The cruelest thing about AI is not that it doesn't give answers to the poor.

On the contrary, it provides answers for everyone.

It provides students with thesis frameworks, employees with email templates, entrepreneurs with business plans, and ordinary people with legal explanations, investment advice, and career planning. For the first time, answers are this cheap, this abundant, and this convincing.

But that’s precisely the issue: when answers are available to everyone, what becomes truly scarce is no longer the answers themselves, but the ability to judge them.

The new information poor are not those excluded from AI, but those who already have the answers but lack the ability to evaluate them or the means to turn those answers into real opportunities.

In the internet age, information poor people are those excluded from the network. The solution seems clear: connect to the internet, distribute devices, and improve literacy rates. In the search engine era, it’s slightly more complex—you need to learn how to extract keywords, evaluate sources, and assess credibility, and ideally, understand some English. But the barriers are visible and measurable.

The information gap in the AI era is structurally entirely different.

Large language models are not search engines—they generate conclusions for you directly. You no longer need to "search" for answers; instead, answers are delivered to you in fluent paragraphs, clear steps, and confident language. On the surface, the barrier has dropped dramatically. But beneath lies a harsh reality: when answers become cheap, so do errors—and the ability to discern whether an answer is trustworthy has become rarer and more valuable than ever.

Throughout history, the diffusion of every general-purpose technology has followed the same logic: new technologies first reward those who already possess complementary capital. The printing press benefited literate individuals first; computers favored those who understood office software and programming; the internet advantaged those with strong English skills and proficient search abilities. Complementary capital for AI includes education, specialized knowledge, critical thinking, organizational empowerment, ability to pay, and perhaps the hardest to quantify of all—judgment.

New technologies rarely reward those who need them most; they typically reward those who can best leverage them.

The first crack of inequality was already drawn before you opened the app.

In April 2026, the AI research firm Epoch AI, in collaboration with the polling company Ipsos, conducted a survey of approximately 5,000 U.S. adults. Over three rounds of questioning, respondents were asked a seemingly straightforward question: Which AI services did you use in the past week? But the responses revealed not simple product preferences, but a complex map intertwining income, access points, and distribution.

Of Claude’s weekly active users, approximately 80% come from households with an annual income of over $100,000; among Meta AI users, this proportion is only 37%. Conversely, about 32% of Meta AI users come from households with an annual income of under $50,000, while only 7% of Claude users fall into this category.

These numbers matter not because they prove that "the rich use advanced AI, while the poor use free AI." That’s the most superficial interpretation. What’s more worth asking is: why do different people encounter different AIs in their daily lives?

One person asks AI to pair leftovers in the fridge with a dinner, brighten the background in a photo, and rephrase a text message to be more polite. Another person asks AI to organize customer interviews, compare vendor quotes, and identify weak assumptions in a report. Both are using the same technology—but one stops at convenience, while the other enters the cycle of income, job security, and negotiating power.

The difference lies not only with users but also with the entry points. Claude’s usage path requires actively searching for, comparing products, understanding capability differences, choosing to pay, and then integrating the tool into your workflow—each step filters users. In contrast, Meta AI’s path is nearly the opposite: it’s built directly into social platforms, free and low-friction, so users often encounter it passively while scrolling feeds, sending messages, or viewing photos.

This is not a market of taste, but a market of distribution. Users appear to be choosing tools, while the tools’ prices and access points are also choosing users.

Source: epoch.ai

Even if you find a good AI tool, a second bottleneck awaits you within the company.

In a typical office, the arrival of AI rarely comes in the form of "layoff notices." Instead, it first takes over meeting minutes, draft emails, spreadsheet organization, customer categorization, and initial reports. For managers, this automation frees up time to focus on judgment-based tasks; for new hires and frontline employees, however, it removes the very opportunities they need to prove themselves, practice decision-making, and advance to higher-level roles.

The data is colder than this scenario: A joint survey by the Financial Times and a research institute on AI adoption among the UK and US workforce (conducted February–March 2026, surveying over 4,000 respondents in both countries) found that 63% of workers in the highest salary bracket use AI on a typical workday, compared to just 17% and 16% in the two lowest brackets. This is not a gentle slope—it’s a cliff.

The more critical finding lies in the drivers. The regression analysis of this workplace survey revealed that, after controlling for other variables, salary has nearly no impact on AI adoption rates—what truly matters are four factors: age, experience, industry, and training. Among these, training has the strongest effect: employees at companies offering formal AI training use AI on average 37 percentage points more daily than those at comparable companies without such training. Even informal guidance results in a 24 percentage point increase.

