Artificial intelligence is no longer a novelty in classrooms. In the article cited by MarsBit, Tencent Research Institute said AI-assisted lesson planning now covers more than 30% of China’s K-12 segment, while adoption at leading public schools is close to 100%. It also said preparation time for new teachers has fallen by more than 60%.
The debate around that shift has stayed fixed on one question: will teachers be replaced by AI? The article’s answer is no. It says the jobs of the world’s 85 million teachers have not shrunk because of AI. What has changed is the definition of a good teacher.
The jump from a 30-point idea to a 70-point answer
The article starts with a classroom moment that is becoming common: a student comes up with an idea, then opens AI.
A teacher at Shenzhen Mingwan School described that moment as a “30-70-90 challenge.” In that framing, the student’s own idea is worth 30 points. AI can instantly push it to 70. The problem is that the 20 points between 70 and 90, the part a child should earn through thinking, can vanish just as quickly.
The example given involved several children at a summer camp who wanted to build a small device that could form an emotional resonance with the plants they were raising. They used AI to generate a first proposal. It looked complete and polished. Then they lost interest in it.
No one was replaced in that scene. AI simply finished the work from 30 to 70, and did it fast. The unresolved question is who tells students that only 30 of those points were really theirs.
The article argues that AI has not replaced the teaching profession, but it has shifted one of the profession’s central tasks. One of the most valuable parts of teaching used to be explaining knowledge clearly. Now it is increasingly about judging when students should be left with the chance to think on their own.
The same model can help or hurt, depending on how it is used
To support that argument, the article cites a randomized controlled study by researchers at the Wharton School of the University of Pennsylvania. The experiment was conducted in Turkey and covered nearly 1,000 high school students.
Students were split into three groups for math practice: a control group with no AI, one group using the standard ChatGPT interface, and one using a teacher-designed version with guardrails that only offered hints instead of direct answers.
During the practice phase, both AI groups outperformed the control group. The standard group scored 48% higher than the control group, while the guardrails group outperformed the control group by 127%.
Then researchers removed AI and gave all three groups the same closed-book test. At that point, the standard AI group scored 17% lower than the control group, while the guardrails group performed roughly in line with the control group.
The researchers said students may have used AI as a “crutch.” They solved the practice problems, but did not learn how to solve them. The study also found that students using AI became overly optimistic about their own learning ability, including top-performing students who wrongly believed they had mastered the material.
The article uses that result to make a narrower point: with the same model, the same students and the same questions, outcomes still depend on whether someone has designed the AI’s response style around how learning works.
What AI takes, and what it does not
Any discussion of teacher transformation has to begin by admitting that AI is very capable.
At Shenzhen Mingde Experimental School, the article says, AI has entered the fine-grained mechanics of teaching. The school uses a “one lesson, three preparations” system for group planning, with AI automatically summarizing teachers’ discussion notes. Classroom recordings are analyzed by large models for time allocation, how often students look up, question quality, and teacher movement heat maps. About 87,000 teacher-uploaded slide decks have entered the school’s resource library after review, and those materials can be used directly by new teachers and partner schools. Even a schoolwide performance analysis report that had once been seen as complicated was completed in 20 minutes with AI.
The article says AI has also made differentiated instruction more concrete. In the past, one teacher could not realistically prepare three separate homework sets for one class. Now, by entering “image formation by convex lenses on the Guangdong senior high school entrance exam,” AI can layer the assignment into “basic consolidation → experimental inquiry → drawing.” The top 20 students can complete all sections, the middle 20 can do the first two, and the bottom 20 can focus on the first section only. That level of granularity was difficult to achieve before.
Cost is no longer the same barrier, either. The article cites research saying AI inference costs have fallen by 97% to 99% compared with 2023.
Still, its central claim is that AI takes standardized output. It does not take judgment, physical presence, or one person’s influence on another.
Research places collective teacher belief above technology and textbooks
That claim is backed in the article by education research. It cites scholar John Hattie, who synthesized more than a thousand meta-analyses ranking factors that shape student achievement.
According to the article, the top factor was neither technology nor curriculum nor teaching materials. It was “collective teacher efficacy,” the shared belief among teachers in a school that they can influence student achievement. The reported effect size was 1.57, more than three times the effect of family socioeconomic status.
Fang Taide, executive principal of Mingwan School, used a stark analogy for that kind of adult conviction. If students are not allowed to make their brains work, he said, it is like placing athletes on an escalator and insisting they are exercising. That is self-deception.
A new benchmark for a good teacher
The article then moves to what it calls a new benchmark for good teaching.
It points to a three-layer AI literacy framework for teachers proposed in joint research by Tencent Research Institute and the Chinese Academy of Educational Sciences.
- The first layer is knowing whether AI is reliable: being able to judge whether its answers are right, understanding that it can fabricate information in a confident tone, and explaining that clearly to students.
- The second layer is being able to direct AI to work: using it for lesson preparation, grading, question design and learning analytics, with AI producing a first draft and the teacher acting as editor.
