, education, procurement, and research institutions.

This is the central reversal behind the “AI for everyone” promise. Distribution becomes easier because global platforms handle the difficult infrastructure. The same convenience makes dependency harder to see because users experience an instant service instead of the concentrated stack beneath it.

The most realistic response is not complete technological self-sufficiency. Few countries can economically reproduce every layer of the AI supply chain. The practical goal is strategic agency, meaning enough skills, infrastructure, and institutional knowledge to choose providers, audit systems, and build local alternatives where stakes are high.

The AI Skills Gap Determines Who Gains Leverage

Access creates an opportunity, but skills determine whether AI expands a person’s agency or increases exposure to decisions made elsewhere.

The OECD reports substantial differences in AI-related training by educational attainment. In survey data highlighted by IEEE Spectrum, 36 percent of respondents with tertiary education reported participating in AI training during the previous year. The figure was 18 percent among people with upper-secondary education.

The pattern matters because AI rewards existing expertise. A trained programmer can use a coding assistant to explore unfamiliar libraries, generate tests, and review repetitive changes. A person without foundational programming knowledge may accept incorrect output because they cannot identify hidden errors.

The same effect appears outside engineering. A lawyer needs legal knowledge to detect a fabricated citation. A clinician needs medical expertise to challenge an unsafe suggestion. A teacher needs subject knowledge and assessment skills to distinguish assistance from plausible misinformation.

AI literacy programs often emphasize interface use because it is easy to demonstrate. Participants learn how to enter a prompt, revise a request, or summarize a document. Those abilities are useful, but they represent the visible edge of a larger competence gap.

Effective use also requires source evaluation, privacy judgment, statistical reasoning, and awareness of automation bias. Automation bias is the tendency to trust a machine-generated recommendation because it appears systematic. That tendency becomes especially risky when an interface hides uncertainty behind polished language.

The OECD’s AI skills research treats the problem as broader than a shortage of specialist engineers. Workforces need general AI knowledge, complementary skills, and advanced technical expertise. Training systems must therefore serve people at different educational and occupational levels.

Employers also influence how evenly those benefits spread. A company can give every worker access to the same assistant while reserving training and experimentation time for managers or specialized teams. Employees with tighter schedules may be told to adopt AI without receiving time to verify its output.

That implementation pattern can intensify workplace inequality. Highly autonomous workers delegate routine tasks and gain more time for judgment. Closely monitored workers may instead face higher output expectations because management assumes AI has removed every bottleneck.

Public-sector deployments carry another asymmetry. A person applying for benefits may be evaluated through an automated system without understanding which data affected the outcome. The agency may lack enough technical staff to explain the decision or correct a systemic error.

The Dutch childcare benefits scandal remains a warning about automated risk assessment, even though it cannot be reduced to generative AI. Families were wrongly accused of fraud, with serious financial and personal consequences. The case shows how data-driven administration can magnify institutional failures when meaningful review is weak.

Amazon’s discontinued experimental recruiting system provides another historical reference. Reuters reported that the system learned patterns that disadvantaged resumes associated with women. The company did not deploy it as an independent hiring authority, but the episode illustrated how historical data can reproduce existing inequalities.

These precedents matter because AI systems increasingly arrive through ordinary procurement. A school district, hospital, or municipal agency may buy an “intelligent” feature inside software it already uses. Staff can inherit algorithmic decisions without conducting a separate public debate about their introduction.

UNESCO has encouraged governments to integrate AI literacy while addressing ethics, teacher preparation, and inclusion. Its education guidance emphasizes that scientific and educational capacity affects whether societies can participate in AI governance, not merely consume finished systems.

Still, education cannot carry the entire burden. Teaching people to question an automated decision is insufficient if the organization offers no appeal mechanism. Training workers to protect data cannot solve a system that requires them to upload sensitive records.

Skills and institutional design must develop together. Individuals need the ability to recognize risks, while organizations need procedures that make responsible action possible. Otherwise, “AI literacy” becomes a way to transfer responsibility from system owners to users with less power.

This is where the IEEE Spectrum argument becomes most consequential for businesses. Enterprise leaders often measure adoption through activated accounts or generated prompts. Those metrics reveal activity, but they do not show whether employees improved decisions, reduced errors, or gained control over their work.

A stronger measure would examine verified outcomes across roles. It would compare who receives training, who can reject an AI recommendation, and who bears responsibility when a system fails. Without that evidence, rising usage can coexist with widening inequality.

Local Models Cannot Fix Weak Institutions Alone

Local-language data and smaller models can redistribute technical participation, but they cannot compensate for weak oversight or careless policy design.

Indonesia offers an instructive alternative to competing directly in the frontier-model race. Its public research institutions have pursued practical systems for local needs, including multilingual models and tools aimed at underserved communities. One application combines satellite data with machine learning to help artisanal fishers locate schools of fish.

That example matters because it begins with a defined problem rather than a prestige benchmark. A smaller system designed around local geography, occupations, and languages can create value without matching the general capabilities of the largest commercial models.

Local-language development also affects who can participate. Many global systems perform best in languages well represented online and in training datasets. Speakers of less-represented languages may receive weaker answers, less reliable moderation, or interfaces that ignore cultural context.

Building a local model does not guarantee fair representation. Developers still make decisions about dialects, data sources, evaluation groups, and acceptable errors. Communities need influence over those choices, particularly when language data carries cultural or political sensitivity.

Open-source software can lower some barriers by allowing researchers to inspect code, adapt models, and share improvements. Stanford’s AI Index says contributions from countries outside the United States, China, and Europe are becoming more prominent. This broadens participation even when compute remains unequal.

