We are introducing AI-based writing correction into a product under development. We had already added experimental features, but before full introduction we checked quality by directly comparing AI and humans. As a result, I recognized AI’s limits and human strengths again. This article considers AI’s strengths, human strengths, and coexistence between AI and humans through AI correction quality checks.

AI’s strengths

With the progress of AI, efficiency in education is improving dramatically. In correction work, AI can quickly and accurately point out grammar errors, suggest better word choices, and propose revisions. Its overwhelming advantage is that it can process large amounts of data instantly.

AI can also learn the rules and patterns behind errors and use them in future correction. For example, we prepare correction tips for each essay topic. Humans tend to skip reading all of them every time, but AI can respect them every time and propose corrections specialized for the topic.

That kind of individualized handling quickly reaches a limit when humans do it manually. A person has to read each essay, check the rules, connect them to the characteristics of the topic, and then produce useful feedback. AI can perform those steps at low cost and with stable speed. At least for “carefully processing a large volume of text according to fixed rules,” AI is already very strong.

In short:

• AI performs far better than humans who work carelessly.
• Humans who work carefully can still beat AI.
• AI can process massive data instantly, so correction costs may become 10 to 100 times more efficient.

This is only the current state. AI will continue improving for several years, and cost will certainly become more efficient. The question is whether it will exceed the quality of humans who work carefully. I do not know yet.

Human strengths

When thinking about the root of education, there will continue to be areas AI alone cannot cover. That is the core of education: facing each student, understanding their individuality, emotions, and growth process, and guiding them appropriately.

AI is limited in deep contextual understanding and emotional care. It can process grammar and vocabulary, but it may not read the background or intention behind a student’s writing. If a student writes about their own experience and expresses emotional conflict, AI may treat it only as a structural issue. Humans can pick up the psychological and cultural context behind it.

I boldly predict that the definition of excellence may shift from high intelligence to high empathy, or high ability to revise one’s own perception. AI behaves as if it recognizes errors when corrected, but this is interaction. Humans can empathize with others in real time and revise their perception introspectively.

Humans also draw out student motivation and encourage autonomous learning. No matter how accurate AI feedback is, if the student is not psychologically ready or motivated, the effect is limited. Teachers recognize effort, praise progress, and communicate the joy and achievement of learning. AI cannot truly transmit empathy, so teachers remain essential.

In other words, the relationship between AI and humans matters. AI can provide efficient feedback, but whether students accept it and turn it into the next action still depends heavily on trust between people.

The relationship between AI and humans

Education depends on human relationships. It is not just knowledge transfer, but exists within trust between students and teachers, and among students. Teachers become role models and help students learn social values and ethics. This is hard for AI to replace.

AI can reduce daily burdens and improve efficiency. But roles that require deep understanding of student growth, empathy, and motivation should remain human responsibilities. AI and humans should complement each other to improve education.

While thinking about this, I encountered J. C. R. Licklider’s paper “Man-Computer Symbiosis.” It proposes a relationship where humans and computers complement each other. Licklider contrasted this with machines extended by humans and warned that in large computer-centered information control systems, human operators would mainly handle functions not worth automating.

• Computing history article: https://kyototextlab.org/2024-12-02-licklider-bbn/
• Man-Computer Symbiosis: https://groups.csail.mit.edu/medg/people/psz/Licklider.html

Current AI built with large language models may be closer to machines extended by humans. In writing-correction checks, I noticed that people tolerate human mistakes to some degree but are strict about AI mistakes. Since large language models are probabilistic, mistakes and hallucinations can occur. But if humans simply check those mistakes, we move toward a future filled with humans doing tasks not worth automating.

Licklider’s symbiosis is a model in which computers support human intellectual activity, while humans use computational power to create new knowledge and discoveries. But if current LLM products are designed carelessly, humans become assistants to machines. AI generates at scale, and humans check the mistakes. Humans end up doing the parts that are not worth automating, but still require someone to take responsibility. That is not the future I want.

In the writing-correction quality checks, the hard problem was how to handle AI mistakes. If AI makes one mistake, users immediately doubt the whole system. If a human makes one mistake, it is often treated as an individual error. To make AI usable, humans need to check it. But if checking becomes the main labor, much of the value of efficiency is lost. This is the point that requires the most care when designing coexistence between AI and humans.

Summary: what we expect from AI

Human strength lies in existing as an organism and in human diversity. We need to think about how to maintain that. In coexistence with AI, it is important to tolerate AI mistakes while not trusting AI too much. Otherwise the future becomes one of machines extended by humans. What future we get depends on what we expect from AI.

We should expect AI to reduce burdens and expand possibilities, not to push humans behind machines as error-correction labor. This is especially true in education. Correction efficiency can improve, but student growth, motivation, trust, and empathy still need human responsibility. The question we need to keep asking is how to place AI where it makes humans more human, not more machine-like.