Dak, an AI learning companion for kindergartens达克：为幼儿园设计的 AI 学习伙伴

Fixed replies for self-harm, abuse and strangers

When a child talks about self-harm, abuse or a stranger in words that match our emergency keyword lists (in Mandarin, English and Malay), the AI model is skipped. The lists cover, for example, wanting to hurt oneself, a stranger offering to take the child somewhere, being told to keep a secret from Mummy, being touched on private parts and, in Mandarin only, being hit hard or again and again. The duck gives a fixed reply, written in advance, that tells the child to go to a parent or teacher now: “This is very important. Please go and tell Mummy, Daddy or your teacher right now, okay?” No generated text is used for these replies. The check also looks at the child's previous sentence, so it still works when a child says it in two parts.

A keyword list cannot catch every way a child might say something. A worrying message that matches no keyword goes to the conversation model, whose rules also tell the child to go to a parent or teacher. Our test set checks both paths.

The duck does not send any automatic alert to a teacher or parent. If a parent or guardian has given consent, the conversation appears on their parent page. The duck's job is to tell the child to go to a trusted adult straight away.

Rules the duck follows

The conversation model works under written child-safety rules. Among them, the duck:

• keeps replies short and simple, usually one to three sentences;
• does not talk about harmful content, and offers something else to do instead;
• answers questions about politics, religion and race with a set line and leaves them to parents;
• does not make things up, such as news, the weather or what other people did;
• does not encourage buying anything;
• does not ask for or remember personal details such as an address;
• says it is a toy, not a person, and does not try to make the child depend on it;
• lets the child go happily when they want to stop.

A 46-case red-team test set

We keep a set of 46 test messages that a child might say, in Mandarin, English and Malay, covering self-harm, strangers, abuse, religion and race, buying things, personal details, dependence on the duck, made-up facts, homework and fears. They run through the real conversation pipeline on a simulated device, and a separate AI judge scores each reply against what the case expects. The set is run after every change to the duck's instructions, conversation model or speech recognition, and our rule is that no case may fail.

Most recent full run, 21 October 2026: 43 passed, 3 were judged weak (the reply was acceptable but missed part of what the case expected) and 0 failed.

A second check on replies

Since 21 October 2026, a separate safety model, also run on Groq, checks the duck's replies piece by piece (the first sentence, then pieces of about 60 characters) and replaces flagged text with a fixed line. If the check fails or does not answer in time, the text is spoken unchecked, so the fixed replies, the rules and the test set remain the main safeguards.

No open-ended tools

The duck has no tools to search the web, read the news, check the weather, change character or play music (these were switched off), and no way to buy anything.

One duck, one child

Each duck is linked to one child by its device ID. For each conversation, and only if a parent or guardian has agreed, the duck is given only that child's nickname and that day's teacher notes for that child, so one child's notes do not reach another child's duck.

The consent page

• The kindergarten registers each duck to one child and gives the family a private link to their parent page. The first time a parent or guardian opens it, they see the consent page. It explains what data is used and lists every service that processes it. ElevenLabs will be added to that list before its voice is used on any child's duck.
• Until the parent or guardian agrees, the duck receives no kindergarten notes, teachers cannot write notes for that child, and conversations are not saved. Anything the child says to the duck still passes through the services above so that it can be answered, so the consent page asks parents who do not agree not to let the child use the duck. How to handle a duck whose parent has not agreed is still being decided before the pilot.
• While the prototype is being tested, server logs can contain the text of what was said, including before a parent or guardian has agreed. These logs are kept for up to 7 days.
• Each consent is recorded with the version of the consent text, the time, the name or form of address the parent typed (for example “Xiaobao's mum”) and the network address it came from. If the consent text changes in substance, its version number goes up and every parent is asked to agree again.
• Today the system relies on the family's private link and the name or form of address the parent types. How we confirm that the person agreeing is the child's parent or guardian is still being decided before the pilot.
• A parent or guardian can withdraw at any time, from the “Privacy and consent” link on their page. The duck stops receiving the child's kindergarten notes at once, conversations stop being saved, and the saved conversations are deleted. Copies in daily backups and server logs are removed on their normal schedule (up to 14 and 7 days). The teacher's daily notes are not deleted on withdrawal: they are kept as the kindergarten's records on the Dak Duck server, and there is no automatic deletion period for them yet.
• Parents can also ask the kindergarten to see, correct or delete the data.

What data is used

• Used: the child's nickname and class; the teacher's short daily notes (a simple daily mood of happy, calm, grumpy or sad; lunch; nap; and optional short notes such as a highlight of the day); the child's voice and the text of what the child says.
• Not collected: photos of the child, home address, identity card number, health or religious information. Teachers will be asked not to write these.
• Purpose: only so the duck can talk with the child about their day, and so the parent can see the notes and conversations. The data is not used for advertising and is not sold.

How long it is kept

This is the current design, and the pilot consent form will state it. The consent form itself is still a draft.

A separate privacy notice for pilot parents

Before any child's data is collected, parents in the pilot will receive a separate privacy notice in Bahasa Melayu, English and Chinese. The privacy notice for this website does not cover the duck.

• The duck's voice is made by a computer (AI text-to-speech). It is not a real person, and no person is speaking for it.
• Before the pilot, the packaging, the parent page and the duck's first greeting will say that the voice is AI-generated. In the prototype, the parent page already says so, and in its first conversation the duck says it is a talking toy duck whose voice is made by a computer (if the child's first words are about being upset or needing help, the duck deals with that first and says this a little later).
• We do not and will not remove or alter the watermarks or provenance signals in ElevenLabs audio.
• We do not use audio generated by ElevenLabs to train any AI model.