Cognitive Disparities in an AI Society: The Shifting 'Bottleneck of Thought' Between Humans and AI
As generative AI permeates every corner of society, how will the disparities in human ability change?
Intuitively, two conflicting future scenarios come to mind. One is a future where high-performance AI is available to everyone, narrowing the gaps in knowledge and skills and leveling the playing field. The other is a future where those who are already highly capable use AI more skillfully, further extending their abilities and causing the gap to widen instead.
Looking at research to date, it is not the case that only one of these two is correct. Rather, while AI may narrow conventional ability gaps in the short term, it has the potential to create new disparities elsewhere in the long term.
The important thing is not to think of AI as simply 'amplifying' human ability. It is to consider that, with the progress of AI, the location of the constraints on the human side of intellectual activity is itself shifting. In this article, I would like to consider this phenomenon from the perspective of the 'shifting cognitive bottleneck'.
1. AI first narrows the gap in 'execution ability'
Generative AI is rapidly replacing tasks that previously required a certain level of knowledge or expertise, such as writing, searching, summarizing, translating, organizing information, and programming. In this sense, AI is not just a device that expands ability gaps. Rather, it has the effect of raising the floor for those who previously had lower abilities.
Even those who are not good at writing can create well-structured documents. Even those with little programming experience can write code. Even without full proficiency in a foreign language, one can access information from abroad. In short, AI first compresses the differences between humans in the 'execution' phase.
However, once execution ability can be easily supplemented, other abilities that were previously less visible become important.
For example, even if AI can write excellent text, deciding 'what should be written about' is a different ability. Even if one can gather a large amount of information, the ability to judge 'which information should be trusted' becomes separately necessary.
AI does not eliminate ability gaps. It shifts the location where the gaps are created. This is the starting point for thinking about cognitive disparities in an AI society.
2. Thinking of humans and AI as a single cognitive system
In an AI society, measuring only 'individual human ability' is becoming insufficient to explain actual intellectual activity. Current intellectual work is established through a combination of not just human knowledge, but also AI, search systems, databases, conversation history, and verification procedures.
Therefore, it is better to shift the unit of analysis from the 'human' to the cognitive system consisting of human + AI + information environment.
The functioning of this system can be broadly divided into seven stages. First, there is problem setting, which determines 'what the problem is.' Next, the problem is broken down, and a judgment is made as to which parts to think about oneself and which parts to leave to AI. From there, execution using AI begins, and the resulting answers are verified.
Furthermore, one must adjust how much to trust one's own judgment versus the AI's, and integrate the obtained results into one's own knowledge. Finally, one must decide what the purpose is in the first place and to what extent to delegate judgment to AI.
These abilities cannot simply be added together. If one part is extremely weak, the performance of the entire system is constrained there. Even if the AI is extremely high-performance, if the way the question is framed is wrong, the returned answer will be off-target. Even if a good question is framed, if one cannot verify the AI's errors, those errors will be adopted as is. This weakest part is the 'cognitive bottleneck'.
3. As AI progresses, the bottleneck moves upward
What current AI is particularly strong at is the 'execution' part: searching for information, generating text and images, writing code, and presenting a large number of candidates. Consequently, the ability to execute itself gradually becomes less rare. The new constraints that emerge then are, first, 'verification'.
AI can generate a large number of hypotheses in seconds. However, it takes time for humans to verify whether those hypotheses are correct. There is a major asymmetry here. Because the speed of generation increases rapidly, but the speed of verification does not increase in the same way. As a result, the bottleneck of intellectual activity shifts from the 'ability to create answers' to the 'ability to evaluate answers'.
Furthermore, if AI begins to support verification as well, the next important things will be judgments such as 'what should be verified,' 'to what extent should it be left to AI,' and 'is this question even worth thinking about in the first place?'
In other words, the progress of AI can be understood not as a process where the human role simply diminishes, but as a process where the core of the cognitive abilities required of humans shifts to a higher level. From possessing knowledge to formulating questions. From providing answers to verifying them. Furthermore, from verifying to governing the judgments of both oneself and AI. This shift is a key characteristic of cognitive inequality in an AI society.
4. 'Inequality of Results' and 'Inequality of Cognitive Capital' Are Not the Same
With the spread of AI, the difference in superficial results may shrink. Even if there is a difference in writing ability, both parties can produce polished documents using AI. Even if there is a difference in information organization ability, the deliverables will look similar if AI performs the summarization and structuring. Looking only at this, the inequality appears to have narrowed.
