How TaleTykes uses AI

TaleTykes uses AI models to help write and illustrate books for each child, and to help choose feedback on children's writing. This page says exactly where AI is used and where it isn't, what each model receives, how we check its work, and what choices you have.

In short:

• Children never type to or talk with an AI model. They choose from pictures, tiles and menus.
• No AI vendor receives a child's name or any other personal information about a child.
• Nothing an AI model makes reaches a child until our checks pass.
• No AI decides on its own whether a child is safe. A person does.
• Nothing from TaleTykes is used to train AI models.

Where we use AI

Writing books

A model writes each book's text from settings our system chooses. We use Anthropic's Claude models. The settings include the child's band and reading level, the sounds and words the child is practicing, the content rules for the band, a topic picked from a menu, and topics a grown-up has blocked. The same model also helps check a draft against our content rules, as one of several checks.

Drawing pictures

A model draws each picture from a description our system writes for it. We use Google's Gemini image models. If Gemini is unavailable, we may use xAI's image model instead. Every book made with AI is marked so grown-ups can see it.

Feedback on writing

When a child builds writing from tiles and prompts, a model may help choose which of our pre-written feedback lines fits best. The child only ever sees lines that people at TaleTykes wrote in advance. The model never writes free text to a child.

Where we don't use AI

Reading aloud

When a child taps Listen, the words are spoken by the device's own built-in voice, using the speech feature of the web browser or of iOS. The text is spoken on the device. No AI vendor or speech service receives it.

Scoring read-aloud recordings

If a parent turns on read-aloud recording, a child can record themselves reading a page. The recording is scored on the device for words correct per minute and accuracy. The audio stays on the device, in memory, long enough for one replay, and is never uploaded to us or anyone else. Only the scores are saved, and they expire, as our Retention schedule describes.

Children's choices and conversations

There is no chat, assistant or open text box for children anywhere in TaleTykes. Children don't send anything to an AI model.

Safety decisions

If a child's writing suggests the child may be at risk, a person on our safety team reviews it and decides what to do. An AI model never makes that decision.

Decisions about a child

TaleTykes doesn't use AI, or any automated process, to make decisions that have legal or similarly significant effects on a child, such as grades, placement, eligibility for services, or discipline. Our system estimates which skills a child is practicing so it can choose the next book or lesson. Those estimates are a check-in for grown-ups. Teachers and parents make the decisions.

What each model receives, and what it never receives

Before any prompt leaves our servers, our code checks it for personal information and stops it if anything is found. Names are added to a book after it is made, on our server or on the device, and only if a parent turned on name personalization.

Free text from a grown-up, such as a request for a book about a favorite animal, is checked and reduced to structured choices from our menus before any of it is used. A model never sees the words a grown-up typed.

How a book is made, step by step

1. Our system picks what the child should practice next from their skill record, such as a new sound or a set of heart words.
2. It picks a topic from the child's menu choices, and removes any topic a grown-up blocked.
3. It builds a request from those settings and the content rules for the child's band. The request contains no name and no information that could identify the child or family.
4. Our code checks the request for personal information. If it finds any, the request is stopped and never sent.
5. Claude writes the text. Our checks run on the text, and a draft that fails is rewritten or thrown away.
6. Our system writes a description for each picture. The descriptions are checked, then sent to Gemini, or to xAI if Gemini is unavailable.
7. The pictures are checked, and the book is put together and checked as a whole.
8. If a parent turned on name personalization, the child's first name is added to the finished book on our server or the device. The name never goes to a model.
9. Only then does the book appear on the child's shelf.

Books in the shared library are made the same way, without any child's settings, and contain no information about any child.

Children's privacy and AI

The Children's Online Privacy Protection Act (COPPA) and state student privacy laws limit what may be done with information about children. Our AI design follows those limits:

• we don't disclose a child's personal information to AI vendors, so using AI to make a book doesn't share a child's information with a third party;
• we don't use children's information to build profiles for advertising, or for any purpose other than running TaleTykes for that child;
• for school spaces, we use student information only for the school's educational purposes, as our School terms describe.

