The 10.6-point gap

A technology gap that widens rather than narrows tells you something about the technology's distribution. A literacy gap that widens between the countries holding the training data and the countries whose data trains the models tells you something about consent.

The January 2026 global adoption figures make the second case unavoidable. AI literacy is no longer an education variable. It has become a proxy for democratic consent, and the state of the literacy is worse than the state of the technology.

The evidence

Global AI adoption reached 66 percent in 2025, up 11 percentage points year on year. Over the same period the Global North–South adoption gap widened to 10.6 percentage points, from 8 points in 2024 ^[1]. Adoption rose everywhere and the distance between the two halves of the world grew.

The underlying literacy picture is not reassuring. The OECD records a participation gap of roughly 2.5 times between high-literacy and low-literacy adults in continuing education ^[2]. Where general literacy participation is already skewed, AI literacy participation will be more skewed, because it sits at the far end of the same funnel.

The legal obligation is now live. The EU AI Act made AI literacy a binding obligation on providers and deployers from 2 February 2025 ^[3]. In February 2026 the United States Department of Labor published an AI Literacy Framework addressing the same question from the workforce side ^[4]. Two of the three largest regulatory blocs have decided that literacy is a duty rather than an aspiration.

And the training-data direction runs against the gap. Most large language models are trained predominantly on English-language text. Most of the world does not read primarily in English. Public administration in the languages of the next billion users remains, in the phrase used by researchers mapping the problem, a language gap that has not been minded ^[5].

Why the gap is a governance problem, not a market problem

The market reading of an adoption gap is that it is a lag. Prices fall, tools improve, distribution reaches further, the gap closes. That reading works for consumer technology adopted by choice.

It does not work for a technology that is applied to people rather than used by them.

In a democracy, a technology that citizens do not understand cannot legitimately govern them. If the 10.6-point gap widens by a further 2 points a year, then by 2030 the same class of models will be shaping public-benefits decisions, welfare eligibility, and immigration adjudication for populations who have had structurally less exposure to those models than the populations that designed them. Consent, in that arrangement, is a formality performed by people who were never given the means to evaluate what they were consenting to.

The Dutch childcare-benefits scandal is the pattern at small scale. Roughly 20,000 parents were wrongly labelled fraudsters by a risk-scoring apparatus, in one of the highest-literacy societies on earth, with a functioning parliament, a free press, and an active ombudsman. It still took years to surface and it brought down a government.

Extend that pattern to administrations with lower baseline literacy, at higher deployment intensity, with weaker channels for contesting a determination. The outcome is not hard to predict. It is only hard to prevent.

What literacy has to mean to be worth measuring

Two definitions are circulating and only one is useful.

The weak definition is familiarity: has the person used a chatbot, do they know what a model is, can they name three risks. This is what most surveys measure and it correlates with almost nothing that matters.

The strong definition is capability at the point of decision: can a caseworker tell when a system's output should not be relied on, can a procurement officer read a model card and identify what is missing, can a citizen recognise that a decision was AI-supported and know what to ask for next. That is a much smaller number in every country and the gap on that measure is almost certainly wider than 10.6 points.

The regulatory obligations now in force point at the strong definition. Most implementation efforts are delivering the weak one, because the weak one is purchasable as a training module and the strong one has to be built into the work.

The bear case

If the 2027 global adoption data shows the gap closing by more than 3 percentage points, this argument weakens materially, and the correct reading would be that distribution economics did the work that policy was arguing about.

Watch the number. It is a single published figure, released annually, and it will settle the question without anyone needing to win a debate.

What ImpactLab is doing

KI for Norge is the Norwegian-language answer: literacy infrastructure in the language of the administration, built for the decisions public servants actually make.

The Norway–Pakistan publication track is the small-state and large-state analogue, testing whether literacy compounds on comparable timescales at radically different scales.

The Middle East ATLAS is the Arabic-language and Urdu-language vector for the same work.

All three sit under Pillar I, Knowledge and Awareness, and run in Mode IV, Blueprints. None of them is a training course.