Transformative AI Use
Develop your thinking.
AI should redefine how you think, not substitute it.
Students are using AI to do their coursework. The question is whether they're using it to think harder, or to skip the thinking. TAU gives students a school-controlled AI chat for coursework, and gives teachers a way to coach that process. Imagine being able to identify where a student pushed back, where they took a shortcut, and where the thinking became their own, for every student, automatically. TAU makes your students' thinking visible.
For teachers and schools. Students are enrolled by their teacher.
Beyond the assignment
A submitted assignment tells you what a student produced. It does not tell you how they produced it: whether they interrogated the AI's first answer or accepted it, whether the argument developed across drafts or arrived complete. Two students can submit comparable work having done entirely different thinking.
TAU starts from the assumption that AI was part of the process, and documents that process. The AI-supported drafting process becomes something you can coach and improve, rather than something you have to infer.
The report
After each draft, TAU produces a report on how the student worked. They receive one overall reading of how much of the thinking was theirs, and the four dimensions that reading is made of.
Bicycle maintenance guide
Draft 1You led with your own ideas and used the AI to develop them.
You set the format, the audience and the five systems in your first message, then took the AI through them one at a time, adding a tool list, a lube guide and a maintenance table. The guide at the end is the one you specified at the start.
The one thing you took without checking was the replacement reference — moments after that same source had been caught inventing one.
This is about how the work got made, not how good it is. Your teacher marks the essay.
The agency level names where the session sits among the four. Beside it, the reading itself: what that level means, what this student actually did, and the one moment that cuts against it.
Direction set in 10 of 11 parts. You opened with a brief rather than a question, and twice refused the frame you were handed.
The moments behind this
Let’s break it down by system.
— before the AI proposed anything.
Remove the citation because specific pages should be referenced.
— it offered two
editions, you took neither. What about the pedals you clip your shoes in?
— the one
thing its five-system structure couldn’t place.
Brakes. The AI decided what mattered there and your four questions stayed inside its list — unlike tyres, where you wrote the brief.
11 of 14 changes altered meaning. Three of those changed how the guide is organised, rather than how it reads.
The moments behind this
What about the pedals… what system do they belong to?
split pedal content across two
systems. Keep the info about future shock because this is for me primarily.
held a real
tension. The closing is very specific to me, lets make it more general.
changed who the
document speaks to.
The m-dashes and the British spellings each took several rounds. Those change how the guide looks, not what it says.
About 8 of 18 claims tested. Twice you checked against something outside the conversation, which is the move that matters most.
The moments behind this
British Cycling (2022). Women’s bike fit guide. This doesn’t exist.
The bottle says 30-40ml for a top up and up to 120 for a refill.
Double check the seallant volumes.
When the AI admitted inventing the British Cycling reference it replaced it with Kotler et al. (2023), and you took the replacement without checking it.
About 3 of 12 ideas are yours. The frame is yours; the five-system spine came from the AI and never moved.
The moments behind this
help me create something usable for female riders who rely on partners to know how to care for
their bikes
— the reader, and the reason the guide exists.
Let’s be honest: a lot of us learned to love cycling through someone else’s
passion…
— your prose, dropped in whole.
I wont be carrying a plug kit, I just carry extra tubes.
— your practice changed
what the guide recommends.
The five-system structure, the maintenance tables and the service intervals arrived in that shape and stayed in it. That spine is the one thing you accepted without argument.
The four dimensions
What the four readings measure
The four are not separate measures. Each reads a different corner of the same question — how much of the thinking stayed the student’s. Two of them read the student’s own material: whether they drove the conversation, and whether the thinking in the essay is theirs. Two read the AI’s: whether the student evaluated what it gave them, and what of it survived into the work. They overlap, and are meant to. Each one shares an edge with two others and owns one corner by itself.
Because they overlap, they are never added up. A total would count the shared part four times over, and averaging a strong reading with a weak one returns the one thing the session was not. A student can drive a conversation well and still finish with little of their own thinking in the essay — that gap is the finding, and one number erases it.
No score stands on its own. Each is stated with the count it was derived from and opens onto the student's own turns, including the place where the pattern doesn't hold. That is what makes it discussable: a teacher can check the reading, and disagree with it, against the same evidence the student can see.
The evidence
The whole conversation, not a sample of it
One square is one student turn, set at the agency level it was read at and left in sequence. Where several turns in a row hold the same shape, the stretch is marked and named — an extraction loop, a challenge arc. A single low-agency turn means little; twelve of them consecutively is a working habit, and that is the scale at which it becomes visible.