However, the results obtained using AI and the abilities formed within a human are not the same. For example, there is a significant difference in what happens on the human side between having AI summarize a paper and using that content as is, versus having AI present multiple interpretations, verifying them against the original text yourself, and reconstructing your thoughts after considering contradictions and counterarguments, even if the externally visible deliverables look similar.
In the former, AI is substituting for cognitive work. In the latter, the human's own thinking is being sharpened through dialogue with AI. What should be distinguished here is the ability to produce results while using AI and the cognitive capital that remains with the human even when AI is removed.
In an AI society, while the inequality of results may shrink in the short term, a difference in the accumulation rate of cognitive capital may emerge in the long term. This is harder to see than traditional inequality. The more AI homogenizes output, the harder it becomes to distinguish from the outside who truly understands and who is merely receiving AI output.
Therefore, cognitive inequality in an AI society does not necessarily expand in a visible way from the start. It can be called a 'latent inequality' that progresses behind the scenes of superficial leveling.
5. The Key Is the 'Ability to Correctly Evaluate Oneself and AI'
What becomes important at this stage is metacognition and 'calibration.' When using AI, a binary choice of 'trusting AI or not trusting it' is insufficient. In areas where AI is more accurate than humans, being overly skeptical is also an error. On the other hand, relying entirely on AI in areas where its reliability is low leads to incorrect judgments. What is needed is to judge how much it is appropriate to trust that AI for a given task.
Moreover, the same applies to oneself. It is necessary to grasp whether one's own understanding is sufficient, what one does not know, and where one can no longer make judgments on one's own.
What becomes important in an AI society is not mere knowledge quantity. More than 'what one knows,' 'how accurately one grasps what one does not know' becomes important.
Here, the difference in metacognition determines the quality of AI usage. Those who can grasp the limits of their own understanding verify AI answers, explore other possibilities, and return to primary sources if necessary. Conversely, if one overestimates one's own understanding or mistakes AI's fluent writing for truth, AI can become a device that reinforces false convictions. The important ability in an AI society is not 'not trusting AI,' but the ability to appropriately adjust trust in both oneself and AI.
6. AI Can Either Substitute for Thinking or Cultivate It
To summarize the discussion so far, there are two ways to use AI.
One is to ask AI when a problem arises and use the answer as is. This method is efficient. Short-term results also improve. However, the more AI takes on, the fewer opportunities humans have to think for themselves.
The other is to formulate hypotheses yourself, have AI provide counterarguments, elicit different perspectives, check materials, and finally reconstruct your own thoughts. In this case, AI is not an answer-providing device, but a device for shaking up and reconstructing thinking.
Even using the same AI, the cognitive results of these two approaches differ significantly. Therefore, the inequality that will become important in the future is not just between 'people who use AI' and 'people who do not.' More important than that is the difference between those who use AI as a substitute for thinking and those who use it for the formation of thinking.
In the former, even if results improve due to AI performance improvements, human ability formation does not necessarily keep up. In the latter, the improvement of AI performance itself becomes material for increasing human cognitive capital. If this difference accumulates over a long period, a large cognitive inequality may form even among people accessing the same AI.
7. What Is Cognitive Inequality in an AI Society?
Based on the above, cognitive inequality in an AI society can be considered in three stages.
First, AI will narrow conventional skill gaps. Next, the scarcity of ability will shift toward higher-level cognitive functions such as problem formulation, verification, judgment, calibration, and delegation to AI. In the long term, differences in how individuals engage with AI will lead to disparities in the speed at which human cognitive capital is formed. In other words, while performance gaps may narrow in the short term, there is a possibility that disparities in cognitive formation will widen in the long term.
The issue here is not merely 'AI literacy.' It is not about whether one can operate AI. More essential is the ability to understand, design, and modify one's own cognitive system, which consists of both human and AI components. This could be called 'cognitive agency.'
As AI becomes more advanced, the areas that humans must process directly will decrease. However, at the same time, the question of what to delegate to AI, where humans should intervene, and where to retain responsibility for judgment becomes even more important.
In this sense, the true cognitive disparity in an AI society is not a difference in simple knowledge or intelligence. It is better to think of it as the gap that arises between those who can govern their own cognitive processes and those who unconsciously externalize them to AI.
However, this is not a predetermined future. There remains room for humans to acquire better ways of using AI through AI interfaces, education, organizational system design, and verification procedures.
That is why the question to be asked in the future is not 'Are you using AI?' but rather how is your own thinking ability changing as a result of using AI? That is likely the most important point when considering cognitive disparity in an AI society.