AI has redefined what it means to be a good teacher.
AI entering classrooms is no longer news. The AI lesson preparation coverage rate in China's K-12 stage exceeds 30%, reaching nearly 100% among top public schools, and lesson preparation time for new teachers has dropped by more than 60%.
The hottest discussion around this issue has always pointed to the same question: Will teachers be replaced by AI?
Over the past few years, the number of 85 million teaching positions worldwide has not decreased due to AI, while discussions on "what makes a good teacher" have emerged in endlessly. More and more people recognize that those most easily overlooked judgments and physical presence of good teachers, as well as the firm belief that "I believe I can influence children", can never be taken over by AI.
This article intends to go a step further: the stronger AI becomes, the necessity of teachers' presence will not weaken, but rise instead.
A 30-point Idea and a 70-point Answer
One action is worth paying attention to: a student comes up with an idea of their own, and then turns to AI.
Teachers at Shenzhen Mingwan School summed up this moment as the "30-70-90 Point Challenge" ¹: the student's own idea is worth 30 points, and AI can instantly push it to 70 points. However, the 20 points from 70 to 90 that should have been earned by the child himself are erased at this moment.
The example given by this teacher is about several children in a summer camp who wanted to make a small device that could generate emotional resonance with the plants they raised. They used AI to generate the first version of the solution, which was very complete and decent. Then they lost interest in this device.
No one is replaced in this scenario. AI only completes the work from 30 points to 70 points, and does it quickly and well.
But in this process, who will remind the children that only 30 points of the result belongs to them?
AI has not replaced the teaching profession, but one of the most valuable jobs for teachers in the past, which is "explaining knowledge clearly", has now become "judging when to leave thinking opportunities to students".
Existing studies have also demonstrated this point. Researchers from the Wharton School of the University of Pennsylvania conducted a randomized controlled experiment covering nearly 1,000 high school students in Turkey ². The students were divided into three groups to do math exercises: the control group that did not use AI, the first group that used the standard ChatGPT interface (hereinafter referred to as the "standard group"), and the group that used the "guardrail" version (the guardrails were designed by teachers to make AI only provide clues instead of giving direct answers) (hereinafter referred to as the "guardrail group").
In the practice phase, both AI groups outperformed significantly: the standard group scored 48% higher than the control group, and the guardrail group scored 127% higher than the control group. Then the researchers took away AI, and the three groups took the same closed-book exam: the score of the standard group was 17% lower than that of the control group, while the performance of the guardrail group was basically the same as that of the control group.
The researchers believed that students treated AI as a "crutch". They did solve the problems during practice, but did not learn the problem-solving skills themselves. The study also found that students who used AI were overly optimistic about their learning abilities, even top students mistakenly thought they had mastered the knowledge.
With the same model, the same students, and the same questions, whether someone has designed the AI's response method in advance according to teaching rules affects students' final performance.
What AI Takes Away and What AI Cannot Take Away
Any discussion about teacher transformation must first admit that AI is really capable.
At Shenzhen Mingde Experimental School, AI has penetrated into every detail of teaching: there is a "three-round preparation for one lesson" system for collective lesson preparation, and AI automatically summarizes the discussion speeches of all teachers; classroom videos are submitted to large models to analyze time allocation, student head-up rate, question quality, and teacher movement heat map; teachers have uploaded a total of about 87,000 courseware, which are stored in the school-based resource library after review, and new teachers and paired support schools can directly use them. For more details, the originally considered very complex school-wide performance analysis report was completed in only 20 minutes with the help of AI.
What is more noteworthy is that the principle of "teaching students in accordance with their aptitude" has for the first time had a specific practical reference. In the past, it was impossible for a teacher to prepare three sets of assignments for one class, but now by inputting "convex lens imaging in Guangdong Province high school entrance examination", AI will automatically stratify the assignments according to "basic consolidation → experiment exploration → drawing", so that the top 20 students complete all assignments, the middle 20 students complete the first two parts, and the last 20 students only complete the first part. In the past, it was difficult to design assignments for students with such fine granularity.
