What an AI Teaching Assistant Can and Cannot Do for Faculty

A lecturer is teaching three sections of the same second-year course, a little under two hundred students between them. The internal exam is nine days away, the question bank has not changed in four years, and last year’s paper is already circulating as a photograph in a student group. She needs fresh items, mapped to the same learning outcomes, at three difficulty levels, by Friday. She also has a laboratory to supervise and a set of project proposals to read.

This is the situation an AI assistant is genuinely good for, and it is the one most conversations about AI in higher education walk straight past on their way to arguing about whether a machine can teach. It cannot. The interesting question is narrower and more useful: which parts of the work surrounding teaching can be compressed, and what becomes of the parts that cannot.

What it does well

First drafts of assessment material

Generating variants of a question, drafting plausible distractors, producing a rubric skeleton, mapping items back to outcomes: this is structured, repetitive work with a clear specification. A draft that needs twenty minutes of editing has still replaced three hours. The honest comparison is not between the model and an ideal exam paper written over a fortnight. It is between the model and reusing last year’s paper, which is what actually happens when the week runs out.

Turning one artefact into several

A set of lecture notes can become a revision sheet, a worked example, a short diagnostic quiz, a slide outline, and a plain-language version for students studying in their second or third language. Each of those has real teaching value and almost none of them get made, because they are translation work rather than thinking work, and translation work is the first thing a busy term deletes.

Reading a cohort faster than a spreadsheet does

After an exam, the interesting fact is rarely the mean. It is that twenty-two students lost the same two marks at the same step, which usually means one explanation in week four did not land. Clustering wrong answers by the misconception behind them, rather than by the score, turns a marksheet into a teaching instruction. That is the whole point of analytics that sit close to the assessment rather than on top of a grade export.

Answering the same question for the fifth time at eleven at night

Faculty attention is not infinitely divisible, and the marginal value of a professor explaining a submission format for the fifth time is close to zero. Routine explanation, revision prompts and worked practice are well suited to a tutor that runs outside office hours, provided the record of what students asked comes back to the teacher. The questions a cohort asks at midnight are a better curriculum diagnostic than most formal feedback forms.

What it cannot do

Know the class

The assistant does not know that this cohort came through two disrupted school years, that the Thursday afternoon lab is the one where attendance collapses, or that the quiet student in the third row is the strongest mathematician in the room and simply will not speak. That knowledge is accumulated by a person standing in front of other people, and it is the input that determines what should be taught next week.

Decide what matters

Every syllabus is longer than every semester. Deciding what to cut when two weeks are lost is a judgment about the discipline, about what these particular students will need in three years, and about what the next course assumes. A model can list options. It has no stake in the outcome and no view of the discipline, which is precisely what the decision requires.

Own the grade

A mark is an institutional claim about a person, and responsibility for it is not delegable. Machine pre-marking is reasonable for structured items and useful as a first pass on written work, where it can flag likely misreadings and inconsistencies for a human to check. What follows has to be a person reading, deciding and signing. An institution that quietly moves that line will find out where it was the first time a grade is formally challenged.

Notice the student who has gone quiet

Analytics will flag a drop in submissions, and that is worth having. What it cannot do is tell the difference between a student who has disengaged, one who is unwell, and one who is dealing with something at home. It also cannot make the approach feel like concern rather than surveillance. The flag is the cheap part; the conversation is the part that works, which is why early-warning systems should be designed to produce a named owner rather than a colour on a chart.

Where faculty rollouts actually fail

Almost never because of model capability. Three patterns account for most of it:

• The mandate without the relief. A tool is added to a workload from which nothing has been removed. Faculty read that correctly as more work and respond accordingly.
• The tool that lives outside the workflow. A separate login, no connection to the course, no link from the system where teaching is already administered. Anything that requires a detour gets used during the pilot and abandoned after it, which is why single sign-on and course-level integration matter more than they sound.
• Unstated rules. If nobody has said whether using the assistant is permitted, whether student work processed through it leaves the institution, or whether anything it records can appear in a performance review, cautious academics will simply not open it. The answer to the last question should be no, and it should be in writing.

A division of labour worth defending

The assistant drafts; the academic decides. It produces candidate questions; the academic chooses which test the thing that matters. It summarizes where a cohort went wrong; the academic decides whether to reteach, move on, or change the assessment. It answers the routine question at midnight; the academic handles the one that reveals a misconception worth a whole lecture.

Framed that way, the technology stops being a threat to expertise and becomes what it actually is: a way of spending less of a scarce, expensive, slow-to-train person on work that does not need them. That framing is also the one that survives contact with a faculty senate, which is not a small consideration. If you are working out where to draw that line for your own institution, start with what the assistant is allowed to do unsupervised and work outwards from there, then look at how it fits the rest of the academic operation.