Nothing is drawn between the turns. There is no trend line and no average, because the reading of a turn is a label, not a quantity. Selecting a turn shows what the student actually wrote there and how it was read, so any claim in the report can be taken back to the sentence it came from — below, turn 30 is open as an example.
Your session, turn by turn
Most of your turns asked for something or edited what came back. 30 of your 44 turns sit in the lower two rows, and 8 push back on what you were told.
Lets add the hint to look for the small Nm on the bolt head after the current sentence about strict torque limits.
Read as refinement. Part of the Extraction → Insight.
The teacher's view
Reports aggregate by class, by draft and by assignment, so what you read before a lesson is the group's pattern rather than thirty separate impressions. Nothing about how you set work or how you mark it changes.
-
You set the assignment
You control the task, the drafts, and the due dates. Assignments are easy to edit and duplicate across classes.
-
Students work in TAU
They research and draft in a school-controlled chat that behaves like the assistant they already use. Nothing to install, it runs in a browser on a computer, tablet, or phone.
-
The reports aggregate
One per student per draft, then rolled up across the class and the assignment. You grade the assignment against your own rubric as before.
Across all your classes and students
Who needs you, before class starts
One screen for every class you teach: who to check in with, whose session the tool could not read, and who has nothing in yet. Sorted by who needs you, not by who is ahead. The three counts answer different questions and deliberately do not add up — an overdue draft and a session too thin to score are gaps in what we can see, not a claim about how a student is thinking.
Inside one class
What the group as a whole could use
Opening a class gives you the same reading one level down, plus the one thing an average hides: whether the number moved because everyone did, or because half the room did twice as much.
What the tool watches
Habits, not scores
Underneath the readings, the tool watches the order of a student's turns — three extractions in a row with nothing in between, a correction taken without testing it, a point pressed twice rather than dropped. A pattern is a shape in a transcript, so it can be pointed at. Every one of them comes with what to try, because a pattern you cannot act on is a label.
Validation Spiral 6 students · English 10 4 · American Literature 2
What this looks like: Taking and confirming, alternating, with nothing breaking the run. A real sequence detector — it reads turn order, which no dimension does.
Extraction Loop 4 students · English 10 3 · Journalism Elective 1
What this looks like: Three or more turns in a row of asking the AI to produce something, with nothing in between. The student is mining for content rather than working on it.
Unquestioned Assertion 3 students · American Literature 3
What this looks like: The AI stated something as established fact and the student took it.
Challenge Arc 4 students · English 10 2 · American Literature 2
What this looks like: Consecutive turns pressing the AI on the same point rather than accepting the first answer.
Rejection → Redirect 2 students · Journalism Elective 2
What this looks like: They rejected what came back and then said what they wanted instead — refusal plus direction, not just refusal.
Across the assignments
Whether the room moved, and which way
Each column counts every student once, at their reading on that assignment. A ribbon is the group of students carried from one assignment to the next, so a ribbon rising is students holding onto more of their own thinking. Nothing is averaged and nothing is scored out of anything — the diagram claims only membership, which is the only claim these readings support.
Where the teaching happens
TAU does not develop critical thinking. Teachers do.
What the tool contributes is evidence about where a group's reasoning is actually breaking down, early enough to act on. If half a class accepted the AI's framing without challenge at draft one, that is a lesson worth planning for Tuesday rather than a pattern you notice in the marking pile three weeks later. The instructional decision stays yours; it is only better informed.
Privacy
Student work is processed by Google's Vertex AI and stored in Google Cloud, in US data centres. Google does not use student work to train models. We sign a data processing agreement with your district before any student data is entered.
Integrity signals are teacher-only, enforced on the server. They never appear in a student's own view, and a school can switch them off entirely.
Common questions
Is this an AI detector?
No. TAU assumes AI was used, and reports on how. When home computers became common, teaching adapted to them rather than trying to keep them out; AI is the same shift. What TAU looks for is whether the student’s thinking and agency survive the process.
How much extra work is this for me?
Setting an assignment takes a few minutes. TAU generates the report on its own, no manual scoring or tagging. Reading it fits into planning your next lesson, it isn't extra work on top of grading.
What does a school have to set up?
Nothing to install. It runs in a browser on a computer, tablet, or phone. An administrator creates teacher accounts; teachers create their classes and enrol students. Students cannot sign themselves up.
Does the AI train on our students' work?
No. TAU runs on Google Vertex AI, which does not use customer data to train models.
Who can see a student's integrity signals?
Only teachers, and a school can switch them off entirely. They never appear in a student's own view.
Does this replace my grading?
No. TAU never sees your rubric and never grades an assignment. It reports on the process; the mark remains yours.