- The third layer is being able to redesign a course with AI: deciding how the course should run, which path each student should take, and how learning should be assessed, instead of simply asking AI to reformat an old lesson plan.
The article says the real issue is not the framework itself, but where training stops. Most teacher training still sits in the first two layers. Training for the third layer is almost blank. Most courses marketed as AI empowerment for teachers focus on the second layer as well, such as writing prompts or generating slides in 10 minutes.
Based on the authors’ fieldwork, that third layer has at least four sides: motivation, where teachers first need to think through how a class should actually be taught; capability, where they move from using tools to making tools; authority, where they are willing to hand the right to ask questions back to students; and perspective, where moral education and mental health enter the course structure.
The article says those demands rise in difficulty one by one, but none of them is merely a technical issue.
The time AI saves can widen gaps between teachers
Time saved by AI is presented as a fork in the road.
Saving time has no built-in direction. Two teachers can both use AI to finish a slide deck in 10 minutes. One then spends the next two hours talking with students who never raise their hands. Another uses the same time to make five more decks.
The “crutch” effect does not stop with students. The article cites the Organisation for Economic Co-operation and Development’s Digital Education Outlook 2026, which summarizes research on what it calls “metacognitive laziness.” Students who ask human experts for help tend to go through a full chain of diagnosing the problem, asking for help, evaluating the help, iterating, and then implementing. With chatbots, many users skip directly to requesting an answer and applying it, leaving out diagnosis, evaluation and iteration.
Teachers can fall into the same pattern, the article says: sending out AI-generated questions without reviewing them, or forwarding AI-written comments without editing them. The tool saves effort, but judgment can drop out with it.
In the authors’ fieldwork, another pattern appeared, described as the “super teacher.” AI expands the radius of influence for a small number of exceptional teachers. A top teacher’s lesson design, commentary logic and diagnostic method can be reused through AI across far more students than before. That creates a widening gap: replacement risk rises for average teachers, while the scarcity of “super teachers” keeps increasing.
The article adds that these “super teachers” will always be few. The real foundation of schools is the much larger number of ordinary teachers who quietly guide each cohort of students through school. AI, it says, should not hang over them like a blade. It should be a ladder in their hands, giving more diligent teachers a chance to be seen by more children.
Economics has tried to put a number on teacher value
The article also asks what a good teacher is worth in measurable terms. It cites research by Raj Chetty and co-authors at Harvard University that tracked school district and tax records for more than 1 million children.
Students assigned to high value-added teachers were more likely to attend college and earned more in adulthood, the article says. The researchers’ quantitative estimate was that replacing a teacher in the bottom 5% of value-added performance with an average teacher would raise the present value of lifetime earnings for that class by about $250,000.
According to the article, AI will only make the transmission of those differences faster and broader. The differences were not created by AI. AI makes choices that were once hard to see show up as visible outcomes.
Every teacher receives the same time dividend from AI, it says, but no one tells them how to spend it.
The article also points to the business side of educational tools. It says about $29 billion has been poured into the U.S. education technology sector, yet venture investment in the field fell to a 10-year low in 2025, and the number of companies with stable profitability can be counted on two hands. There are many reasons, but one important one is that school districts bought the tools and teachers did not use them.
Its conclusion on that point is blunt: tools are never the variable. People are.
Bloom’s 40-year-old question gets a new answer
The article closes by returning to a classic education question.
In 1984, University of Chicago education scholar Benjamin Bloom published the paper that framed the well-known “2 sigma problem.” In an experiment conducted by two of his doctoral students, researchers compared regular classroom instruction, mastery learning and one-to-one tutoring.
The result was that ordinary students who received one-to-one tutoring scored about two standard deviations above the average student in conventional classes, outperforming roughly 98% of students in those regular classrooms.
Bloom then asked the question that has followed education research for four decades: since one-to-one tutoring is too costly for most societies to provide at scale, can researchers and teachers design instructional conditions that allow most students in group settings to reach the level now seen only under tutoring?
The article says AI has now made “one tutor for every child” economically possible for the first time. But the University of Pennsylvania experiment adds a condition Bloom may not have anticipated: cheap one-to-one tutoring does not automatically become effective one-to-one tutoring.
That gives Bloom’s question a new version. Society no longer lacks tutors. It lacks people who know how tutoring should be done.
The article’s answer is simple: that person is still the teacher.
Its final point is that this is what it really means to say stronger AI creates a greater need for teacher presence. Machines can pave the road from 30 to 70. The last 20 points, the part that determines how far a child can really go, still require a person standing nearby and deciding when to step in and when to step back.
Source and byline
The article originated from the WeChat public account Tencent Research Institute (ID: cyberlawrc). It was written by Zhang Hongru, researcher at Tencent Research Institute, Wang Peng, senior expert at Tencent Research Institute, and Lyu Jiayi, researcher at Tencent Research Institute. It cited work and papers from Tencent Research Institute, the Chinese Academy of Educational Sciences, the Wharton School of the University of Pennsylvania, John Hattie, Raj Chetty and Benjamin Bloom.