The Divide We Learned to Name

I. The question of the week

The phrase “digital divide” is old enough to have grandchildren. It was coined for telephone lines, inherited by broadband, and has now been handed down, barely altered, to artificial intelligence. What changes across those generations is not the structure of the worry — some people get the tool, some people get left holding the old one — but the confidence with which each era believes it will finally close the gap this time. AI’s turn has produced a discourse fluent in equity: fluent in naming who is missing, fluent in modeling how the gap will widen, fluent above all in the vocabulary of divergence, exclusion, and inclusion. The question worth asking after three years of that fluency is whether the naming has produced anything a student, a teacher, or a ministry of education can actually use.

This is a topic where the conversation and the conditions have moved at different speeds. The conversation learned quickly. By early 2025 it could describe the equity problem in AI adoption with real precision — gender gaps in who uses the tools, language gaps in what the tools understand, regional gaps in who trains them, disability gaps in who is designed for at all. The conditions moved more slowly, and mostly in the wrong direction: adoption raced ahead in exactly the populations already advantaged, capital concentrated, and the classroom-level questions — the ones about the student learning in a language the model handles poorly, or the student whose disability the interface never anticipated — arrived last and funded least.

The arc that follows traces both clocks. First, what we have been saying about AI equity, and how the saying shifted from opportunity to alarm across three quarters of 2025. Then, what has actually been happening to the numbers underneath. Then the more uncomfortable synthesis: where the rhetoric and the reality genuinely meet, and where the discourse has substituted the act of naming a divide for the far harder work of building the bridge across it.

II. What we’ve been saying

The opening register of 2025 was opportunity. In February, the Wharton visiting scholar Dr. Cornelia Walther published “How AI Exclusion Impacts Humankind”, framing the equity question not as a risk to be managed but as a resource being wasted — the human potential locked out of the system. A few weeks earlier the World Economic Forum had issued its “Blueprint for Intelligent Economies”, which treated AI maturity as a journey with unequal starting points but a shared destination, reachable through regional collaboration. The grammar of these pieces is the grammar of the seizable opportunity: the gap is real, the gap is closable, and closing it is a matter of will and coordination. Equity, in this telling, is upside left on the table.

Almost immediately a second register answered it. In March, researchers at the Free University of Brussels published work — carried under the headline “Unmasking inequalities in AI: Research challenges idea that bias is a technical flaw” — that refused the premise the opportunity discourse depended on. Bias, they argued, is not a bug in an otherwise neutral system awaiting a patch; it is the sediment of historical data and existing power, learned faithfully by machines doing exactly what they were built to do. This is the same insight the journalist and researcher Meredith Broussard has pressed for years; her Artificial Unintelligence (2018) traces the founding of NYU’s AI Now Institute in 2017 precisely to study AI as a matter of social power rather than engineering hygiene. The distinction matters for the equity conversation because it determines what kind of solution the problem admits: an opportunity is optimized, a power relation is contested.

Through the second quarter of 2025 the vocabulary hardened around a single word — divide — and the mood tilted, though not evenly. In April, reporting on a United Nations projection warned that “AI could affect 40% of jobs and widen inequality between nations”. In May, coverage of workplace adoption found that “Generative AI like ChatGPT is at risk of creating new gender gap at work”, the gap opening not in access to the tool but in willingness to use it. That same month the Fraunhofer survey “Who Uses AI in Research, and for What?” put empirical texture on the same anxiety inside the academy. What is striking about this quarter is that the pieces were still, by the discourse’s own self-labeling, “optimistic” — they framed the divide as a problem AI-driven growth could yet solve, as when Google’s UK policy blog argued that “AI’s potential to tackle the UK’s productivity puzzle” hinged only on getting workers to adopt. The alarm and the optimism were, for a season, the same sentence.

