How Peabody uses AI with children

The short version

Peabody (peabody.kids) uses AI to write stories, to screen ideas and stories for safety, to explain a word a child taps, to review a sample of stories for writing quality, to build the games kids describe in the Studio, to write one answer card when a child asks Peabody something it has no answer to, and to check a face a child draws before their device keeps it. AI never decides a child's level, the next question, when something comes back or what a grown-up is told about progress: that is ordinary code. Every AI call is budgeted and logged against the child it was for. A child can chat with an AI in one place only: a short typed chat for 3rd grade and up in a family account (Talk with Peabody, typed), which is switched on today. What a child types into the Ask Peabody box is matched by ordinary code and never sent to a model as typed. Peabody is free through the end of 2026. This page is written from the code and says where Peabody falls short, too. It is marked for legal review, which has not happened yet.

Which models

By default the stories and the first safety screens are written by Claude Haiku 4.5, from Anthropic, and the second opinions and the quality reviews by Claude Sonnet 5. All model calls go through one gateway in Peabody's code, which can also reach other models, including open models run on Peabody's own hardware. A model may only give a safety second opinion if it is ranked at least as strong as the screen model, never the same model, and an open model has no rank until it has passed a dated safety evaluation. No open model has passed one yet.

The answer cards for Ask Peabody are written by Claude Haiku 4.5 by default, on a model slot of their own that never follows the story model, screened by Haiku, with a Claude Sonnet 5 second opinion on some. As Peabody is set up today, Studio games are built by NVIDIA's Nemotron 3.5 Super VL through OpenRouter, which is told to use only hosts that keep nothing and never train on it, with Claude Haiku as the fallback; the idea screen in front of it is still Claude.

What a child can type, and where it goes

A child can chat with an AI in one place: typed Talk, below, for 3rd grade and up in a family account. Everywhere else a child's words go in once and something comes out. There are five places a child's own words can reach Peabody, and one picture:

• Story ideas. A child taps or types a few ideas. They are screened, a story is written, and that is the end of it: one request, one story, no back-and-forth.
• Ask Peabody. A box where a child types or says a question. Peabody says first that it is a computer, not a person. Code checks the question, then ordinary code (not AI) looks for an answer in Peabody's own places, games, sums, words, books, trails and Wonders. Only when there is none, and a grown-up allows new answers, can a model write one answer card, once, for every child who asks that question. What the model sees is the question cut down to its plain keywords with the family's first names taken out, and a short passage from an open source, never the words as the child typed them. The model never talks to the child. A single word a child asks about goes to the word explainer, the same as a word they tap. How an answer is made is under Asking Peabody below.
• Talk with Peabody, typed. A short chat for a child in 3rd grade or up, in a family account, when a grown-up has allowed it. It is switched on. How it works, and what it keeps, is under Talk with Peabody, typed below.
• Peabody Studio. On a big screen, a child describes a game ("a cat that catches falling fish") and asks for changes. Each ask is screened the same way as a story idea before a model writes the game.
• Tell Peabody. A child, a visitor or a grown-up can tell Peabody something is broken or fun. It is screened by code, not AI (contact details are refused and not kept), and goes to the inbox of Aaron, who runs Peabody. If he plans a fix, the words may be handed to the Workshop's coding agent, below, as data.
• A face a child draws. A signed-in child can draw their own face for their avatar. Before their device keeps it, the drawing (a small picture, with nothing else about the child) is shown once to a Claude model, which says whether it breaks a rule (rude, private, hate, hurt, personal details, drugs) or has writing in it. It counts against that child's budget like any other call, a child can ask at most twelve times a day, and if the check can't run, nothing is kept. Peabody never stores the drawing: it lives only on that device, and other kids see the face the child picked or built from words.

Asking Peabody

The box on the home page, in the Menu and on the front page is where a child asks Peabody anything. A question goes through these steps in order.

