Seed · October 2026

AI makes you think you’ve learned.
Dr.Turing makes sure you really learn.

Most AI cares whether the answer is right. Dr.Turing cares whether the student understands. Remembering and understanding look exactly the same in the moment; Dr.Turing can tell them apart, and teaches by how much the student truly understands.

First, an experiment: students learned with AI. Did they actually learn?

Problem

AI helped students get the answers right, but didn’t help them learn.

High-school students practised with AI, and their practice scores rose a lot. On an exam where AI wasn’t allowed, they did no better than classmates who never used it, and the group using plain GPT-4 did worse. Yet they thought they had learned.

Bar height: how far above or below the no-AI control group (relative to the control mean).

Even when the AI only gave hints, never answers, practice scores rose a lot and exam scores didn’t. Doing well on the spot isn’t the same as having learned.

Bastani et al., PNAS 122(26), 2025. High-school maths at one school in Turkey; about 1,000 students randomised by class into three groups. Limits: one school, four sessions, exam and practice in the same session, no test of long-term retention. This study shows the problem exists. It doesn’t prove Dr.Turing works; that is what the pilot below is for.

Why doesn’t doing well on the spot count as learning?

Insight

The way of studying that worked best on the spot
scored worst a week later

A classic experiment: university students read the same short passage. One group only reread it; another read it once and then recalled it from memory three times. Tested five minutes later, rereading did better. Tested a week later, the result flipped. And the rereading group was the most confident they would remember.

Rereading only40%
Read, then recall61%

Following this pattern, Dr.Turing splits “knowing it” into two kinds

Surface-level

Right after the explanation, the student seems to get it. But that may only mean they still remember what they just heard. Like the five-minute score in the experiment, it doesn’t show they’ll still understand it after some time.

Truly understood

After some time, without hints, the student can still explain why, and can still use it in a new situation. Like the one-week score in the experiment, this is the understanding that counts.

Roediger & Karpicke, Psychological Science 17(3), 2006, Experiment 2: university students, N=180, 30 per study method and test delay; figures are the proportion of idea units recalled. A third group (read three times, recall once) fell between the two. Predicted recall a week later, on a 7-point scale: rereading only 4.8, read then recall 4.0. Limits: prose passages, lab setting. It shows in-the-moment scores and feelings are unreliable; it doesn’t prove Dr.Turing works. New situations count too: Butler, JEP: LMC 36(5), 2010, where students who tested themselves repeatedly also did better on inference questions in new contexts. Forgetting curve: Ebbinghaus 1885, replicated by Murre & Dros, PLoS ONE 2015; forgetting is fastest in the first day after learning. So Dr.Turing waits at least 24 hours after a student first gets it on the spot, an assumed value set from this curve and not yet calibrated on real data.

If the feeling in the moment can’t be trusted, what does real understanding look like?

Product

Every knowledge point, taught until it’s truly understood

In Dr.Turing, every knowledge point in the course materials is a star: grey before it’s learned, half-lit at the surface level, and fully lit only when truly understood. Below, follow one student learning “Deadlock”. Tap each stage to see the star light up step by step.

DeadlockNot started

Example dialogue, written to the product’s current rules; the star lights the same way as in the real product.

If stage 4 goes wrong, this is the real interface

Product screen (Chinese interface): on the left, a red banner saying “You were very sure, but got it wrong. This is thinking you’ve learned it”, and the tutor re-explaining with a kitchen example; on the right, in the black star map, the Deadlock star is circled in red and marked as thought they knew it, not lit
Left: explanations that read like a textbook

The student is very sure but has it wrong. The tutor doesn’t hand over the answer: it goes back to the kitchen example with a counterexample, starting from “locking”, which the student already knows.

Right: this course’s star map

At the same moment, the “Deadlock” star is circled in red and marked “thought they knew it, not lit yet”: the red circle means the student was very sure but had it wrong.

How this moment is scored: stage 4 of Technology, “Misunderstood”
Real screen recording of the Chinese interface, made on 24 October, before the tutor rules changed. Now, when the student has it wrong, the tutor first offers a counterexample or a guiding question, and next turn asks the student to explain it again. The top bar’s 「看懂」 is now called “surface-level”. The demo student’s earlier learning history is pre-seeded.

From not knowing to truly understanding: who does what at each step?

Technology

Whether the student has learned it is computed by code, not decided by the model

Every step the star above takes runs on the same five-step loop: the student says something, Claude understands it, code scores it, rules pick the action, and Claude speaks again. Claude doesn’t score, and the rules can’t see the scores. Tap each stage to see what happens behind the scenes.

Before classClaude reads the course materials, extracts the knowledge points and how they build on each other, and builds the course’s star map. All students share one map; how far each star is lit for each student is recorded separately by code.

