The Next Digital Divide: Who Can Turn AI Into Useful Work
As access to AI expands, a second gap is emerging between having powerful tools and possessing the judgment, local knowledge and systems needed to turn them into reliable work.
The first digital divide was about who could get online. The AI debate has mostly repeated that story: who has electricity, broadband, computing power, and a model in a language they speak.
That gap has not closed. United Nations Secretary-General António Guterres has warned that about one-third of humanity remains offline, while computing power, data, technical expertise and investment stay concentrated in a handful of countries and companies.
Intensity of use is uneven even among people who are online. The World Bank reports that, as of May 2025, ChatGPT use per internet user was nearly 50 times higher in high-income countries than in low-income countries.
A second gap is now visible among those who already have the tools. Access to a model is getting cheaper. The ability to turn that model into work that holds up outside the chat window is not.
What the development record already shows
The World Bank’s World Development Report 2026 argues that developing countries do not need to build trillion-dollar general-purpose models to benefit. Importing tools is also not enough. Systems, the Bank says, need to be adapted to local languages, institutions, data and development needs, while countries build the skills and infrastructure to do more than run a demo.
The Gates Foundation’s 2026 Goalkeepers report supplies the language half of that problem. More than 90 percent of the data used to train early large language models came from English-language sources. In English, leading speech-recognition systems err less than 6 percent of the time; in Yoruba, the Foundation reports, the error rate exceeds 60 percent.
Johannes Jütting, executive head of the PARIS21 Secretariat at the OECD, has told IPS that many low-income countries still lack the conditions that make any of this usable. “First, connectivity: without it, there is no practical AI application. Second, technical infrastructure such as data centers and reliable data transmission. Third, human capacity and skills, which require sustained investment. And fourth, governance and legal frameworks that must be updated to reflect new technologies.”
“How can we talk about fancy AI models when basic population data is missing?” he asked. “We have to start with the fundamentals.”
Development economist Johanna Choumert-Nkolo, also speaking to IPS, put the same constraint in operational terms. “Digital tools offer huge opportunities. But they must be rooted in context, evidence and local needs.” Mobile money worked, she has argued, because it solved a local problem and fit local realities. “The key is context. We must understand what problem we are trying to solve and whether digital tools are the right fit.”
That is the global problem as the institutions covering it now describe it: access, language, local data, skills, institutions. What is less visible in those reports is the mechanism that appears once a tool is already on someone’s desk.
One operator, one mechanism
Ken Ashe is a New Jersey-based CPA and project management professional who spent roughly two decades working across accounting, finance, project and product management at S&P Global, EY, Noom and Prudential Financial. In the later part of that career, he built Lucky Domains, a business focused on domains, websites and SEO, before leaving corporate work to build applications with commercially available AI models, APIs and automation tools documented at KenAshe.ai.
He does not train foundation models. After a year of combining tools that vendors now describe as democratized, the pattern he documents is not a shortage of chat windows. It is a shortage of judgment, workflow design, verification and domain knowledge — and of people who can tell output from a finished task.
“Most users do not want an agent,” Ashe has written. “They want a repeatable way to run a standard business process with fewer clicks.”
If that is the real demand, handing someone a general model is not the same as handing them productivity. It is closer to handing them a fast junior assistant who does not know the local rules, does not know what a mistake costs, and will not always say when a step was skipped.
His own systems have failed in ways familiar to anyone who has put a model inside a live workflow. A pipeline skips a step and does not complain. A draft cites a source that was never opened. Fluent text arrives built from events that did not happen.
People who already possess craft — accountants, nurses, agronomists, logistics managers, teachers who know their syllabus — can use a general model as an assistant and still recognize a bad answer. People who have access to the model but lack the relevant domain knowledge may generate output they are poorly equipped to audit.
That raises a harder question than another round of digital-skills pilots usually asks: how much can prompt training accomplish if users lack the domain knowledge needed to evaluate the answer?
The barrier moved
None of this makes the first divide irrelevant. Power, connectivity, compute, language and governance remain binding for billions of people. Jütting’s four conditions still come first. No workflow lecture substitutes for electricity.
Among offices that do have connectivity, a different failure is taking shape. “AI adoption” can be logged when a chatbot appears on a government site or a classroom runs a pilot. World Bank evidence suggests that much current use in developing-economy firms remains concentrated in functions such as search, writing and translation rather than deeper reorganization of work. Fluent errors are harder to see than a blank page.
AI does lower some barriers. A one-person firm can now attempt a first website, a first research pass, a first draft of correspondence. But lower financial barriers shift more responsibility onto the operator: defining the task, checking the result and deciding what the system is allowed to do.
In published assessments of systems that can spend money or change records, Ashe has argued that the model is rarely the first thing that breaks. Permissions break. State breaks. Retries break. A system can move funds and then be unable to explain why. If that is visible inside a small private practice in a high-income country, it is a warning for public systems being told to adopt AI at speed.
Choumert-Nkolo’s caution sits next to that warning rather than against it. Tools fail when they are not matched to a defined problem. Ashe’s year of building is one illustration of that mismatch at the level of a single workflow. Jütting’s point is the same mismatch at the level of a national statistical office. The scale is different. The missing ingredients are not.
What literacy would have to mean
The Gates Foundation argues that doctors, farmers, teachers and developers need the chance to evaluate tools and adapt what works to their own contexts. The World Bank’s complementary finding is that adaptation — to language, institutions, data and actual development needs — is the work, not an afterthought.
If AI literacy is going to mean much in that setting, it has to extend well beyond prompt writing. It is the capacity to define a bounded task, supply the local facts a general model does not have, keep a human on irreversible actions, and measure whether time and error rates moved.
Widely available models do not necessarily erase the advantages of larger organizations. They may give smaller organizations new leverage on tasks they already understand well enough to bound, check and measure. But where that knowledge is thin, organizations with reviewers, controls and specialized expertise may be better positioned to absorb the same technology.
Jütting has told IPS that digital transformation “is not just a technology issue. It is a change management issue, a capacity development issue, a skills issue, and a political will issue.” That sentence travels. It describes a ministry without population data. It also describes an office that has been handed a model and still cannot say what “done” means.
AI is getting cheaper. Knowing what to do with it is not. The first fact is in every access report. The second is visible wherever someone already has the tool and still has a job that has to be right.