In January 2000, at the dawn of the new millennium, then-President Kim Dae-jung said in his New Year's address that it was "an era in which competitiveness based on knowledge and information matters more than natural resources." His vision of making South Korea one of the world's 10 leading knowledge and information powers extended even to the phone bills of poor students. The government pushed a policy to supply personal computers to 50,000 students from low-income households and to subsidize their internet fees for five years. Owning a computer meant little if the cost of connectivity kept them offline, leaving them cut off from information all the same. A program to provide computer training to 10 million people followed the same logic. Raising national competitiveness through knowledge required improving the public's access to it. The spread of the internet eased information asymmetry by diffusing data that had been concentrated in the hands of a few.
Today, however, we face another crisis of knowledge inequality. Tom Wheeler, former chairman of the U.S. Federal Communications Commission (FCC), argued in a Brookings Institution essay last month that the "token divide" differs from the information divide of the past. With flat-rate broadband, there was no need to worry about extra charges each time one looked up more material. Artificial intelligence, by contrast, incurs computing costs with every additional query and round of reasoning.
Even for those who are connected, the chance to reach knowledge now depends on which AI they use and how much of it. AI helps with reading, comparing and interpreting material. A small business owner who cannot afford specialized staff can analyze a market, and a student studying alone can keep asking questions until something clicks. But using it as much as needed requires being able to afford the fees. In other words, how many tokens — AI's unit of information processing — one can use may determine the results that students, small business owners and companies are able to achieve.
The more tokens available, the greater the capacity to discard a weak answer and ask again, to gather additional material and to examine counterarguments. When someone with little token purchasing power hits a usage cap and stops searching, those with more can keep going. The chance to try again until a better answer emerges hinges on purchasing power.
The token divide also bears on fairness in education and selection. On an assignment where AI use is permitted, the same deadline means one student rations a usage cap while polishing a draft, and another has several models generate alternatives, then compares and revises them. In competitions that serve as a springboard to employment, if AI use is allowed, the ability to marshal tokens shapes the opportunity to refine proposals, videos and code. Token purchasing power, in short, can affect outcomes in admissions and hiring.
Early-stage technology startups run into the same problem. In developing prototypes and verifying code with AI, more tokens bought mean more chances to run additional experiments and make fixes. It is difficult to judge the gap between the output of a team that halted further verification for lack of funds and that of a team that tested multiple alternatives as a gap in technical capability alone. When evaluations lead to investment and commercialization support, the ability to marshal tokens becomes a decisive factor in whether a company reaches the next stage.
If the internet once lowered the threshold for reaching information, AI broadens the ability to analyze information and put it to use as knowledge. Even with the same public data and research reports in hand, the knowledge one extracts can differ depending on the capacity to compare and verify with AI. It has become hard to speak of equal access to knowledge on the strength of disclosure alone. The more sophisticated the analytical and reasoning power that money can buy, the greater its influence on the final product. And if people and companies are judged on output alone, the risk grows that differences in financial circumstances will be mistaken for differences in ability.
The success of an AI powerhouse cannot be measured by model performance and data center capacity alone. Students must be able to use AI enough to keep learning, and founders enough to validate their ideas. Policies that lower the cost of using AI are not merely welfare but an investment in nurturing talent and technology. If the state pours vast resources into building national AI capability while leaving each citizen's access to his or her own means, the new knowledge divide of the AI era will be hard to narrow. If economic means still determine who gets to learn and who gets to take risks, our society will have advanced only in technology while regressing in equality and fairness.