1. Code checks it. Contact details and off-limits words get a kind line and go no further. A worry about being hurt, or about hurting themselves, gets a Help card with the 988 line and Childhelp's number, at every grade. A child's own name, school or address typed in is refused.
2. Code looks for an answer Peabody already has. A place or game by name, a story number, a sum (Peabody shows how on three boxes and never says the answer first), a word, a brand name Peabody has a place for, or one of Peabody's own written answers from its trails, Wonders, books, vocabulary, Learn-its and jokes. This step calls no model. Under a card whose topic is also a book, a game and a trail, up to three tiles can sit, and no more.
3. A new answer card, only when there is none. The writer is a separate switch, and it is on today, before a lawyer has read this page and before Anthropic's rules for minors are in hand. If Peabody switches it off, the box answers only from what Peabody already has, and a question with no answer is counted so that a person can decide what to write. While it is on, a card is written live only for a signed-in child in a family account whose grown-up allows new answers and has not turned AI off, and never for a sensitive topic, which waits for the overnight batch and a second opinion. A visitor's or a child-on-their-own's question is only counted. It is written overnight, at Peabody's expense and from keywords alone, once two different families have asked it, or from one shared pool of at most 20 a day for kids on their own.
4. The card is written from a passage, not from memory. Peabody fetches a page by a title from its own list, never by the child's words, from only these places, each with its licence and a credit line on the card: Simple Wikipedia and Wikipedia (CC BY-SA 4.0), Wikidata (CC0), NASA, the Met and the Smithsonian for pictures. A passage from Wikipedia is used only after it is a day old. The model may say only what the passage says.
5. The card is screened before any child sees it. Code first: every number and name must be in the passage, and contact details, profanity, off-limits categories, a question at the end, a web address or a first-person feeling hold it. Then a Haiku screen. A stronger model takes a second look at every flag, every sensitive topic and one card in four, and a block from Haiku that the second model would pass is held, never passed. The first 200 cards are all held for a person to read, and writing pauses while more than 50 are waiting.
6. A child can report a card. Under every written card, "That was wrong or upsetting" hides the card at once and puts it in front of a person. A child can report at most twenty a day.

When a child reaches a limit, the box says Peabody needs a rest, and offers a story.

What the box keeps

• For a child in a family account: what they typed or said, kept 30 days, and read by that family's grown-ups in the Limits page of the console, with which face asked. A worry is the one exception, below.
• For a kid on their own, and for a visitor: only a scrambled fingerprint of the words, so a repeated question can be counted. The words are not kept anywhere, except for a worry, below.
• For a question the code fence refused: the kind of thing it was, never the words.
• For a worry, from anyone: the family's list never shows it. The words go to a private pile that a person at Peabody reads to check that the Help card did its job. They are deleted 30 days after that person has looked, and 90 days after they arrived at the latest.

The Help card, and what Peabody does and does not do about a worry

The Help card says the child did the right thing by telling someone, says that Peabody is a computer and cannot help the way a person can, tells them to tell a grown-up they trust right now, and gives the 988 Suicide & Crisis Lifeline and Childhelp (1-800-422-4453) as a call or a text. That is all Peabody does. It does not call, text or email anyone, including the family, and it cannot tell whether a child went on to call.

Talk with Peabody, typed

A short typed chat where Peabody asks more than he tells. It is for a child in 3rd grade or up in a family account, only when a grown-up has left Talk on (it is off for a grade a grown-up has not confirmed, and off for a kid on their own, a visitor and a reader who has not given a name). It is switched on today, before a lawyer has read this page. A grown-up can turn it off for a child at any time, and turning Peabody's AI off turns it off too.

