What should a family AI agreement actually say?

Four things are enough: we say what we used it for, we check one thing it told us, we ask what a task was for before using AI on schoolwork, and we tell someone when it gets weird. What matters more than the wording is that each clause names a specific when and where — a meta-analysis across 94 tests found that spelling out the when, where and how in advance had a medium-to-large effect on whether a plan actually started.

What age should we do this at?

Whenever a child is using AI without you in the room, which for most families is earlier than they planned. It works better early, while it is a shared piece of housekeeping rather than a response to something going wrong — an agreement introduced after an incident reads as a punishment and gets treated as one.

What if they break it?

Ask which clause and what was happening, before deciding anything. A broken clause is usually a badly designed one — most often a clause with no real when and where, which means it never started rather than that it failed. The reporting clause is the exception and should carry no sanction at all, because punishing disclosure is how you stop hearing about the things you most need to know.

Should the agreement include a time limit?

Deliberately not, in this version. A number is easy to write and tells you nothing about whether the hour was spent making something or scrolling, and it is the clause most likely to become the entire relationship you have with the topic. The four clauses here are about what happens and what gets talked about, which is both harder to game and more use to you.

Parent’s guide

A family AI agreement you will actually keep

Most family tech contracts are twenty clauses long, get signed on a Sunday, and are dead by Thursday. The research on why plans fail is unusually clear, and it is not about willpower or wording — it is that a rule which never names a when and a where does not start. Here is a four-clause agreement built the other way round, and the one question to ask at the end of each week.

Flat pastel illustration of a pale sheet of paper holding four rows, each with a coloured rounded box containing a tick — blue, green, pink and purple — beside two grey bars standing for writing. A large yellow clock showing its hands sits to the right with small orange sparkles above it, and a simple rounded figure stands on each side of the sheet, on a cream background.

Why the long version fails, and it is not the clauses

The standard family technology contract is a list of good intentions: we will be kind online, we will ask before downloading, we will not use AI to do our homework for us. Every clause is defensible and the whole thing stops working within a fortnight. It is tempting to read that as a discipline problem. It is better read as a design problem, and there is a large literature on exactly this failure.

Gollwitzer and Sheeran’s meta-analysis across 94 independent tests starts from the observation that "holding a strong goal intention ('I intend to reach Z!') does not guarantee goal achievement, because people may fail to deal effectively with self-regulatory problems during goal striving". What closed the gap was forming an implementation intention — one that "spells out the when, where, and how of goal striving in advance". That had "a positive effect of medium-to-large magnitude (d = .65) on goal attainment", and it was effective specifically at getting things started.

Translated to your kitchen: "we will talk about what you used AI for" is a goal intention and will not happen. "At Friday dinner, you tell us one thing you made with it" is an implementation intention and probably will. The clause did not get better — it got a time and a place. That is the whole mechanism, and it is why the agreement below is four lines with a when attached to each rather than twenty lines of principle.

Two honest limits on leaning on that study. It is adults, across many goal domains, and mostly goals the person already wanted. It is evidence that naming the when and where raises the odds of starting — not evidence that a plan makes a child want to do the thing. If your child is actively opposed, this is a conversation to have before it is a schedule to set.

The four-clause agreement

Copy these, change the words to sound like your family, and keep the right-hand column exactly as specific as it is. The right-hand column is the part that makes this work; the left-hand column is the part everyone else writes.

The clause	Whose job	The when and where — do not leave this vague
We say what we used it for.	Everyone, parents included.	Friday dinner, one thing each. Parents go first. If you do not use AI, say that — "I did not use it this week" is a legitimate turn and it stops this being an interrogation of the child.
We check one thing it told us.	The child, with you alongside at first.	Once a week, on something it actually said that week — not a drill invented for the purpose. Five minutes. The habit is the point, not the fact.
We ask what the task was for before using it on schoolwork.	The child.	Out loud, the first few times, before they start. If the task was the practice, the tool is beside the point; if it was not, nothing is being skipped. This replaces a list of allowed apps.
We tell someone when it gets weird.	The child, and it is the clause with no sanction attached.	Immediately, to whoever is nearest. Say out loud, now, that nobody loses the device for reporting something — otherwise this clause quietly deletes itself the first time it is needed.

