AI teacher vs AI teaching assistant: where the line actually is

The phrase "AI teacher" does a surprising amount of damage in a staff meeting. It puts every teacher in the room on the defensive, it makes leadership sound naive, and it turns a practical procurement conversation into a debate about whether machines can care about children. Nobody wins that debate and nothing gets bought or rejected on the merits.

The split that makes the conversation productive again is not philosophical. It is a job description. An AI teacher would have to hold the things teaching actually consists of: relationship, judgement, accountability and the authority to make decisions about a child. An AI teaching assistant holds none of those and does not need to. It does the bounded, repetitive, cognitively expensive work that surrounds teaching, and it hands the results to a person who is accountable for what happens next.

What an AI teacher would have to be able to do

Run the list honestly and the category collapses on its own.

• Hold a relationship over time. Notice that a normally chatty student has been quiet for four days and that this matters more than the missing homework.
• Exercise professional judgement under uncertainty. Decide that this class needs to abandon the plan and spend the period on something else, and be right often enough to be trusted.
• Take responsibility. Be answerable to a parent, a principal and a regulator for a decision about a child's placement, progression or wellbeing.
• Read a room. Detect that the silence is confusion rather than concentration, and that the loud group in the corner is anxious rather than disruptive.
• Make and keep commitments. Promise a student that this will be different this term and then arrange the conditions that make it true.

None of that is a capability gap that a larger model closes. They are properties of a person occupying a role inside an institution that holds them accountable. A system with no stake in the outcome cannot occupy that role regardless of how well it explains fractions.

What an AI teaching assistant can genuinely take off the desk

The useful work is unglamorous and enormous. A teaching assistant, human or otherwise, is defined by the tasks it can be handed without transferring accountability.

Before the lesson

• Draft differentiated versions of a task at three reading levels from the one you already wrote.
• Produce ten more practice items in the same style as the six you use every year.
• Turn a set of learning objectives into a rubric you then edit rather than write from scratch.
• Predict the three most likely misconceptions for a topic so the lesson plan has a plan for them.

During the lesson

• Coach individual students through practice with hints while the teacher works with a small group.
• Answer the eleventh identical procedural question so the teacher can answer the one that is different.
• Keep a student who finished early genuinely extended rather than idle.

After the lesson

• Read every student's work and name the specific misconception rather than just marking it wrong.
• Produce a per-skill mastery read with the student's own words as the evidence.
• Group the class by fault line so tomorrow is two small groups rather than one whole-class reteach.
• Draft the sentence that goes into a progress report, for the teacher to approve or rewrite.

Notice that every item in the third list is a task teachers already do badly, not because they are bad at it but because there are 140 students and one of them. That is the honest case for the category. It is not that AI does it better than a teacher would with unlimited time. It is that a teacher does not have unlimited time and currently does a version of this work at 9pm, for a subset of students, from memory.

The line: who is accountable for the decision

The clean test for whether a task belongs on the assistant side is not difficulty or subject matter. It is whether the output is a decision or an input to a decision.

Assistant (input)	Teacher (decision)
Names the misconception in a piece of work	Decides whether the diagnosis is right and what to do about it
Produces a per-skill mastery read	Decides the grade and what it means for this student
Suggests a five-minute reteach	Decides how tomorrow is spent
Flags that a student has stalled on the same skill three sessions running	Decides whether this becomes an intervention referral
Drafts a sentence for a progress report	Signs the report

Written down, the split is obvious. It stops being obvious in practice when a product's output is opaque, because an unverifiable suggestion quietly becomes a decision. If a teacher cannot check the reasoning in a few seconds, they will either ignore the tool or defer to it, and both failure modes are bad.

This is why evidence lines matter so much more than confidence scores. A mastery read that says "Developing, 47%" is a number to be trusted or ignored. The same read with "paraphrases the passage but never anchors a claim to a line" attached is a claim a professional can check against the work in five seconds. That mechanic is described in detail on how it works.

Three failure modes to watch for

1. The assistant that quietly teaches

A tool that answers student questions freely, with no teacher-facing output, has taken over instruction without taking over accountability. Nobody chose this; it happens by default when the student-facing surface is good and the teacher-facing surface is an afterthought. The symptom is that the teacher cannot say what any individual student learned this week.