Cost is no longer a threshold. Studies show that the cost of AI reasoning has dropped by 97% to 99% compared with 2023 ³.
What AI takes away is standardized outputs, and what it cannot take away is judgment, physical presence, and the human-to-human influence between people.
These contents sound somewhat comforting, but in fact there are very solid data supports in educational research. Education scholar John Hattie synthesized thousands of meta-analyses to rank various factors affecting student performance. The first place is neither technology, nor curriculum or teaching materials, but "collective teacher efficacy", which means the belief shared by all teachers in a school that "we can influence students' achievements". Its effect size is 1.57, more than three times that of family socioeconomic status ⁴.
Fang Taide, Executive Principal of Mingwan School, made a sharp metaphor for this "adult belief": if we do not let students' brains work hard, it is like letting athletes stand on an escalator while insisting that they are exercising, which is self-deception.
The New Criteria for "Good Teachers"
So, what are the new criteria for judging a "good teacher"?
Tencent Research Institute and Chinese Academy of Educational Sciences proposed a three-tier framework for teachers' AI literacy in a joint study, which can be used as a measurement scale.
The first tier is to know whether AI is reliable. Teachers can judge whether the answers given by AI are correct, know that AI may generate plausible but wrong content, and can explain this clearly to students.
The second tier is to be able to direct AI to work. For lesson preparation, grading, designing assignments, and learning situation analysis, let AI generate the first draft, and teachers act as reviewers.
The third tier is to be able to redesign a whole course with AI. Re-decide how the course is arranged, which path each student takes, and what evaluation method to use, instead of letting AI just reformat the old lesson plans.
The tiered structure itself is not new. What is really worth mentioning is a judgment we repeatedly confirmed in the survey: most teacher trainings stay at the first two tiers, and the training for the third tier is almost blank. Most of the AI-enabled teacher courses on the market also teach the second tier: how to write prompts? How to generate courseware in ten minutes?
What does the third tier look like? We have observed four dimensions through our research: clarify the teaching plan at the motivation level, evolve from tool users to tool creators at the competence level, be willing to transfer the right of questioning to students at the power level, and integrate moral education and psychological guidance into the curriculum structure at the vision level.
The difficulty increases sequentially, and none of them are technical problems.
The Gap Is Widened by the Saved Time
The time saved by AI is a fork in the road.
The act of saving time itself has no position. With the same ten minutes spent on generating a full set of courseware by AI, for the remaining two hours, some teachers will chat with those few children who never raise their hands, while others will generate five more sets of courseware.
The "crutch" effect does not only occur on students. The OECD summarized several studies in the *2026 Digital Education Outlook*, mentioning a phenomenon called "metacognitive laziness" ⁵: students who turn to human experts will go through the complete chain of "diagnosing the problem → seeking help → evaluating the help → iterating → implementing"; when facing chatbots, many people directly ask for answers and use them directly, skipping the three steps of diagnosis, evaluation and iteration.
Teachers will fall into the same trap: they distribute the assignments generated by AI without checking, and send out the comments written by AI without modification. The tool saves effort, but also makes people lose their judgment inadvertently.
We observed a phenomenon in the survey, which we tentatively call "super teachers": AI greatly expands the influence radius of a small number of excellent teachers — the teaching design, evaluation ideas and diagnosis methods of a top teacher can be reused for far more students than in the past through AI. This creates a scissors difference: the risk of substitution for "average-level teachers" is rising, and the scarcity of "super teachers" is also increasing.
The so-called "super teachers" will always be a minority. The real foundation of every school is thousands of ordinary teachers who are unknown and carefully teach every cohort of students. AI should not be a knife hanging over their heads, but a ladder handed to them — so that more conscientious and responsible good teachers can be seen by more children, and get the recognition they deserve.