By the third quarter the sentence broke. The framing inverted — critical pieces began to outnumber optimistic ones — and, more importantly, the conversation finally descended from the altitude of nations and economies to the level of the classroom. The Auburn study “Evaluating the Economic Impact, Equity Implications, and Long-Term Prospects of AI-Powered Personalized Learning” is representative of the shift: it asked not whether AI would widen inequality between countries but whether the personalized-learning systems now entering schools would tailor themselves to every student or only to the ones the training data already understood. This is the register in which the sharper equity questions live — language equity, the prioritization of students with disabilities, the regional access gap — and it is the register our own coverage had been circling since spring. Our social-aspects briefing of 2025-06-29 noted the recurring claim that AI could “democratize access to resources” in education; by the 2025-09-28 briefing that democratization rhetoric was being read against its own results, a spectrum “ranging from enhancing educational equity to addressing systemic biases” that no longer took the enhancement for granted.

What the discourse acquired across these three quarters, then, was not a new fact but a new mood and a finer resolution. It learned to distrust the word “democratize.” It learned that a divide could be named at the level of the individual learner. What it did not obviously acquire was any corresponding account of who would pay to close what it had become so skilled at describing.

III. What’s been happening

Underneath the rhetoric, the most legible movement has been in regulation, and the legibility is deceptive. Stanford’s AI Index Report 2024 records that the number of AI-related regulations in the United States rose from a single measure in 2016 to twenty-five in 2023 — a 56.3 percent jump in the final year alone — while the European Union reached agreement on the terms of its landmark AI Act. Read one way this is the state catching up to the equity problem; read another it is the state catching up to itself only where the state was already strong. The same report notes that the regulatory action concentrated on “both sides of the Atlantic,” which is precisely where the resources to regulate already sit. Policy proliferated fastest in the jurisdictions least likely to be on the losing end of the divide.

The adoption numbers tell a more paradoxical story than the “access gap” framing predicts. The United Nations Development Programme reports in “The Next Great Divergence” that AI reached 1.2 billion users in three years, with nearly 70 percent of them in developing countries — a distribution that ought to embarrass any simple narrative of the poor locked out. But raw access is not the operative variable. The joint UN and International Labour Organization report “Mind the AI Divide” argues that without cooperative international action the revolution “will only widen the gap between high and low-income countries,” because the value does not accrue where the users are; it accrues where the infrastructure, the models, and the capital are. On that measure the concentration is stark: an analysis of “AI, Wealth Inequality, and the Future of Retail Investing” found that 85.87 percent of the $29.29 billion in tech venture capital flowed to AI infrastructure. A billion new users in the Global South, and the money for what they use pooling almost entirely elsewhere.

The demographic gaps that predate AI have been reproduced inside it with unsettling fidelity. The AI Index Report 2024 finds that “substantial gender gaps persist” among European informatics and computer-science graduates — every surveyed European country reported more male than female graduates at bachelor’s, master’s, and doctoral levels — even as the proportion of Hispanic computer-science bachelor’s graduates in the United States grew by a modest 5.2 percentage points. These are the pipelines that produce the people who build the systems, and they remain skewed at the source. The point that the Free University of Brussels researchers made rhetorically, the graduation data makes structurally: the exclusion is upstream of the algorithm. As John Maeda writes in How to speak machine (2019), “the tech industry has an ongoing tendency toward exclusion” — a tendency that no downstream fairness patch reaches, because it operates before the code is ever written.

At the level where the discourse arrived last — the classroom, the region, the individual student — the reality is thinnest and least measured. Experts quoted in “Experts warn AI could widen digital divide in the Philippines” described a country that could be transformed by AI or left further behind by it, depending on reforms not yet made. The scholarship coming out of the Global South — a qualitative framework on the “Determinants of Ethical and Scalable AI for the Sustainable Development Goals” — reads less like a report on what has happened than a set of conditions for what might. And the education-specific equity questions the discourse learned to ask by Q3 — whether a personalized-learning engine serves a student who speaks a low-resource language, whether it accommodates a disability — remain largely at the stage of study design rather than deployed evidence. The “Artificial Intelligence in the Global Economy” study frames the whole matter as an open question in its title — “Propelling Development or Deepening Inequality?” — which is an honest admission that the outcome is not yet in the data.

The most recent entry closes the loop the Brussels researchers opened. An October analysis of “Biases in Artificial Intelligence and Implications for AI Use in Public Administration” returned to the mathematics itself, showing how bias enters through the technical foundations of large language models — not to excuse it as a mere glitch, but to demonstrate that the technical and the social are the same substance viewed from two angles.