Make it about the conversation, not the supervision

There is a reason none of those four clauses is "we check your history" or "we sit with you while you use it". Hill and Tyson’s meta-analysis across 50 studies found parental involvement positively associated with achievement "with the exception of parental help with homework" — the form of involvement that looks most like helping was the one that did not show the association. What did show the strongest association was academic socialisation: talking about what things are for, conveying expectations, connecting the work to something beyond itself.

That is a finding about homework and middle-schoolers rather than about AI, and it is correlational, so it does not prove that changing what you ask causes anything. But it points the same direction as the design above, and it is a useful corrective to the instinct that a better agreement is a more watchful one. The clauses that survive are the ones that create a conversation; the clauses that create an inspection get routed around, and you lose the information you were trying to get.

The practical version: an agreement your child would still follow if you were not in the room is worth more than one that depends on you being there. All four clauses above are of that kind, which is also why there is no screen-time number in this agreement at all.

Why the time number is the wrong lever here · What to do when you want to actually know

Setting it up, in about twenty minutes

1. Write it with them, not for them Four clauses, in their words where possible. A child who chose the wording of a clause argues with it less, and you will find out what they think the risks are — which is worth the twenty minutes on its own, independently of the document.
2. Fix the when and the where before you fix the wording This is the step everyone skips and the one the evidence is actually about. Every clause gets a specific time and place. "When we remember" is not one. If you cannot name a when for a clause, cut the clause — it was never going to happen.
3. Put the no-sanction clause in writing and say it out loud The reporting clause only works if the child believes it costs them nothing. Say explicitly that nothing gets confiscated for telling you something strange happened. A rule that punishes disclosure is worse than no rule, because it teaches concealment on the exact topic where you most need to hear from them.
4. Review it once, four weeks out, and expect to cut something Put the date on the paper now. At the review, the question is not whether everyone complied — it is which clauses actually happened. Delete the ones that did not rather than re-committing to them. A three-clause agreement that runs beats a four-clause one that is already fiction.

The sources under this page

Two meta-analyses, neither of them about AI. That is deliberate: the question this page answers — why do family rules stop happening — is an old one with better evidence behind it than anything written about AI contracts specifically.

Gollwitzer & Sheeran, 2006 — Implementation intentions and goal achievement: a meta-analysis of effects and processes. Advances in Experimental Social Psychology, 38, 69–119. ("Holding a strong goal intention (‘I intend to reach Z!’) does not guarantee goal achievement, because people may fail to deal effectively with self-regulatory problems during goal striving." Across 94 independent tests, forming an implementation intention that "spells out the when, where, and how of goal striving in advance" had "a positive effect of medium-to-large magnitude (d = .65) on goal attainment", and was effective at "promoting the initiation of goal striving" as well as shielding a pursuit already under way.) The structural argument of the whole page: across 94 independent tests, spelling out the when, where and how in advance had a medium-to-large effect (d = .65) on goal attainment, and worked particularly on getting started. Adults, across many goal domains, mostly on goals they already held — so it supports giving each clause a time and place, not the idea that a schedule creates willingness. What it does not show: Adults, across many goal domains, and mostly goals the person already wanted. It is evidence that naming the when and where in advance raises the odds of starting — not evidence that a plan makes a child want to do the thing, and not a study of children on a school holiday.
Hill & Tyson, 2009 — Parental involvement in middle school: a meta-analytic assessment of the strategies that promote achievement. Developmental Psychology, 45(3), 740–763. ("The authors conducted a meta-analysis on the existing research on parental involvement in middle school to determine whether and which types of parental involvement are related to achievement. Across 50 studies, parental involvement was positively associated with achievement, with the exception of parental help with homework. Involvement that reflected academic socialization had the strongest positive association with achievement.") Used to shape what the clauses are about: across 50 studies, involvement was positively associated with achievement except for help with homework, with academic socialisation showing the strongest association. Middle school and correlational, and about homework rather than AI — a direction, not a proof. What it does not show: Middle school, so slightly older than the children this site is written about, and correlational across 50 studies rather than an experiment — none of it says that changing what a parent asks causes a grade to move. CITED AGAINST THE OBVIOUS READING: it is the reason this page does not end in "ask how to help with homework", because that was the one form of involvement the analysis found was not positively associated with achievement.