  1. 1Student says

  2. 2Claude understands

  3. 3Code scores

  4. 4Rules pick the action

  5. 5Claude speaks

Mmastery0.6 half-lit
Otruly understood0.7 fully lit
  • Surface 0.2
  • Explains why 0.3
  • New situation 0.3
  • Explained again after full +0.1
All code passes to the rules

The rules never see M or O

Example record · numbers computed from the parameters below, which are not calibrated on real data

The rules can’t see scores

The rules that decide whether to explain, probe, guide or explain fully this turn only know how far along this point is, whether it’s time to come back to it, and which side is still missing. Which point to teach and in how much detail is set by a second group of code rules that do read mastery. Claude only puts it into words and can’t change the action the rules chose.

Rules pass automated tests before launch

The tutor rules are code that never calls a model: the same situation always gets the same action. Before going live they passed 33 step-by-step tests, then 6,000 randomly generated 30-step simulated conversations, checking every turn for “at most one question”, “a student’s question is answered first” and “a full explanation after the second miss”. All passed. These test whether the rules hold, not whether teaching works; what Claude actually says isn’t covered either, and prompts keep it in line.

Every score traces back to the student’s words

Every score traces back to the day, the conversation and the sentence the student said. After each session Claude rereads the whole conversation, corrects any on-the-spot misjudgements, and code recomputes the scores. After a model swap, Claude’s labels need re-checking.

Mastery M uses a classic formula (BKT, Corbett & Anderson, 1994; adjusting the starting point by prerequisites is our own rule) with prior 0.3 (raised to 0.4 when every direct prerequisite has been studied and is half-lit, lowered to 0.2 when one has been studied but isn’t half-lit), guess 0.2, slip 0.1 and learn 0.15. Truly understood, O, is recorded by side: 0.2 for first getting it without hints, 0.3 for still explaining why when a condition changes, 0.3 for still using it in a new situation, and past 0.7 the star is fully lit; explaining it again after some time once fully lit adds 0.1, then 0.05. Both sides must come at least 24 hours after first reaching the surface level; intervals for revisiting a fully lit point are preset by knowledge type (about 3, 7, 14 and 30 days for concepts). All of these are assumed values, not calibrated on real data. The numbers above are worked out step by step from a starting point of 0.4: an example, not a real student’s record.

That covers what’s learned in class. What about what comes up outside it?

Student model

What comes up outside class becomes a star on the student’s map too

The map isn’t split by course: concepts that come up outside class go into the student’s own map too, linked to related points. By design, the longer a student uses it, the better it knows what they know. After each session, Claude also notes the background, goals and preferences the student stated in their own words and saves them to “My profile”, each with one quote and the date; it’s instructed not to record guesses. The tutor uses them for analogies and examples; goals stop being used once they expire. The student can see and edit the profile; a deleted item is no longer used.

My star map: a course star cluster linked by “related to deadlock” to an off-course star, the dining philosophers problem
Details of the off-course star: the evidence section lists the two times the student mentioned it, with dates and their own words

Every star traces back to the day, the conversation and the student’s own wordsWhile working on deadlock, the student asked in passing about the dining philosophers problem from an operating systems course. It became an outside-course star linked to “Deadlock”.

How it checks itself

Before the student speaks, it guesses whether they’ll get it, and afterwards checks how good the guess was. It also notes “which kind of explanation works for this student”, passes that to the tutor only once it proves right, and deletes it if it keeps being wrong. Tap each step to see what it records.

Every time the student speaks without hints, there is something to check. But to tell whether it can guess that the student will still know it later, it has to see, after a delay, whether the student still gets it without hints. Products that only check on the spot never get this data.

The student’s own sense of it is recorded too. After they say “I get it”, whether they really get it next time without hints is recorded; so is how sure they said they were, paired with whether they were right. Very sure but mistaken is called out on the spot (see “Product”); the accumulated records are only stored for now. The tutor doesn’t read them and students don’t see them, until there is real student data to decide how to use them.

Real screen recording of the Chinese interface. The code is live, but only validated on demo data; there is no real student data yet. The “23 October” session is a simulated earlier session.

How much of this is built today?

Progress

Live now. The map for the first 3 lessons matched in full; no real students yet

61/61

Star map auto-generated from the first 3 lessons: all 61 hand-labelled knowledge points covered, no duplicates, in 109 seconds.

How we tested

We uploaded the first 3 lessons of COMP5348, built the map online with a real model, and checked it point by point against a hand-labelled list of 61 knowledge points. We checked only for missed and duplicated points; extra points beyond the key weren’t graded.

34/34

Every source quote extracted from Lesson 1 was found on the matching page of the materials.

How we tested

Code looked up each quote on its source page; we also spot-checked 10 at random by hand, and all were correct.

5/5

Two students with different learning histories ask the same question: the tutor taught differently in 3 ways, and the same 3 differences appeared in all 5 runs.

How we tested

On the online comparison page, with both students’ histories pre-seeded, we ran the same question 5 times and compared how the tutor taught. This shows teaching adapts to the student, not that it teaches better. It was tested on 22 October, before the tutor rules changed. The new rules have only been through automated tests (see “Technology”); how they teach together with the real model hasn’t been re-tested.

