Can AI Bridge the Digital Divide?
Sean wang
Talkoot Director of AI
“What do you mean, they can’t access the curriculum because they don’t have Google accounts? Aren’t they free?”
That was the first time I realized how deeply inaccessible the modern digital world had become. I was speaking with a group of refugee educators desperately combatting an education crisis in their home country. Over several months, they had developed a nearly complete video curriculum using smuggled textbooks and relying on donations from the diaspora. When they released the curriculum, however, they noticed almost no one was accessing the content – nearly 80% of the users had dropped off at the login screen, and another 40% had dropped off at the onboarding page.
It turned out, their philanthropic partner who hosted the content placed a simple Google login and intake form before users could access the content. This seemingly innocuous step, which takes no more than a few minutes for experienced users of the internet, presented an insurmountable barrier for a population lacking in digital literacy and deeply wary of divulging metadata online.
The current iteration of artificial intelligence represents a much more transformative development than previous iterations of AI breakthroughs. Unlike earlier innovations, confined largely to corporate labs with near infinite resources, large language models (LLMs) are far more accessible. You no longer need a math or computer science degree to take advantage of these powerful technologies. Everyone, from financial analysts to copywriters, can tap into AI to dramatically boost their productivity.
Yet AI has worsened accessibility for a different group of people: the billions already left behind by previous waves of digital transformation, who still worry about creating a Google account and putting their name on a form on the internet. With each iteration of AI, the barrier to entry gets higher, and the digital world gets farther away from the people who would benefit from it the most.
While many AI developers and startups have focused on using AI to boost the productivity of knowledge workers in advanced economies, a very different question emerged for me: what happens to the billions of people already left behind by previous waves of digital transformation? As the technology landscape matures, we must confront the uncomfortable reality that the AI revolution risks deepening the digital divide.
This does not have to be the case. With intentional effort and strategic investment, AI can become one of the most powerful tools in the arsenal for reducing global inequality. The path we choose now will determine whether AI entrenches privilege or extends opportunity.
What is the digital divide?
The digital divide is the large and growing gap in technology adoption between people living in advanced, often western economies, and those living in less developed economies. The statistics are stark and deeply geographical. Approximately 95% of the population in Western Europe has access to affordable, high-speed internet, while only about 25% of sub-Saharan Africa is online at all.
This gap has far-reaching social and economic implications. Workers with lower connectivity lose out on remote work, digital marketplaces, online education, and many other catalysts of economic mobility.
There is also a self-reinforcing quality to the digital divide. Workers who lose out on the opportunities tend to have less incentive to develop the skills that would allow them to benefit from increased connectivity, further reinforcing the disparity between the connected and disconnected worlds.
How AI is Exacerbating the Digital Divide
Currently, AI is increasing the gap in three interconnected ways: labor productivity, physical infrastructure, and cognitive biases.
First, labor productivity in many sectors is booming as a result of AI, but the productivity gains are in concentrated sectors traditionally dominated by those who have benefitted from the information age. Software engineers can choose from a crowded field of “vibe-coding” applications like Claude Code, Codex, and Antigravity; finance workers have powerful tools like Hebbia, 9fin, or Carousel; and scientific researchers have tools like OpenEvidence and Research Rabbit. This development closely follows market logic – after all, fields like technology, finance, and science have the deepest pockets to pay for commercial AI applications. But the yawning productivity growth further compounds economic power in the hands of those who have already benefitted the most from the technological developments of the last half century.
Physical infrastructure plays a role as well. Increasingly, foundation labs like OpenAI, Anthropic, and Google are focused on building bigger, more complex models that require tremendous computing power and specialized equipment. Dominant techniques like retrieval-augmented generation, agents, and LLM-driven evaluations all assume near-infinite access to cheap and fast tokens, further increasing reliance on datacenter-driven AI. This reality restricts AI to advanced economies that can support the data center infrastructure required for the most cutting-edge models, leaving lower-resourced economies behind in the AI transformation.
Lastly, the cognitive biases of AI favor those who are already digitally connected. Almost all large language models are trained on a dataset derived from publicly available text from internet archives and social media, which makes it much better at understanding the linguistic and cognitive constructs favored by the already-digitally-literate. I encountered some particularly thorny examples of this when I was advising a group of Senegalese engineering students who were building a legal aid chatbot to help small businesses navigate the complex and overlapping regulations in West Africa. Most of the foundation AI models we used at the time struggled with understanding local legal jargon and regional French variations, due to underrepresentation of West African French in the LLM’s training data.
How can AI Shrink the Digital Divide?
Despite the current trajectory, AI still has the potential to dramatically shrink the digital divide, as long as a moderate amount of resources are put into developing models and applications that are specialized for this purpose. There are three key areas of investment: accessible user interface, small democratized models, and targeted, culturally responsive applications that solve problems for people left behind by the digital revolution.
One of the most promising applications of AI for digital equality is the ability to build new, accessible user interfaces. The built digital world is full of invisible barriers that reinforce access controls. In my work with less-connected populations, I found that even something as simple as a Google OAuth login can present a significant barrier to access, since getting a Google account requires navigating an unfamiliar workflow with novel concepts. UI elements like the scroll bar, so ubiquitous and easily taken for granted, often present accessibility problems for children whose only interaction with a screen is through a touch-screen interface. AI has the ability to bypass all of that by providing natural-language and voice interfaces that people understand intuitively. By connecting more services to AI-powered interfaces, we have an opportunity to dramatically lower the barrier to entry for many digital services and flattening the on-ramp for many populations to participate in the digital revolution.
For these interfaces to be economical, they must be powered by smaller, less expensive AI models. Cutting-edge foundation models like ChatGPT, Claude, or Gemini, even at subsidized rates, are cost-prohibitive for users in less developed economies. Additionally, many of these models have interfaces that assume constant internet connectivity, which may not always be a given. Certain emerging techniques like miniaturization and quantization can address this problem by creating AI systems that can run on locally available hardware, eliminating the need to pay for additional AI services or more stable internet connectivity. Using these models, less-connected populations can participate in the benefits of the AI revolution without needing to spend massively on AI infrastructure.
Lastly, developers should build AI applications that respond more directly to the needs of less-connected populations. Currently, AI applications are developed primarily to meet the needs of highly connected, often western, knowledge workers, with feedback collected through means that are locked under the same accessibility barriers. AI application developers could use methodologies that intentionally collect more user information from less connected users and create AI applications that more deftly navigate communication and cultural gaps.
Inclusive AI
Increasingly, we live in a world where educational, professional, and economic attainment is decided not based on one’s ability, but on one’s place of birth. Technological advancements increasingly benefit a narrower and narrower set of the already-successful. AI presents a crossroads: It can either further entrench structures of economic and technological inequality, or it can potentially lift up billions who have been left behind by the information revolution of the last half century.
I believe the impact of the AI revolution, indeed any economic transformation, will ultimately be measured by how many people it lifted out of poverty, rather than how much profit was generated for shareholders. As AI developers and users of AI applications, we have the power to shape the next stages of the AI revolution, and contribute to making it an inclusive, rather than a dystopic, one.