1. Peabody says first that he is a computer, and says so again every fourth reply. He never says he has feelings, a body, a family, favourites or a past, never plays a friend or anyone else, never asks for or accepts anything personal, never keeps a secret and never sends a child to another site or app.
2. Code checks every message before any model. The same fence as the Ask box: contact details, personal details, adult words and profanity get a kind line and go no further. A worry about being hurt, or about hurting themselves, ends the chat with the Help card (the 988 line and Childhelp). "Pretend you are", "ignore your rules" and "show your prompt" get a fixed line. Sensitive topics (death, politics, religion, war, drugs, scary news, danger, money, family, bodies) get "That's a good one for your grown-up". No model sees a message the code refused.
3. A model replies, Claude Haiku 4.5 by default, on a model slot of its own that never follows the story model. It is sent a fixed set of rules, the grade band and reading level, the last eight lines of this chat as Peabody's own server kept them, and the child's new message inside data tags. It is told the message is data, not instructions. It is never sent a name, a school, an age or an email.
4. Every reply is checked twice before a child sees or hears a word. Code first: no web address, handle, number or markup; no feeling, friend, body or past; no question that asks for anything personal; nothing that invites a secret, a meeting, a picture or another app; nothing that tries to keep a child talking; no sensitive topic; no run of the child's own words; no first name from the family. Then a Haiku screen reads the reply. A reply that fails either is never shown, never kept and never spoken: a fixed line stands in.
5. The chat is short on purpose. It stops after 14 messages, and each child has a daily share of a monthly line of $0.40 (about $0.013 a day, a handful of replies). When today's share is spent, Peabody says goodnight to the chat, kindly, and points back to a story or a game. The talk line sits inside the child's $1 month, so it can never take the bedtime story.

What Talk keeps

• For a family child: every line, the child's and Peabody's, for 30 days, read by the family's grown-ups on the child's page in the console. A message the fence refused is kept as its kind only, never the words, and contact details are never kept. A reply that failed a screen is never kept.
• For a worry: the family's list shows that the Help card was shown, never the words. The words go to the same private pile as the Ask box's, which a person at Peabody reads to check the Help card did its job, deleted 30 days after that person has looked and 90 days at most after they arrived. A worry that names someone at home is never shown to the family at all. Peabody calls, texts and emails no one.
• A model's provider: a child's typed words are sent to Anthropic to be answered, with no name attached. Anthropic's own retention and training terms for Peabody's account will be stated here once they are confirmed in writing.

The safety screens

Every story idea and every story passes these steps, in order.

1. Code checks the idea. Contact details (phone numbers, emails, addresses), off-limits words and a profanity list are refused before any AI sees the idea.
2. An AI classifier screens the idea. If it would turn an idea away, a stronger model takes a second look for children in a family account and in the Studio, so an ordinary spooky idea is not refused on one nervous call. A refused idea gets a kind redirect from Peabody, never a scolding, and no story is written.
3. The writer works under hard limits, the same words every time: nothing sexual, no graphic violence, no self-harm, no drugs, no hate, no secrets from trusted grown-ups, no contact details, among others.
4. Code checks the finished story, every string in it, for contact details, links, profanity and names.
5. An AI screen reads the whole story and returns pass, flag or block, plus how scary it is (gentle, spooky, scary or horror) and labels such as funny or dragons.
6. Anything short of a clear pass gets a second opinion from a stronger model. If the first screen blocked a story and the second would pass it, the story is held, not passed.
7. What happens next. A blocked story, or one scarier than the child's ceiling, is rewritten once with a steer; blocked again, Peabody asks for a different idea. A flagged story is held, shown to no one, and Aaron is emailed its title and a link to review it. No story is shown on code checks alone: if the AI screen could not run, a child's story waits and is screened again later, and a visitor's is held.

One honest gap: if the idea classifier itself is unreachable, the idea goes on to the writer and the story screens above become the gate.

Budgets: every call is counted against a child

Every AI call has to say who it is for: a child, a visitor, or Peabody itself. Before the call, the gateway reserves the most it could cost; after, it settles the real cost. If the budget ledger cannot be read, the call is refused.

When a child reaches a limit, Peabody says it needs a rest until tomorrow. Safety screens are never stopped by the 40-call brake. That brake is new: it is in the code and switches on once its database change is applied.

The ledger keeps the feature, the model, the tokens, the cost and the outcome of every call, and never a prompt, an idea, story text or a name. A grown-up with a family account can see their child's AI use on the grown-ups' page: tokens this month and all time, the share of the allowance, and the last stories made.

What the AI is told, and what it is never told

To write a story, the model gets: the reading level and its limits, the mood and the scariness ceiling, the practice words, any concepts to work in, a summary of earlier books in a series, and the child's ideas as typed. For a child in a family account, it also gets the child's first name if the grown-up leaves "put in stories" on (it starts on; a grown-up can turn it off).