So when do the big labs’ study modes check?

Why now

The big labs’ study modes check on the spot, not whether it stuck

Study-mode launch dates · when the check happens

  1. Apr 2025Claude learning modeSocratic questioning; public materials don’t mention a later review
  2. Jul 2025ChatGPT study modeSocratic questioning and on-the-spot quizzes
  3. Aug 2025Gemini Guided LearningStep-by-step explanations, checking understanding along the way
  4. Jun 2026Gemini study notebooksDiagnoses strengths and weak spots; the official announcement doesn’t mention an independent review after a delay
  5. Oct 2026Dr.TuringCounts as truly understood only when, at least 24 hours after first reaching the surface level (an assumed value set from the forgetting curve), the student can still explain why and use it in a new situation without hints

Timing

Students already learn with AI, and there is now evidence that AI creates an illusion of having learned. Models can also, for the first time, build a star map from course materials automatically (measured on COMP5348: 3 lessons in 109 seconds), so mastery learning is no longer limited to hand-built courses.

Stance

We aim for real learning, which means asking students to come back later and see whether they still understand. That conflicts with the in-the-moment satisfaction general assistants aim for.

Compounding

Every time, the student model guesses first and checks afterwards; an explanation style it has noted is used only once it proves right. Knowing whether it can guess “still knows it later” takes seeing the student again after a gap, which products that only check on the spot can’t accumulate. See “Student model”

And what we can’t defend: explanations, map building and spaced review are not moats, and we depend on Claude’s models. Our bet is that “still knows it after some time” becomes the standard for this category, and we aim to build it first.

Put the big labs next to the learning products: who has what?

Competition

Each feature exists somewhere; we haven’t seen them combined

The closest competitor is Google’s Gemini study notebooks. Tap a column to see who has it; tap a row to see what that product has and lacks.

Product

Based on each product’s website and launch posts, as of October 2026. The Gemini row also draws on a third-party review by Educators Technology (2026-10-09). The Dr.Turing row reflects what is already in the code; there is no student usage data yet.

Who do we start with, and how big can it get?

Market

Start with Chinese students abroad, then expand ring by ring

They read English course materials in a second language; Dr.Turing explains in Chinese and keeps quotes in the original English. They are concentrated in a few universities, pay high fees and will pay for grades, and they gather in WeChat groups, on Xiaohongshu and in course study groups. The founder is one of them. Tap each ring to see how it is calculated and which numbers are estimates.

Not to scale: 269M is about 800 times 338k.

Later

Bottom-up estimate. Sources: UNESCO 2024, Australian Department of Education 2025, Open Doors 2025; some counts are second-hand or estimated. Studiosity figure from a UK public tender.

Who pays first, and how much is left per student?

Business

Student subscriptions first; sell to universities once we have evidence

Student plan · per active student per month$14.99/mo

The free tier is limited to one course. Reference: Chegg is about $15.95/mo (second-hand source). Selling to universities is the “Later” ring; see the benchmark above.

Rough cost estimate: everything online runs on Opus today. Assumes 12 study sessions a month and one model call per turn; in practice each turn makes several calls, so cost could double, and free users aren’t counted.

Who is building it, and why us?

Team

The founder is the user:
taking this course in Sydney and building the mastery engine

Founder

Viper Xu (徐奥峰)

Founder / CEO · Master’s student in Computer Science, University of Sydney

“I learned a whole course with AI and felt I understood all of it. When asked to explain it back, I couldn’t say a single sentence.”

That sentence later became a paper and a product. The paper was accepted to the AACL-IJCNLP 2026 main conference with Viper as first author; it asks what evidence AI tutoring can give that students really learned. Viper wrote all of Dr.Turing’s code, and its first course is COMP5348, which Viper is taking now. Viper also leads Vibe Coding Club, a student organisation at the University of Sydney with more than 300 members, over 90% of them international students.

A team of three: Viper leads product and engineering and goes full-time after graduating in December 2027; Shibo Meng (孟士博) has just joined; the third member, a university lecturer, leads education.

1,310

comparison tests: the learning stages and review scheduling ported from open-source DeepTutor, and the BKT mastery formula, each checked input by input against a reference implementation, with error under one in a million

6,000

random simulated conversations check the tutor rules, all passing. Code decides what to do each turn; Claude reads what the student says and explains

3

student-model parts: explanations adjust to the student; guesses made beforehand and checked afterwards only score, they don’t change parameters. Plus “My profile”, and self-ratings that are recorded but not yet used. Live online, no real student data yet

Starting from the pilot, how does this year go?

Next

The next 12 months: evidence of real learning

Months 0–2

A three-group controlled pilot: one-week test results, the share who come back to review, and willingness to pay. Bring model cost down.

Months 3–6

Open to Chinese international students in Sydney, for any course.

Months 6–12

A first university partnership; a second city.

AI makes you think you’ve learned.
Dr.Turing makes sure you really learn.