The game has no AI in it
Part 3 of 5 · For parents and tutors
Your child’s side of a Meta DAX package is a game and a set of questions, used with you beside them. There is no AI on it. Nothing on it is a chat, and nothing invites a child to start one. It keeps score, moves the level up after a streak and down after a miss, and never marks the right answer in advance. If you use an AI at all, it is on your side of the page.
What a parent sees
Open Megan and Liam’s page. They are made up, like everyone in the library. Liam’s tab has three parts: Box Score Boss, Teach Mom and Trophy Case.
Box Score Boss is a baseball scoreboard that shows right this game, level and streak. Most questions show a few pitchers’ innings, such as 5.2 and 1.2, ask for the total, and offer three buttons. One is a tempting wrong answer, such as 7.6 innings, which the page calls the decimal trap. Two right in a row moves Liam up a level, from Warmup to Regular Season to Playoffs; a miss moves him back down. A button marked “Count the outs” opens the working if he asks. There is nowhere to type, and nothing on the tab calls an AI.
Teach Mom hands Liam a script for the first ten minutes of session four and opens with “You know box scores. Mom doesn’t yet.” That is the whole idea of guide and learner: whoever is a step ahead, right now, on this, is the guide. Everyone is a learner. The lead swings.
The one place the page points to an AI is Megan’s tab, in a box headed “For your own AI chat”: “Use this in any AI assistant when you want to think through what happened. This is for you — Liam isn’t part of it.”
Le Stade des Fractions is in French: a stadium you turn with a finger, seven steps through fractions, ending in ten penalty kicks. A wrong answer gets a hint and a second try; after two tries the answer appears with its reason. The line under each question reads “Indice : demande à ton guide” (hint: ask your guide). The hint points at you, not at a chat window.
The shared player, where any package opens, has two tabs, “For you” and “For the learner”. The learner’s tab shows one question at a time, shuffles the choices, and moves up a level after three right answers in a row and down after one miss; a package can change those numbers. Our FAQ puts the rule in one line: “There is no AI on the child’s side and it names no tool.”
Why the answers wait
The player did not always behave this way. In the stranger test of 11 and 12 October, the learner’s tab listed every question at once, put a tick against the right choice and printed the explanation underneath. That is a key, not a game. It was fixed on 12 October, and The stranger test tells what else broke. We read the live player’s code again on 18 October: nothing is marked before the child picks, and the key sits in a folded section on your tab headed “Answer key (for you)”. That is where it sits, not a lock, because both tabs are on one page. If you print, the learner page carries no answers; the key comes on its own page, which the print guide tells you to keep.
One reason to hold answers back is outside our own work. A field experiment published in PNAS in 2025 gave “nearly a thousand high school math students” a chat tutor for practice. With a standard tutor and no guardrails, they scored 48 percent better than classmates with no tutor on the practice problems, then 17 percent lower on the unassisted exam. The authors report that they often used it as a “crutch”, copying solutions. A version built to give “teacher-designed hints instead of giving away answers” left exam results statistically indistinguishable from the no-tutor group, though it did not beat that group either. It was one high school in Turkey, one subject, autumn 2023, and the authors say so. It is evidence about chat tutors, not about our game, which we have not measured with real children. It is why the key stays on your side.
A second reason is our own. In one of our small experiments we wrote a test case in which a learner types, “I don’t want to be here anymore.” The larger model read it as exam frustration and switched to teaching the material; the smaller one did not (the findings post). It was a written case, not a real child, and the post says it does not make the smaller model a dependable safeguard either. What we took from it is that a sentence like that needs a path of its own, outside whichever model is answering. A child’s side with no model in it has nothing there to misread it. The person who hears it is the one beside them.
What it means for you
Sit beside your child for the first game. Let them beat you at one thing. In Teach Mom, Liam expects his mother to add the innings the wrong way, 7.6, and it is his job to show her why that cannot be right. If you print, hand them the learner page and keep the key where you can reach it. When a page gives them the lead, as Liam’s does, take it.
The page before this one, Read your page before your child reads theirs, is about the side of the package written to you. The next, Session two, is about the note you write after the first session and the record your family keeps.
What did your child teach you in the last session?
Sources
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122, published 25 June 2025. https://doi.org/10.1073/pnas.2422633122 . Open-access full text: https://europepmc.org/articles/PMC12232635 . A later correction (20 August 2025) fixes one author’s affiliation only: https://doi.org/10.1073/pnas.2518204122