Economics has given a specific number to the value of a good teacher. Three scholars from Harvard University, Raj Chetty et al., tracked the school district and tax records of more than 1 million children, and found that students assigned to high value-added teachers are more likely to go to college and have higher income in adulthood. Their quantitative conclusion is: replacing a teacher in the bottom 5% of value-added level with an average-level teacher can increase the present value of the lifetime income of the students in this class by about $250,000 ⁶.
The arrival of AI will only make this gap spread faster and cover a wider range.
The gap is not caused by AI. AI only turns the originally invisible choices into visible results.
AI distributes the same amount of time dividend to every teacher, but does not stipulate how this dividend should be used.
The US edtech sector has accumulated about 290 billion US dollars in investment, and venture capital in this sector fell to the lowest level in ten years in 2025. The number of companies that can achieve stable profitability can be counted on two hands. There are many reasons, and one of the key reasons is that school districts bought the tools, but teachers did not use them.
Indeed, the tool itself is never the variable. People are.
A 40-Year-Old Question and Its Answer Today
At this point, we have actually encountered a very old question in pedagogy.
In 1984, Benjamin Bloom, an educationalist at the University of Chicago, published a famous paper ⁷. His two doctoral students conducted a set of controlled experiments to compare three scenarios: regular classroom teaching, mastery learning, and one-on-one tutoring. The results showed that ordinary students who received one-on-one tutoring finally scored about two standard deviations higher than the average level of the regular class, surpassing 98% of the students in the regular class.
Thus he put forward the "two sigma problem" that has been questioned for 40 years: one-on-one tutoring is too expensive for most societies to afford on a large scale... Can researchers and teachers design such teaching conditions so that most students can achieve the achievement level that can only be achieved under tutoring conditions in collective teaching?
Today, for the first time, AI makes it economically possible to have "one tutor for every child".
The experiment of the University of Pennsylvania complements a premise that Bloom may not have anticipated back then: cheap one-on-one tutoring does not automatically equal effective one-on-one tutoring.
So Bloom's problem has a new version: we no longer lack tutors, but we lack people who know how to tutor.
That person is the teacher.
This is the full meaning of "the stronger AI is, the more teachers' presence is needed". The machine paves the way from 30 points to 70 points, and the last 20 points that determine where a child can ultimately go still require a person standing by, judging when to step in and when to step back.
References:
[1] Tencent Research Institute. Three Things AI Cannot Teach [EB/OL]. (2026-06-26)[2026-09-10].
[2] BASTANI H, BASTANI O, SUNGU A, et al. Generative AI without guardrails can harm learning: evidence from high school mathematics[J]. Proceedings of the National Academy of Sciences, 2025, 122(26): e2422633122.
[3] Tencent Research Institute. Co-Education with Humans, Move Forward for Good: Education Reform in the AI Era [R]. 2026.
[4] HATTIE J. Mindframes and maximizers[Z]. 3rd Annual Visible Learning Conference. Washington DC, 2016-07.
[5] OECD. OECD digital education outlook 2026: exploring effective uses of generative AI in education[R/OL]. Paris: OECD Publishing, 2026[2026-09-10].
[6] CHETTY R, FRIEDMAN J N, ROCKOFF J E. Measuring the impacts of teachers II: teacher value-added and student outcomes in adulthood[J]. American Economic Review, 2014, 104(9): 2633-2679.
[7] BLOOM B S. The 2 sigma problem: the search for methods of group instruction as effective as one-to-one tutoring[J]. Educational Researcher, 1984, 13(6): 4-16.
Zhang Hongru Researcher, Tencent Research Institute
Wang Peng Senior Expert, Tencent Research Institute
Lv Jiayi Researcher, Tencent Research Institute
This article is from the WeChat official account "Tencent Research Institute" (ID: cyberlawrc), author: Tencent Research Institute, authorized for release by 36Kr.