Vendors' terms

Our AI vendors operate under terms that bar them from training on what we send and limit retention. We use their business services, not their consumer apps. The vendors we use are listed, with what each receives, on our Subprocessors page. We don't use AI vendors other than those listed there.

No training on your data

We don't use your information, your child's writing, answers, recordings or reading records to train AI models, ours or anyone else's. We don't build our own AI models from user data.

How every book is checked before a child sees it

Every book passes automated checks before it reaches a child's shelf. A book that fails any check is never shown to a child. It is fixed and checked again, or thrown away.

1. Personal information. The request is checked for personal information before it is sent to a model, and the finished book is checked again.
2. Decodability. For early readers, every word is checked against the sounds and heart words the child has been taught, using a pronunciation dictionary. A book with a word the child can't yet sound out is sent back.
3. Readability. Sentence length, word length and vocabulary are checked against the band.
4. Content and safety. The text is checked against the content rules for the band in our Content standards, including blocked topics, and for things books must never contain, such as web addresses, phone numbers and real brands.
5. Pictures. Picture descriptions are checked before drawing, and pictures are checked after drawing.
6. Structure. The book's pages, questions and answers are checked for shape and correctness.

We never stream text from a model straight to a child. A child sees a book only after the whole book has passed.

Human review

People on our review team read books as well as code:

• books that come close to a limit in any check go to a person before a child can see them;
• people read a sample of books that passed, to find problems the checks missed;
• every book a grown-up reports is read by a person;
• every piece of children's writing held for safety is reviewed by a person.

When a person finds a problem, we fix the book and, where we can, the rule or check that let it through.

We plan to let parents approve each new book before their child sees it. Until that is available, you can read any book on your child's shelf from your own account.

Model and version changes

AI vendors release new models and retire old ones. When we change the model or version we use for a task, we test the new one against the same checks before it makes books for children. If we add a new AI vendor, or start sending a vendor a new kind of information, we update this page and our Subprocessors page first, and give schools notice as our School terms require.

The limits of AI

AI models are useful, and they make mistakes. Our checks are designed to stop unsafe or unreadable books, but they can't catch everything.

• A book can be plain, repetitive or odd. Checks catch unsafe books, not dull ones.
• A nonfiction book can get a fact wrong, despite our rules.
• A picture can get a detail wrong, such as the wrong number of legs or a picture that doesn't match the text.
• A word can be harder than it should be, or a question can have more than one fair answer.
• Feedback on writing is chosen from a list and may not fit perfectly.

Please read with your child when you can. If something isn't right, report it.

Skill estimates and reading-speed comparisons are a check-in for grown-ups, shown as ranges. TaleTykes does not diagnose or screen for any condition.

Research comparisons

To learn which settings help children most, we may compare two versions of a lesson or book setting, both of which we would be comfortable using for every child. Nothing in a comparison holds back teaching. A child takes part only if their parent turned on the optional research choice in the Notice for parents, which is off by default, or under a school's agreement with us. The child doesn't see which version they got. Comparisons don't change which AI vendors receive information or what they receive.

Your choices

• Read-aloud recording, putting your child's first name in books and messages to you, and research comparisons are each off until a parent turns them on.
• You can block topics for each child, and books will avoid them.
• You can report any book, and it will be read by a person.
• You can ask us what information we hold about your child, and ask us to delete it, as our Privacy policy explains.

State AI laws

Several states have passed or are considering laws about AI and automated decisions. We have built TaleTykes so that:

• no automated process makes decisions with legal or similarly significant effects on a child or a family;
• it is clear when a book was made with AI;
• people can always ask a person to look at a book or a decision;
• this page and our Subprocessors page say which models we use and what they receive.

If a state law gives you a right about AI or automated decisions, write to privacy@taletykes.com and we will respond as the law requires.

How to report a problem

• Use Report this book on any book's page.
• Email safety@taletykes.com about a book, a picture, or writing feedback.
• Email privacy@taletykes.com about how information is used.

Our Content standards explain what happens after you report and how to appeal.