However, the reality is that as of early 2026, only 14% of employees reported having received formal AI training from their employers, while two-thirds had received no training of any kind.

AI training is not a technical issue—it's an allocation issue. Those selected for training are granted access to the path of productivity growth; those who aren’t, the tools remain merely unopened icons on a screen.

AI is an application on the consumer side and a privilege on the workplace side—and privileges have never been evenly distributed.

Source: Focaldata

This is the most subtle分流, and also the most fundamental one.

Imagine a recent graduate who has just joined a consulting firm. He uses AI to generate a first draft of an industry analysis report—well-structured, data-rich, and confidently worded. His supervisor, someone with a decade of experience in the industry, glances at it and points out that two of the data citations stem from sources with methodological flaws, and that the causal reasoning behind the third conclusion is flawed. The supervisor isn’t more hardworking—he simply has that foundational layer: he knows where mistakes are likely to occur, and he can distinguish genuine fluency from the hollow smoothness of a machine filling in blanks.

This is the true meaning behind the counterintuitive finding in workplace survey data: the heaviest users of AI at work are not the youngest employees, but those who have been in their current roles for 2 to 10 years. The relationship between AI usage and experience remains significant even after controlling for age. This is not because younger employees don’t want to use AI, but because AI’s value heavily depends on the user’s existing judgment and expertise.

Experience is AI’s most important complementary capital, and experience cannot be subscribed to.

AI has lowered the cost of "sounding knowledgeable," but not the cost of "truly understanding." An even more dangerous consequence is that users with weaker foundational knowledge are more likely to accept AI outputs uncritically—and the more they do so, the harder it becomes for their judgment to develop. When an agent makes decisions for you, you’re consuming intelligence, not accumulating it.

Nobel laureate and MIT professor Daron Acemoglu is blunt about it: using AI tools requires a certain level of education, abstract thinking, quantitative skills, and familiarity with technology. "It is almost certain that AI will increase inequality," he says.

The new reality is here: the poor are not those without AI, but those who have AI, access, and answers—yet lack the training to judge those answers; they have tools and scenarios, but not the authority to turn tool outputs into opportunities; they consume intelligence daily, yet never accumulate it.

However, the relationship between AI and inequality is not only about widening the gap.

Multiple experimental studies have found that, under controlled conditions, AI tends to provide greater improvements for low-skilled individuals—such as call center workers, junior writers, and entry-level consultants. This is not surprising: top experts gain only marginal benefits from AI; for someone who has never been able to afford professional services, using AI to understand a contract for the first time represents a qualitative leap.

But there is a key distinction to note: experimental studies measure "improvement after use," while real-world data measures "who is actually using it," "who is permitted to use it," and "who can turn the results into opportunities." Neither set of data is lying—they are measuring entirely different things.

A technology can narrow the gap in the lab while widening it in the real world—if its adoption is unequal, if the contexts are unequal, and if judgment itself is unequal.

AI possesses technical characteristics of equality, yet operates within unequal social structures. Both of these truths coexisting define the true nature of the problem.

Each generation tends to believe that the general-purpose technologies of their time will disrupt the old order.

After the invention of printing, literate individuals benefited for centuries. At the early stage of computer普及, it amplified the abilities of those already skilled in office software and coding. The early benefits of the internet flowed to those who understood English, knew how to search, and had the time and motivation to arbitrage. In every technological wave, the voice of "this time is different" has been loud, yet structural disparities often take decades to become visibly apparent.

AI's divergence may occur faster and run deeper, because it affects not just one category of tasks, but nearly all work that relies on judgment and language—precisely the types of abilities that are hardest to standardize and hardest to reallocate.

Some believe the gap will eventually narrow. Economic historian and professor at the Oxford Internet Institute, Carl Benedikt Frey, holds this view, based on history: the inequality caused by the adoption of computers gradually diminished over decades as barriers to usage declined. This analogy is not without merit.

The issue is that even accepting this optimistic historical analogy, Frey himself acknowledges a key caveat: "It depends on how long it takes for the gap to close. If it takes ten or twenty years, that’s more concerning."

Ten or twenty years is not a timeframe one can easily wait through—especially for those who need to find jobs, negotiate salaries, and gain experience during that period.

This is a peculiar moment in history: for the first time, we have a technology that makes everyone feel like they’re becoming smarter.

This feeling is often the end.

The problem is that, in an era where outcomes are truly determined by judgment, treating intuition as the end goal may be the most costly mistake of all.