Why meaningful educational AI must remain teacher-led

Teacher judgment is an educational requirement

“AI in education” can describe very different things: drafting a lesson outline, organizing administrative work, suggesting a question, delivering an answer, recommending a learning path, or making a consequential evaluation. Grouping all of these together hides the central question: who can understand and change what happens to a learner?

Teaching requires contextual judgment. A teacher can connect a student's response to prior conversations, confidence, language, relationships, and the purpose of the lesson. An AI system works from the information and model it has been given. That can be useful, but it is not the same kind of responsibility.

The U.S. Department of Education's 2023 guidance argues that people should remain in the loop: educators should retain agency over instructional decisions, be able to inspect and override automated recommendations, and provide human consideration and recourse.¹

The human loop must be real

A teacher is not meaningfully “in the loop” if the system's conclusion is opaque, the interface nudges approval, or changing the recommendation creates more work than accepting it. A systematic review of automation bias, written with a focus on healthcare but drawing on studies from aviation, human-computer interaction, and other fields, found that people often follow automated advice even when it is wrong, and that workload and time pressure make this more likely.² A policy analysis of human-oversight requirements adds a caution: a nominal review step can give false reassurance about a flawed system, so the earlier question is whether the tool should be adopted at all.³

In Dwyl's framing, meaningful control has at least six observable dimensions:

1. Visibility. Teachers and students can tell where AI is involved and what information it used.
2. Editability. A teacher can change generated plans, feedback, or scaffolds before they shape learning.
3. Override. A recommendation can be rejected or stopped without hidden penalties.
4. Traceability. Important actions and changes can be reviewed later by the people accountable for them.
5. Appropriate defaults. The safest pedagogical choice does not depend on every teacher finding a buried setting.
6. Recourse. A learner can reach a person who can reconsider an error or harmful outcome.

These dimensions make “teacher-led” testable. They also reveal failure modes. A dashboard that shows a score but not its basis offers visibility without explanation. An editable suggestion that has already been sent to a student offers editability too late. An appeal route that no learner can find is not meaningful recourse.

Four loops educators should retain

Retaining these loops is not only technical work. UNESCO's AI competency framework for teachers organizes teacher competency across a human-centred mindset, ethics, AI foundations, AI pedagogy, and professional learning.⁴

Planning

AI may help a teacher surface options, draft a sequence, or notice a missing consideration. The educator still defines the learning goal, decides what belongs for this class, and checks the result against curriculum, evidence, and student context.

Student support

Learning support should protect productive struggle. A useful scaffold can ask a question, point to evidence, offer a strategy, or help a learner break down the next step. Simply producing the answer may complete the task while bypassing the capability the task was designed to build.

AI can also support the adult rather than the learner. In a US randomised trial of Tutor CoPilot, reported in a working paper that has not been peer-reviewed, tutors could request expert-informed suggestions and choose or edit them. They used more guiding questions and gave away answers less often, and their students were 4 percentage points more likely to master the session's topic. The study found no significant effect on end-of-year standardised maths scores.⁵

Evaluation

AI can help organize evidence or flag patterns for attention. It should not become the unaccountable final judge of a student's work, placement, opportunity, or need. Teachers need access to the underlying work, the reasoning behind a recommendation, and the ability to make a different decision.

Visibility for the teacher can matter in its own right. In a classroom-randomised experiment with 286 US middle-school students, giving teachers real-time analytics from an intelligent tutor increased students' learning with that tutor. The effects were small and the study lasted about one week.⁶

Recourse

Students and families need to know when AI affected an interaction and how to ask for human review. Recourse is especially important when a system is wrong, biased, confusing, or applied outside the situation it was designed for.

Learning support is not answer delivery

The OECD's 2026 outlook distinguishes immediate performance from durable learning. It warns that general-purpose GenAI can improve the quality of an output without necessarily producing a learning gain, and recommends purposeful, pedagogically guided use that does not replace the cognitive effort a task is meant to elicit.⁷

For educators, that suggests a simple design test: after the AI interaction, what thinking does the learner still have to do? If the tool removes the exact reasoning, retrieval, explanation, or revision the task was meant to exercise, convenience has displaced the learning goal.

Teacher-led use can look different:

• ask for a hint that leaves the decision with the learner;
• compare two possible strategies and explain the trade-off;
• identify evidence that needs checking;
• rehearse a question before asking a peer or teacher;
• receive feedback on clarity while preserving ownership of the idea;
• make AI involvement visible in the student's process record.

Questions to ask an educational AI provider

Can educators see what the student saw?

Ask whether prompts, responses, sources, and interventions are visible in a form that supports teaching rather than surveillance.

Can a teacher change or stop the behavior?

Look for controls that work at the right level—class, task, learner, and moment—not only a global preference.

Does the system explain recommendations?

An unexplained score or next-step suggestion is hard to evaluate. Ask what evidence informed it, what the model cannot know, and how uncertainty is communicated. The automation-bias review lists showing confidence and presenting information, rather than directive recommendations, among the factors that may reduce over-reliance.²

What happens when the system is wrong?

Look for correction, escalation, audit, and human-review paths. Ask whether a learner can contest an outcome without being forced through the same automated process.

Does the default preserve student thinking?

Test whether the tool tends to ask, hint, and scaffold before it answers. “Helpful” should not quietly mean completing work for the learner.

What is logged, retained, and reviewable?

Teachers and leaders should understand which actions are recorded, who can access them, how long data is retained, and how the design aligns with the school's privacy obligations. Our privacy page explains Dwyl's current public commitments.

A practical standard

Teacher-led educational AI should make professional judgment easier to exercise, not easier to surrender. A useful system helps educators see more clearly, prepare thoughtfully, and support learners at the right moment. It leaves goals, consequential decisions, and responsibility with people.

That standard is more demanding than putting a person near an automated process. It asks whether educators and learners can understand the system, challenge it, change it, stop it, and reach a human when it matters.

Continue with Dwyl.studio's public responsible AI principles, the practical guide to responsible AI in project-based learning, and the educator and school pathways. Our building-stage invitation explains how to join the conversation. For the wider learning argument, return to real-world learning in the AI era.