Personally, I think AI is essentially a tool coefficient with a lower cost of acquisition. It lowers the material threshold for expert-level resources: one-on-one access that used to require money or people now costs less. That is real progress, but what it does is multiplication, not avg
Multiplication depends on the base. The base is the person, not the tool. The base is the minimum knowledge threshold at which a person can make an effective judgment about AI output. Take coding: someone who knows nothing about programming asks AI to write a complete application and gets a black box they can neither evaluate nor fix, with the result entirely down to gacha. Someone with a foundation asks AI to complete a specific module, and when something goes wrong they can read it and fix it. When the base is zero, no matter how large the coefficient, the product is still zero
The threshold varies by field. This threshold differs by orders of magnitude between fields. In fields whose output is natural language and evaluable with everyday common sense — legal basics, explanations of medical symptoms, administrative procedures — the threshold for effective judgment is close to basic literacy, and avg in these fields may really be close to universal. And these happen to be exactly the places that traditionally had the most jargon and the most unequal information
So my judgment is this: on the dimension of “elite knowledge,” AI does multiplication and widens the gap between people. On the dimension of “basic information,” it does averaging, and is dissolving some of the oldest and most universal information exploitation. The two proceed in parallel without contradiction
For people with an active desire to learn, AI can also help raise the base itself. Make a mistake, follow up, understand, use it again — this loop is a real upward spiral. But the starting point of the spiral is always the person: you have to be willing to ask one more why when something goes wrong. The admission ticket is subjective initiative, and that itself is not distributed equally
The inclusive effect of AI is also constrained along three dimensions: on the supply side, the infrastructure gap has not been closed — infrastructure in mountainous areas is poor, and AI’s threshold has not been lowered for them at all. On the demand side, the gaps in ability and will between people persist. The quality of the tools themselves is unequal, and they flow structurally toward the stronger: the people who can reach better models often already have stronger information-acquisition ability and willingness to pay. The gaps between models like Doubao, deepseek and Claude really do exist
But the more fundamental problem is that knowledge equalization, as a systemic promise, was oversold from the very beginning
A hardworking person in a high-cost environment might be 50×1＝50. A lazy person in a low-cost environment might be 1×50＝50. The result is the same; the paths are completely different. A real leap only happens when subjective initiative and a high-quality environment work together. The policy logic of knowledge equalization is essentially optimizing the environment coefficient while pretending to solve the problem of the person. It’s just whack-a-mole
The upper layer doesn’t passively benefit from inequality; it actively produces and maintains it. When degrees lose value, people compete for elite schools; when elite schools become common, people compete for overseas ones; when AI lets more people write code, the competition shifts to architecture, judgment and taste — things that are harder to toolify and harder to equalize. Every step forward in equalization rebuilds the barrier at a higher level
This narrative keeps circulating because it is useful to every party: it gives internet companies moral endorsement, gives policymakers a lever, and gives ordinary people hope. Its function is not to solve the problem but to give inequality an acceptable explanatory frame — turning structural problems into personal ones, and power struggles into educational issues
“Equalization will never happen” is indeed hard to claim, but it is undeniable that it is far smaller, slower and more local than the way it was sold, and that it always faces structurally opposing forces
AI is genuinely dissolving the oldest information exploitation in some fields, and genuinely accelerating existing divisions in more places. It has not changed the fact that the person is the decisive variable, nor the mechanism by which barriers keep being rebuilt. It is worth pursuing, but as an achievable systemic end state, it has never truly held in the structure of human society
The arrival of AI will not make it hold; it will only make this clearer sooner. It is an accelerator, but an accelerator doesn’t consider direction