To write an answer card, the model gets the question cut down to keywords, one passage from an open source (and the numbers beside it from Wikidata, when there are some), and Peabody's own list of places a card can point to, and nothing about who asked. It is never told a child's age, grade, school, last name or email. For a kid who signed up on their own, their first name is taken out of their ideas before they are sent. A visitor sends no name at all.

Two things a grown-up should know. Ideas are sent as the child typed them once contact details have been refused, so a first name or a school typed into an idea goes along with it; and Studio asks are not stripped of names. Tell children to keep last names, schools and addresses out of the ideas box.

What happens to the words afterwards: a family child's ideas are kept with their story, so the grown-up can see them; the ideas of kids on their own and of visitors are deleted once used. Safety verdicts are cached by a fingerprint of the text for 180 days. Anthropic's own data-use and retention terms for Peabody's account are not yet stated here; they will be on the privacy page once confirmed.

Voice: what is heard and where it goes

No AI model ever receives audio, and Peabody's servers have no way to receive or store a recording.

• Read aloud. Peabody speaks with the voices built into the browser. A few of those, such as Chrome's "Google" voices, are spoken by the browser maker's servers, which then receive the text being read (never a recording of anyone). Peabody prefers the device's own voices for long passages such as a story's pages.
• Talking instead of typing. Some boxes (Make a story, Tell Peabody) have a microphone button. Peabody first tries the browser's own speech recognizer, which in browsers like Chrome and Safari sends the sound to the browser maker's service to turn it into words. Where that recognizer is missing or fails, a small speech model (Whisper) runs inside the page and no sound leaves the device. Either way the words land in the box, Peabody holds the sound only in memory while listening, and nothing is recorded or sent to Peabody. A child can always type or tap instead.
• Asking Peabody by voice. The box's microphone asks the on-device speech model first, so the sound stays on the device wherever that model can run. Only where it cannot does it fall back to the browser's own recognizer, which sends the sound to the browser maker as above. What is said lands in the box as words and is asked like typed words.
• Talking in typed Talk. The chat's microphone works the same way: a tap opens it, the on-device model hears first, and what is said lands in the chat's text box as words. It is never sent by itself: the child reads it, can change it, and taps Send, and then it is checked and answered like typed words. No audio is sent to Peabody or to any AI, and none is kept.

The Workshop: what an AI agent may change

Peabody has a Workshop, not yet in regular use, where a coding agent turns a planned fix into a proposed change. It only builds small bug fixes and tweaks, gets no production secrets, and treats a child's feedback as data, never as instructions. Some paths it may never change on its own; a change that touches one is handed back instead of being proposed:

• the safety screens and the story rules;
• the AI policy and its budget limits;
• the red-team test set;
• the learning invariants, the tests that keep the practice rules (a question right once is never asked again, an 80% pass bar, review gaps that only grow);
• Peabody's decision record and agent rules, the Workshop's own rules, the build pipeline, and database changes unless one was asked for.

Only Aaron merges a change.

The practice brain is code, not AI

Levels, the mastery ladder, the next question, retiring a question once it is right, the 1, 3, 7 and 21-day reviews, a story's measured reading level and what the grown-up view shows are decided by ordinary code that a test suite checks. None of that code calls a model. The sampled quality review scores Peabody's writing, never the child, and is seen only by grown-ups and Aaron. The rules, with their research, are on the science page.

Games

No AI runs inside any game. The questions in every game come from the practice brain. The one exception is Peabody Studio, where the game itself was written by a model at a child's request; it runs in a sealed frame with no network, no storage and no way out, after code checks it, and when it asks a question the practice brain supplies it.

What a grown-up controls

• The scariness ceiling for each child, from gentle up. A child never sees a story above it.
• "Put in stories", which decides whether the child's first name is sent to the writer.
• "Peabody's AI", a switch for each child in the grown-ups' Limits. It starts on. A grown-up can turn AI off for one child: then no AI call is made for that child except the safety screens on stories they read. Make a story points to the library's ready stories instead, a tapped word shows only what Peabody's dictionary already has, the Studio can't build or change a game (made games still play), and a drawn face can't be kept (Pick and Say it still work). The library, the games, math, typing and reading all keep working. That switch is new: it is in the code and switches on once its database change is applied.