However, an available model is not a complete public capability. Institutions need secure hosting, evaluation data, engineers, procurement standards, and long-term maintenance. A pilot can attract attention while failing to survive staff turnover or a change in government priorities.

South Africa’s 2026 draft AI policy exposed the governance side of that problem. The government withdrew the document shortly after publication when nonexistent academic citations were identified. The communications minister called the incident an unacceptable lapse.

The episode should not become a simplistic argument that governments in developing economies cannot govern AI. Wealthy institutions also publish unsupported claims and deploy poorly tested systems. The more useful lesson concerns the mismatch between policy ambition and verification capacity.

A national strategy can promise sovereign compute, responsible innovation, inclusive data, and workforce development. Delivering those goals requires specialized staff who can check evidence, coordinate agencies, negotiate with vendors, and maintain independent oversight.

The same institutions may already face shortages in cybersecurity, procurement, statistics, and basic digital service delivery. Adding AI responsibility without expanding operational capacity can create impressive policy language with little ability to enforce it.

This risk also affects private organizations. Companies can establish AI principles while leaving individual teams to select models, upload data, and judge accuracy. Governance exists on paper, but the people closest to deployment lack practical review processes.

Local models present a similar tension. They can improve linguistic relevance and reduce dependence on a single provider. Yet a model trained under weak privacy controls can exploit communities even when its developers operate locally.

Ownership is therefore only one dimension of inclusion. The decisive questions include who approved the data, who can inspect performance, and who receives a remedy after harm. A domestically hosted system can still concentrate authority inside an unaccountable institution.

The World Bank’s emphasis on “Small AI” offers a useful path because task-specific systems can require less compute and provide clearer evaluation targets. A crop-disease detector, translation model, or administrative assistant can be tested against a defined local need.

Smaller scope also makes failure easier to observe. Evaluators can compare outputs with expert judgments and measure performance across languages or regions. General-purpose assistants produce a much wider range of answers, complicating claims about overall reliability.

Even so, small systems should not receive small scrutiny. A narrow model used in benefits administration or clinical triage can directly affect essential services. Its limited technical scope may increase accountability, but only if institutions collect evidence and respond to complaints.

The debate should therefore avoid a false choice between global platforms and complete national self-reliance. Countries can combine imported infrastructure, open models, local datasets, regional partnerships, and targeted public investment.

What matters is bargaining power. An institution with technical expertise and credible alternatives can demand clearer contracts, data protections, and performance evidence. One without those capabilities must largely accept the provider’s design and assumptions.

Three Signals Will Show Whether the Divide Is Closing

The next test is not whether AI adoption keeps rising; it is whether infrastructure, training, and accountable local deployment rise with it.

The first signal is the distribution of usable compute. Announcements about sovereign AI facilities should be judged by operational capacity, access rules, and research availability. A data center does little for inclusion if its accelerators remain unavailable to universities, startups, and public-interest teams.

Regional capacity deserves particular attention. Shared facilities and procurement agreements can give smaller economies more leverage than isolated national projects. They can also support common evaluation work across languages that cross national borders.

The signal would strengthen the inequality warning if compute remains concentrated while global usage grows. It would weaken the warning if more regions gain dependable access to advanced hardware and use it for locally directed development.

The second signal is who receives AI training and what that training covers. Governments and employers should publish participation data across education levels, occupations, regions, ages, and income groups. Completion counts alone will not show whether people gained practical judgment.

Useful programs should cover verification, privacy, uncertainty, and contestability alongside interface skills. Contestability means a person can question a decision, understand its basis, and seek meaningful review. That concept links individual literacy to institutional responsibility.

Training would narrow the divide if it reaches workers and communities currently excluded from technical development. It would reinforce inequality if advanced instruction remains concentrated among already advantaged professionals while everyone else receives basic product demonstrations.

The third signal is evidence from locally governed deployments. Indonesia’s public-interest tools, African regional initiatives, and other language-focused projects now need transparent evaluations. Policymakers should look for measured outcomes, documented limitations, community participation, and stable funding.

Success should not be defined by launching an application or publishing a national strategy. A credible deployment should solve a real problem, work across intended populations, and provide a process for correcting harmful errors.

This signal would weaken the article’s central concern if local institutions repeatedly build and sustain systems shaped around their priorities. It would strengthen the concern if pilots depend indefinitely on outside vendors, disappear after grants end, or lack public evaluation.

IEEE Spectrum is right to treat the AI divide as a contest over influence. The stakes go beyond whether a student, worker, or government employee can generate text. They concern whether that person can understand the system, shape its use, and challenge it when necessary.

For developers, this changes what responsible product work requires. Translation is not enough when a system lacks relevant evaluation data. Low latency is not enough when a user cannot appeal an automated result. An open model is not enough when local teams cannot afford to operate or audit it.

Enterprise buyers should ask equally concrete questions. They need to know which employees receive training, which data crosses borders, and which vendor claims have independent evidence. They should also identify who can stop a deployment when performance differs across groups.

Knowledge workers can act at a smaller scale by retaining source material, checking generated claims, and documenting how AI contributed to important decisions. A searchable AI knowledge base can preserve context, but human review remains essential.

Policymakers face the hardest assignment because they must expand access without locking institutions into fragile dependencies. That means funding ordinary capabilities such as education, procurement, statistics, cybersecurity, and administrative review alongside visible AI projects.

The critical question for the next several months is simple: will investment follow the parts of the stack that distribute agency? More users will arrive regardless. The meaningful progress will appear when more communities can build, evaluate, govern, and reject AI systems on their own terms.

That is the standard readers should apply to the next national strategy, corporate rollout, or education initiative. Ask who controls the infrastructure, who receives advanced training, and who can challenge an error. If those answers remain concentrated, wider access will not close the digital divide. It will scale the divide through a new interface.