We may have spent two decades trying to solve the problem of internet access, only to be creating a new challenge: AI access. Only around 3% of US households currently pay for any AI service, according to the Bank of America Institute, even as headlines suggest generative AI adoption is already universal. In fact, the gap between the attention AI receives and its actual use is opening into what looks like the next digital divide. Telecom operators, who built much of the infrastructure that helped narrow the last one, have a specific stake in how this one closes too.
The AI Access Ladder
Call it the AI Access Ladder: connectivity, a capable device, access to a genuinely useful AI assistant, the quality and integration of that access, and the literacy to use it effectively, with each rung dependent on the one below it. Society has learned to subsidise the bottom rung. It has barely begun to ask whether it needs to subsidise access to the tools higher up the ladder. If it eventually decides that it does, operators will almost certainly become part of the delivery infrastructure, in the same way they already are for broadband.
The old divide has not closed
The foundational divide is still open. ITU’s Facts and Figures 2025 report counts 2.2 billion people offline globally, with internet use at 94% in high-income countries against 23% in low-income ones. Any conversation about an AI divide sits on top of a connectivity divide the ITU itself calls stubborn.
AI adoption is diverging even among the connected
Microsoft’s AI Economy Institute found that in the first quarter of 2026, 27.5% of working-age people in developed economies had used a generative AI tool, against 15.4% in developing ones. That gap widened by 1.5 percentage points in six months, even as global adoption reached 17.8% of the working-age population. The same structural factors behind the original digital divide – connectivity, digital skills and, in some regions, electricity – are now widening a second one.
Income decides who climbs the ladder
Income predicts AI use almost as cleanly within wealthy countries as it does between rich and poor ones. A UK-focused agency study found regular ChatGPT use roughly twice as common among higher earners as lower earners. In the US, Epoch AI’s analysis of Ipsos survey data found the income mix of weekly users varies sharply by assistant: Claude skews towards higher earners, Meta AI towards lower earners. Which tools people can reach, not simply whether they use AI at all, is itself splitting along income lines.
Affordability looks like the mechanism behind the split. The Bank of America Institute figure cited above, at a median spend of $20 a month, is already out of reach for a meaningful share of households. Economists Alex Imas and Soumitra Shukla make the gap starker still at a global level: a premium ChatGPT subscription equals roughly 38.6 months of income in some low-income countries, highlighting the affordability and quality-of-access issues described by the AI Access Ladder.
Not every signal points the same way
OpenAI’s own research with Harvard economist David Deming found ChatGPT’s demographic gaps, particularly by gender, have narrowed as its user base has broadened, and usage has grown faster in low and middle-income countries over the past year than in wealthy ones. Growth from a low base and a widening usage-rate gap can both be true at once, and the data currently suggests they are.
Policy is moving on AI skills, but barely on access
Governments have started to respond, mostly through skills programmes rather than access entitlements. India’s Yuva AI for All is a free training course targeting ten million citizens with AI literacy, but it teaches skills rather than granting tool access.
Singapore’s 2026 budget goes further on paper, offering six months of free premium AI access to citizens completing selected courses from the second half of the year. Estonia’s AI Leap programme already gives around 20,000 upper-secondary students and 3,000 teachers free AI tool access alongside training, expanding to 38,000 more students in September 2026.
What still appears to be missing is a policy framework that distinguishes between universal access to useful free AI tools and targeted support for low-income households where paid services offer materially greater capability, privacy or integration.
Where this leaves telecom operators
Operators already run the administrative machinery behind the one AI-adjacent subsidy that widely exists. Eligibility verification, billing integration and distribution for social broadband and mobile tariffs often sit with the network operator, not a separate agency. A future AI-access entitlement would plausibly sit on that same infrastructure. It would not need a parallel system built from scratch.
This matters more than it sounds. Verifying who qualifies for a subsidy, at scale, without excluding the people it is meant to reach, is a genuinely hard operational problem, one operators already solve for broadband. Identity verification, income-tier eligibility checks and fraud prevention around subsidised access are capabilities built over years, not something a ministry or an AI vendor could stand up quickly.
There is also a distribution role. If a government or employer eventually wants to bundle AI-tool access the way Singapore is proposing to bundle it with training, operators are natural distribution partners, already positioned to bundle a subscription onto a mobile plan the way many now bundle streaming services. That puts operators in a position to help shape commercial terms with AI providers, rather than simply reacting to whatever those providers decide to offer consumers directly.
Operators also occupy a more neutral position in this debate than most other participants in the AI value chain. Their commercial interest lies in people having enough connectivity and device capability to use whatever AI tools they choose, not in which specific assistant wins. This gives operators a distinct perspective to bring into a policy conversation currently dominated by the AI vendors, and one they should contribute before the policy framework is set rather than after.
An open question, not a settled one
None of this settles what a fair AI-access policy should look like, only that the gap it would need to close is real, and still widening in some dimensions even as it narrows in others. The next twelve months of data, from Pew, Microsoft, Epoch AI and others, should show whether AI access converges the way mobile broadband eventually did, or settles into a more durable two-tier structure. For now, this is an emerging trend worth tracking rather than a case for any single policy response.