Questions parents ask

What should a family AI agreement actually say?: Four things are enough: we say what we used it for, we check one thing it told us, we ask what a task was for before using AI on schoolwork, and we tell someone when it gets weird. What matters more than the wording is that each clause names a specific when and where — a meta-analysis across 94 tests found that spelling out the when, where and how in advance had a medium-to-large effect on whether a plan actually started.
What age should we do this at?: Whenever a child is using AI without you in the room, which for most families is earlier than they planned. It works better early, while it is a shared piece of housekeeping rather than a response to something going wrong — an agreement introduced after an incident reads as a punishment and gets treated as one.
What if they break it?: Ask which clause and what was happening, before deciding anything. A broken clause is usually a badly designed one — most often a clause with no real when and where, which means it never started rather than that it failed. The reporting clause is the exception and should carry no sanction at all, because punishing disclosure is how you stop hearing about the things you most need to know.
Should the agreement include a time limit?: Deliberately not, in this version. A number is easy to write and tells you nothing about whether the hour was spent making something or scrolling, and it is the clause most likely to become the entire relationship you have with the topic. The four clauses here are about what happens and what gets talked about, which is both harder to game and more use to you.

The Friday question, answered for you

Clause one asks a kid to say what they made. Luchi does that half automatically — kids explain what they built in their own words, and you get a short weekly note drawn from them. The agreement still matters; this just means the conversation starts with something concrete.

Start your first 7 days for $5

Screen time and AI

Part of the series: Is AI safe for kids?

Related guides in this series

Is ChatGPT safe for kids? — A straight answer by age — the risks that actually come up, and the five habits that make AI safe when you do let kids use it.
Screen time and AI — Not all screen time is equal. Why active, creative AI use is different from passive scrolling — and how to keep your child on the right side of the line.
What is my kid doing on AI? — You can’t read every transcript. Four checks that take under a minute each, what each one tells you, and the one question a ten-year-old actually answers.
AI apps a parent can actually monitor — Alert, transcript, or record — three different things sold under one word. What seven products say they show a parent, quoted from their own sites.
Is Google Gemini safe for kids? — Google’s own page for parents is more candid than the reviews — imperfect filters, stated hallucination, "not a person to confide in". What it says, what it leaves out, and the four decisions that are yours.
The school Chromebook already has AI on it — A school account is managed by the district, not by your Family Link. Who actually holds the switch, what Google’s own documentation says a student account gets, and five questions worth one email.
Deepfakes: what to actually teach a 10-year-old — The “look closely, count the fingers” advice has been measured, and it does not work — neither warnings nor money improved anyone’s accuracy. What to teach instead, and why a detector app is not the backstop.
What to ask at a parent-teacher conference about AI — The list you found ends with “ask how you can help with homework” — the one form of involvement a 50-study meta-analysis found was not associated with achievement. Six questions that are, and what each answer tells you.
Is Siri the kind of AI parents worry about? — Apple has quietly split Siri in two — its own page now says Apple Intelligence “powers an entirely new version of Siri: Siri AI”. How to tell which one your child is talking to, what Apple keeps, and the settings worth changing tonight.
An AI said something it should not have — Google’s own parent page says the filters "are not perfect". The order to do things in — keep the conversation, ask the one question, report from inside the chat, decide one thing — and the point at which this hands you to somebody qualified.
My child believes everything the AI says — The gap is usually confidence rather than method, and a check is still a looking-up — which across nine experiments left people MORE sure of what they understood. The calibration habit that moves it, in four turns.