UMI

How Umiko uses AI

Umiko teaches children about AI, so it should be honest about its own. Here is where AI writes the words, where a person does, and what the AI can get wrong.

Umiko is an AI, and says so

Umiko is not a person, and children are told that plainly. In the storybooks a child makes, the child's words and the AI's words are printed in different colours, so anyone can see who wrote what.

Written by a person, not by AI

  • Every lesson, every question in a lesson, and every Pearl Dive test.
  • Everything about right and wrong: the life lessons and the stories about manners and kindness.
  • The daily words, the story serial, the spot-it games, and the facts on Umiko's postcards.
  • The one fixed line a child sees if their words suggest they are upset.

Written by AI, then checked

  • Umiko's answers when a child talks to her.
  • Her replies inside a lesson when a child practises asking a question.
  • Homework help: the method and the next step to try, never just the answer.
  • Small makes: the name of a dish in Recipe Lab, feedback on a sentence in Word time.
  • The stars and the one tip after a practice attempt, marked against a checklist a person wrote.
  • The notes around a parent's four-week plan. The plan itself is built by fixed rules.

Every one of these passes a safety check before it is shown. If the check cannot be completed, nothing is shown.

Which AI

  • Words: Claude, made by Anthropic, writes the answers and helps check that they are safe.
  • Listening and live voice: Sarvam AI, an Indian company, turns a child's speech into words and speaks lines that were not recorded in advance.
  • Recorded voice: Umiko's own lines are recorded ahead of time with ElevenLabs. It is not used live in a child's session.

We chose these with children in mind, and we name each one in the consent a parent reads.

What the AI is told, and is not

  • It is told what the child just typed or said, and what the lesson is about.
  • It is never told the child's name or age.
  • Phone numbers, emails and addresses are removed before anything is sent.
  • A homework photo is sent once, to read the page, and is not kept.

What AI cannot do

  • It can be wrong and still sound sure. Catching that is one of Umiko's lessons.
  • It does not know your child, your family or today's news.
  • It is not a doctor, a counsellor or a teacher. For health, safety or a heavy heart, a child is pointed to a grown-up they trust.
  • It makes no decisions about your child. The most it does is mark a practice attempt against a checklist a person wrote, to give stars and one tip. It never ranks children against each other, and nothing it says goes on any record but your own family page.

Your child's work and AI training

We do not use your child's work to train AI models, and we do not sell it to anyone who would. The optional "Help make Umiko better" box is only about us learning which lessons work, from answers with the name and age removed. See Parental consent.

If the AI gets something wrong

Tell us. We look at what was said and fix the cause. The guardrails are described on the Child safety page.