Safety
How story time stays picture-book safe
Three of the 14 games write with an AI: The Dragon's Library, Story Potion and Rhyme Time. This page follows one sentence through them, lists what kids can and can't see, and ends with what the checks can't promise.
What kids can and can't see
A kid can see
- Their own games, stars, levels and stories.
- The avatar picker on their device, which they tap to find themselves. It only shows the kids that device was set up for: their own family at home, or their own class at school.
- "Sponsored by talisim.com" with Tali on menu screens. Not during games.
A kid can't see
- Anything another kid wrote. Stories are never shared or published.
- A chat, messages or a friends list. None of those exist.
- Another kid's stars, levels or progress.
- Any other family's kids. Names never show up outside the family or class.
Brothers and sisters, or classmates, do see each other's display names on the avatar picker, so a first name or a nickname is plenty. The parent or teacher who manages a profile can read every story that kid wrote, both sides of it, in the dashboard.
The path of one sentence
Both what the kid types and what the AI writes go through the same two layers, one after the other.
- 1
A kid types a sentence
In The Dragon's Library that might be "The dragon found a map under the bridge."
- 2
Layer 1: the word filter
A fixed list of words that are never OK in a kids' story, whatever the context: swearing, slurs, sexual words, self-harm, hard drugs and gore. It also catches those words spelled with symbols, numbers, spaces or stretched-out letters.
- 3
Layer 2: the AI check
A separate AI, used only for checking, reads the sentence along with the last few lines of the story and answers one question: would this fit in a picture book? Context matters here, so a knight fighting a dragon reads differently from a real-world threat. If the check can't run, the sentence waits; it's never let through unchecked.
- 4
The storyteller writes the next line
Only a sentence that passed both layers reaches the AI storyteller.
- 5
Both layers again, on the AI's line
The storyteller is checked exactly like the kid. It gets no free pass for being a machine.
- 6
The kid sees the result
Either the storyteller's sentence, or, if it failed, a gentle fallback line in its place.
If the kid's sentence fails
It's never sent to the storyteller. The kid gets a friendly nudge to take the story in a different direction and tries again. Nothing is lost, and no stars are taken away.
If the check can't run
If the AI check is unavailable, it fails closed: the kid's sentence waits and they're asked to try again in a moment. It isn't passed on unchecked.
If the AI's sentence fails
The kid never sees it. A fallback line takes its place, like:
- “A tiny star peeked out and winked, as if it knew a secret.”
- “A little door in the library glowed with a warm, curious light.”
- “Somewhere nearby, a page turned all by itself with a gentle flutter.”
What "picture-book level" means
Kids' stories need a bit of danger to be any fun, so the AI check isn't tuned to block every dragon. It's tuned to what you'd find in a picture book.
Passes
Fairy-tale adventure: knights on a quest, a dragon guarding its treasure, a spooky castle, getting lost and finding the way home.
Blocked
Gore, sexual content, self-harm and real-world threats, along with swearing and slurs caught by the word filter.
Where story text goes
In the three AI games, what a kid types goes to the safety check and, if it passes, to the AI storyteller. It's used for that and nothing else: we don't use kids' stories to train AI. Stories are stored securely so the parent or teacher can read them later.
The other eleven games don't send anything a kid types to an AI.
What it can't do
These are the gaps worth knowing about before you hand over the keyboard.
-
The AI check makes mistakes, both ways.
It's an AI making a judgement, so it can let through a sentence it should have stopped, and it can block one that was fine. The word filter catches the obvious cases, but something unsuitable written in innocent words could still get past both layers.
-
The word filter is short on purpose.
It matches words, not meaning, so it's kept to words that are never OK in a kids' story. Everything else, including ordinary adventure words like "sword" or "dragon-slayer", is left for the AI check to judge. That puts most of the judgement on the AI layer.
-
Only the AI games are checked this way.
The other games are fixed puzzles and word lists with no AI check on what kids type. In Silly Mad Libs a kid can put a rude word into a blank and see it in their own silly story. No other kid sees it.
-
The sponsor line is a link.
"Sponsored by talisim.com" on menu screens links to talisim.com. A kid who taps it opens another website.
-
It doesn't replace a grown-up.
That's why the dashboard shows every story in full. If you read something you don't like, you'll see exactly what the kid typed and what the storyteller wrote back.