2. The teacher tool with no student in it

The opposite failure. A suite of planning and drafting tools for adults is genuinely useful and does not touch the hardest part of the job, which is knowing what each of 140 students actually understands. Do not mistake a prompt library for a tutoring platform. The category comparison lays out where each kind of product actually sits.

3. The record that nobody can defend

A per-skill dashboard with no visible evidence behind it is worse than no dashboard, because it will eventually be used in a meeting where a parent asks how you know. If the answer is "the system said so", the meeting goes badly and the tool is finished at that school. Any progress claim needs the student's own words underneath it, which is also what makes it useful in an MTSS context. See student progress tracking.

How to run the conversation in your school

Take the phrase off the table first. Ask the staff meeting a narrower question: which parts of your week are cognitively expensive, repetitive and currently done badly because of volume? The answers cluster tightly, and almost all of them are on the assistant side of the line.

1. List the tasks, not the tools. Twenty minutes with sticky notes produces a better requirements document than any vendor demo.
2. Sort each into decision or input. Anything that is a decision stays with a person, permanently, and that is written into the policy.
3. Pick the two most expensive inputs. Usually "work out why each student got it wrong" and "produce something defensible for the progress meeting".
4. Evaluate tools only against those two. Not against a feature matrix; against the artifact you actually need on Monday.
5. Check the evidence trail. Whatever the tool claims about a student, ask to see what it is basing that on.
6. Check the data posture before the pilot. Whether student work trains a model is a procurement question, not an IT detail, and it is the one that kills pilots late. Ours is on the student privacy page.

The version to say out loud

"We are not replacing teachers. We are giving every teacher the thing that only tutored students have ever had: someone who works out why this particular child got it wrong, and tells the teacher in time to do something about it tomorrow."

That sentence is defensible in a staff meeting, at a board presentation and in front of a parent. "AI teacher" is not defensible anywhere, including on the merits. If you want to see the assistant-side artifact rather than read about it, the Mastery Check demo produces one from a piece of student work in a few seconds, and you can compare the hint mode against the worked-explanation mode in the same panel.

A worked example: the same student answer, both sides of the line

The split is easiest to see when the two sides are handed the identical input. Take one grade 8 answer to a percent-change question: "the shirt is $40 and it goes up 20% so thats $48 then down 20% so the 20s cancel out and its back to $40."

What the "AI teacher" framing produces

Asked to teach, a general chat product does what teaching looks like from the outside: it explains. It tells the student the answer is $38.40, walks through why the second percentage is taken of $48, and offers a similar problem. It is clear, correct, and patient in a way a tired human at 4pm frequently is not.

It also ends the diagnostic. The student now has a correct method and no memory of having held a wrong one; the teacher has no record that nineteen students in the room shared the same fault; and next week the same error comes back because nothing was reteaching, only re-explained. The expensive part of teaching is not the explanation. It is knowing which explanation this particular child needs, and that is precisely what got thrown away.

What the assistant framing produces

Given the same answer and told to assist rather than teach, the output is a different kind of object entirely. The next move to the student is a question: "twenty percent of $40 is $8; what is twenty percent of $48, and why is that a different number?" It contains no corrected value and forces the student back into the reasoning.

The output to the teacher is the part that does not exist on the other side of the line: a named misconception ("applies both percentages to the same base"), four skill rows with states and an evidence line quoting the student's own words for each, and one five-minute activity for tomorrow. Nothing in that record is a decision. It is material for a decision, which the teacher then makes.

Why the difference is structural, not a setting

You cannot get the second output by prompting the first product more carefully. The chat product has no teacher as a party to the conversation, so there is nobody for a record to be addressed to and no reason to produce one. It is built with the student as the customer, and it serves them well by finishing the problem. The assistant is built with the teacher as the customer, which is why the artifact it produces is a record rather than an answer. That is the whole line, and it shows up in the output long before it shows up in the marketing.

Where to look next, by subject and by review

Math is where the assistant/teacher distinction is sharpest, because the misconception vocabulary has to be specific enough to act on. "Struggled with percentages" is a summary; "chose the wrong base for the second percentage" is a lesson. The AI math tutor page sets out how the skill names stay mathematical rather than generic, and which fault lines the diagnosis is built to separate across fractions, ratio, algebra and measure.

The other question that decides this in a real school is not pedagogical. Somebody in the review will ask where minors' work goes, whether it trains a model, and who controls the record. Our answers, including the formal attestations we do not hold, are on the student data privacy and COPPA compliance page.