III½. — one caution the numbers carry

What the reality does not show, across any quarter, is the closing of a single named gap. It shows regulation multiplying, adoption spreading, capital concentrating, and pipelines staying skewed. Movement everywhere; convergence nowhere.

IV. Where they meet, where they miss

The rhetoric and the reality meet on the diagnosis and part on the cure. Both now agree — the optimistic register having finally caught up to the critical one — that AI is at least as likely to widen educational and economic inequality as to narrow it, and that the widening runs along the familiar seams of gender, language, region, and disability. That is real convergence, and it should not be undersold: a discourse that in 2023 was still selling frictionless democratization had, by late 2025, learned to say “divergence” without flinching. The naming is genuine progress.

The miss is in what the naming licenses. The dominant framing still treats equity as a downstream adjustment — a matter of extending access, tuning a model, adding a language, coordinating a region — when the evidence points upstream, to who is in the room before the system exists. The skewed graduation pipelines in the AI Index Report 2024 and the industry’s structural “tendency toward exclusion” that Maeda names in How to speak machine describe a problem that no amount of downstream access will resolve, because the tool arrives already shaped by the absence. To promise inclusion through distribution is to promise to share a room that was designed without you.

Policy exhibits the same slippage in a different key. The MIT Press Essential Knowledge volume AI Ethics observes that “in spite of cultural differences, it turns out that AI ethics policies are remarkably similar” — a convergence that sounds like consensus but may be closer to copying, the same principles circulated between the same well-resourced jurisdictions. The same volume warns that beneath the shared language lie real “disagreements and tensions,” including “on how much new legislation is needed.” The twenty-five regulations the AI Index counts are almost all Atlantic; the students the discourse worries about by Q3 are disproportionately not. A policy vocabulary can be globally uniform and its enforcement, its funding, and its benefit remain sharply local. Uniform rhetoric is not distributed capacity.

Here the skeptical reading earns its place. It is convenient for the institutions producing the most AI — the venture-funded infrastructure that absorbed 85.87 percent of the sector’s capital, the two regulatory blocs writing the rules, the universities and firms whose surveys define who “adopts” — to hold the equity conversation at the level of description. Describing a gap costs a report; closing one costs a budget line, a redesigned pipeline, a model trained on a language with no commercial market. Our social-aspects briefings tracked, across March through September of 2025, how reliably the promise to “democratize access” recurred in the sources; what recurred far less was any account of who bears the cost of the democratizing. The UNESCO module Think Critically Click Wisely reminds its readers that the deeper concern with these systems “raises questions about control” — and control is exactly what the equity discourse tends to leave unspecified. Whose gap, closed on whose schedule, with whose money.

The classroom is where the miss becomes concrete. The Auburn and Global South studies ask the right questions — does the personalized tutor serve the student in the margins? — but they ask them in the future tense, as research agendas, while the systems are already being installed in the present tense. The discourse’s finest resolution arrived at the exact moment its subjects had the least leverage over what was being built for them.

V. The longer view

Three years in, the AI equity conversation has become genuinely good at one thing and conspicuously untested at another. It can find the gap — by gender, by language, by disability, by region, by the mathematics of the model itself — with a precision the broadband era never managed. What it has not yet done, in any figure across the arc, is close one. The AI Index Report 2024 can show regulation rising twenty-five-fold and the graduate pipelines staying skewed in the same breath, which is the whole story in miniature: motion in the vocabulary, stasis in the structure. A society can become fluent in describing its exclusions long before it becomes willing to pay for their remedy, and fluency, left alone, has a way of feeling like progress while nothing underneath it moves.

That is the caution to carry into the next quarter, when the personalized-learning systems now in trials become the ones in schools. The test of this discourse will not be whether it can name the divide more finely — it has proven it can — but whether the naming ever converts into a budget, a redesigned pipeline, a model that speaks a language with no market. Until it does, we should be honest about what we have built: not a bridge, but an exquisitely detailed map of the water.

We have learned to name the divide with real precision; naming it is not the same as being willing to pay to cross it.