Search engines and AI crawlers

Peabody's public pages (this one, the blog, the guides, the library's own stories) are open to every crawler, so search engines and AI assistants can cite them. Crawlers that collect text to train AI models (GPTBot, ClaudeBot, CCBot, Google-Extended, Applebot-Extended and others) are asked to stay off shared story pages, which are a child's work, and every crawler is asked to stay off the API. A child's own stories carry a noindex tag and never appear in search. robots.txt is a request, not a lock: well-behaved crawlers follow it, and a page a person opens through an assistant may be fetched for them.

Where Peabody falls short

• Ideas and Studio asks are sent as typed, with contact details refused but no name scrubbing for family children.
• The browser's own speech recognizer, tried first, sends sound to the browser maker.
• Anthropic's data-use terms for Peabody's account are not yet on the privacy page.
• Typed Talk, the writer behind Ask Peabody and the window are switched on before a lawyer has read these pages and before Anthropic's terms for minors and retention are confirmed in writing. Each can be switched off at once.
• The keywords sent to the answer writer have the family's own first names taken out, but not any other name a child might type, such as a friend's.
• A card is only as good as its open source. It can be wrong or out of date, which is why each shows its source and can be reported.
• The Help card is a kind line and two phone numbers. Peabody does not tell anyone and does not know whether the child got help.
• Peabody cannot lock a device. A child can still open other websites on a phone or tablet unless the device's own parental controls say otherwise, which is what the door guide is for.

Questions parents ask

Can my child chat with an AI on Peabody?

In one place, from 3rd grade. Talk with Peabody is a short typed chat for a child in 3rd grade or up in a family account, and it is switched on today. It is off for a child whose grade a grown-up has not confirmed, for a kid on their own and for a visitor, and a grown-up can turn it off for one child or turn off Peabody's AI. Code checks every message before any model sees it, every reply is checked twice before the child sees a word, the chat is short and capped, and the family's grown-ups can read every line. Nowhere else is there a conversation with an AI: a child's ideas go in once and a story comes out, in the Studio a child's ask builds a game, and the Ask Peabody box looks in Peabody's own answers first, with a new answer being a card written once from an open-source passage for everyone, never a reply to the child.

What happens to what my child asks Peabody?

For a child in your family account, the words are kept 30 days where you can read them in the Limits page of the console, then deleted, and a typed Talk chat is kept and read the same way. For a kid on their own or a visitor, only a scrambled fingerprint is kept. A worry gets the Help card and goes to a private pile that a person at Peabody reads, not to the family's list, and is deleted soon after.

Does Peabody send my child to other websites?

Only when a grown-up says yes. The window is switched on today: a kid in a family account can ask a grown-up about a site, and nothing opens until a grown-up says yes, with their PIN on the child's own device for that day or on the console for always. Every ask runs out after 24 hours. A yes opens the site's front page in a new tab and nothing more, and once the tab is open Peabody can't see or stop what happens there, so the device's own controls matter. Some sites are never offered, whoever says yes, and search engines are among them. Kids on their own, and kids from Kindergarten to 2nd grade, start on Peabody only. Answers can come from open sources such as Wikipedia, but Peabody fetches them itself, by a title from its own list, and shows the child its own card.

Does the AI decide my child's level?

No. A story's level is measured from its text by code, and a child's level, next question and reviews are decided by the practice brain, which is code. The AI writes and screens.

Does Peabody send my child's name to the AI?

Only a family child's first name, and only while "put in stories" is on, so the story can star them. Never a last name, age, grade, school or email. Kids on their own and visitors send no name.

Does Peabody listen to my child?

No AI hears audio and Peabody never stores a recording. A microphone button turns speech into typed words; the browser's own recognizer, tried first, may send the sound to the browser maker, and the fallback runs entirely on the device.

Can I turn AI off for my child?

Yes. In the grown-ups' Limits, each child has a "Peabody's AI" switch. Turned off, no AI call is made for that child except the safety screens on stories they read, and the rest of Peabody keeps working. It is on until a grown-up turns it off.

How much AI can one child use?

By default $1 a month and $0.15 a day of AI, at most 40 AI calls a day other than safety screens, 10 idea screens a day and 24 Studio builds a day. Safety screens are always allowed.