LLMs and the Purpose of Education

An answer is evidence of learning only when the student has done enough of the thinking to understand it. That has always been an issue in education. Large language models (LLMs) make the distinction harder to overlook because producing a plausible response can now require very little engagement with the underlying problem.

What interests me is the question beyond whether schools should permit a particular tool. It concerns what students should become capable of doing, how that capability develops, and which forms of assistance support it.

Education involves acquiring knowledge, exercising judgment, and learning to participate with other people. LLMs may change how some of that work is carried out. Their value depends on whether the changes strengthen those capacities, including when the tool is unavailable.

Assistance and Learning Are Different Outcomes

An LLM-based tutor can offer another explanation, adapt an example, or provide practice at a pace that suits a student. Those are plausible advantages, and research gives reasons to investigate them carefully.

A randomized study in a Harvard undergraduate physics course, published in 2025, found stronger immediate learning gains with a purpose-built LLM-based tutor than with the comparison active-learning lessons. The analysis covered 194 eligible students and two lessons. That supports a claim about the tested intervention and setting, rather than a general claim that chatbots outperform teachers.¹

A different study in high-school mathematics, also published in 2025, found that access to an unrestricted LLM interface improved assisted practice performance but reduced performance on a subsequent unassisted exam relative to the control group. A tutor with learning safeguards largely removed that disadvantage, without establishing an unassisted-exam advantage over the control.²

Together, the studies illustrate why “LLMs in education” is too broad a category for a useful verdict. The design of the assistance, the task, and the measure of success all matter. Helping a student complete work during a session is distinct from improving immediate understanding or supporting knowledge they can later use independently.

Those outcomes need separate evaluation. A system can make practice look more successful while leaving the student less prepared for the next task. Another can provide useful guidance without demonstrating every benefit its supporters might hope for.

Decide What Students Still Need to Practice

Ready access to explanations does not remove the need for subject knowledge. Students need enough understanding to recognize a relevant method, notice a contradiction, and decide whether an answer fits the question. An explanation they cannot evaluate may be useful as a starting point, but it cannot establish that they understand the material.

That makes the placement of assistance important. A hint that helps a student identify the next step has a different educational role from a completed solution. Asking a student to explain why a method works provides different evidence from checking whether the final answer is correct.

The appropriate balance will vary. Someone encountering a topic for the first time may benefit from a worked example. Someone practicing a skill may need space to attempt it before receiving guidance. A student with a disability may require assistance that is essential to access the task. Independence should not be defined as withholding an accommodation.

A useful assessment therefore starts with the capability it is meant to examine. It can then distinguish permitted support from the part of the work the student must demonstrate. That distinction is more informative than assuming either that all assistance is cheating or that an accurate output settles the question.

Teaching Requires Judgment About the Situation

Teachers do more than deliver explanations. They choose what deserves attention, interpret misunderstandings, organize shared work, and decide when a student needs a challenge, reassurance, or a different approach.

Some of those activities may be supported by software. Their purpose does not disappear because an explanation can be generated. A student's hesitation might reflect a missing concept, unclear instructions, embarrassment, or something happening outside the lesson. Similar visible behavior can require different responses.

I would want better tools to preserve the classroom as a setting in which students' reasoning has to meet other people's reasoning. Explaining a method, listening to an objection, and contributing to a shared project are educational activities in their own right. Personalization should leave room for them.

UNESCO's guidance takes a human-centered approach to generative tools, including LLMs, in education, with attention to privacy, age limits, inclusion, and the pedagogical suitability of tools.³ Those considerations matter when deciding how a tool should be used, who is responsible for it, and what evidence would justify expanding its role.

Teachers need a part in those decisions. Treating their experience as feedback on a finished procurement choice misses an opportunity to shape the intervention around the actual work of teaching.

Families Need Context and Workable Responsibilities

Better communication between home and school requires more than a higher volume of messages. A parent may receive frequent updates while remaining unsure which changes matter, what they mean, or how to respond.

Useful communication identifies the issue and the responsibility attached to it. It gives a family enough context to participate without expecting continuous availability from either the parent or the teacher. Students should have an increasing role in explaining their work and understanding the decisions made about it.

Participation also has to fit different circumstances. Attendance at an afternoon meeting is a poor stand-in for commitment when working hours or caring responsibilities make attendance difficult. Several ways to participate may be necessary to make the relationship practical.

Tools can support this coordination, but the expectations need to be settled first. Otherwise, a new channel simply becomes another place where someone is expected to monitor, interpret, and respond.

Visibility Has an Educational Purpose and a Limit

Grades and assignments provide useful information. They do not capture every kind of progress worth discussing. A student may become more willing to ask questions, take responsibility for a project, or persist through an unfamiliar task.

Sharing appropriate parts of that experience could improve the conversations around the student. The purpose should determine the information collected and who receives it. A useful update is not a reason to build a comprehensive record of ordinary development.

Students need room to make mistakes, revise a view, and ask for help without assuming that every moment will be preserved or circulated. Access also needs to account for different family circumstances. Disclosure cannot be treated as automatically safe simply because the recipient is an adult.

Knowledge Needs Contact with Possibility

Schools also give students opportunities to encounter subjects, activities, and forms of work they might never have thought to search for.

The OECD's 2025 report on teenage career preparation describes uncertainty about future work and unequal access to career-development experiences. It reports associations between such experiences and later outcomes, which support further attention to exploration without proving that any one activity causes a particular result.⁴

A placement, a practical project, or a conversation with someone doing unfamiliar work can give a student something concrete to consider. LLMs may help them prepare, investigate, or reflect. That assistance has a different role from the encounter itself.

The educational question remains what students come to know and what they can do with it. Tools should be judged against that purpose, alongside the relationships and opportunities through which it is pursued. Automating more of the school day is an activity. Improving education requires evidence of what the student gains.

Sources and Notes

1. Greg Kestin, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti. AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports 15, 17458 (2025). ↩
2. Hamsa Bastani and colleagues. Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences 122(26), e2422633122 (2025). ↩
3. Fengchun Miao and Wayne Holmes, UNESCO. Guidance for generative AI in education and research. 2023. ↩
4. OECD. The State of Global Teenage Career Preparation. 2025. ↩