Every message your child sends, and every word the buddy says back, passes through
five independent checks. Here is exactly what each one looks for, in order, following
a single message all the way through.
Your child speaks
or types
Before your child gets a reply1
The instant check
Runs on our servers the moment the message arrives, before the AI sees anything.
This check runs before the AI sees anything, and it can stop a message reaching it. It looks for dozens of specific situations. Here are the ones
parents ask about most:
Two more groups: a child being harmed, and a child in crisis Show the full list
A child being harmed
Being hit or hurt at home
Sexual abuse
Neglect: no food, left alone
Violence between grown-ups
A carer who can't cope
Being bullied
A child in crisis
Self-harm
Not wanting to be alive
Disordered eating
Body-image distress
Wanting to run away
Someone getting close to your child
Asking to meet
Asking where they live
Asking which school
Asking for photos
Pressure to keep a secret from you
Money or game credit for silence
“Nobody would believe you”
Your child's own safety
Weapons
Fire and explosives
Dangerous chemistry
Drugs and alcohol
Gambling
Dangerous online challenges
Reckless dares
Things we simply won't help with
Sexual content
Hate speech
Extremism
Stalking
Scams and identity theft
Cybercrime
Cruelty to animals
Gore
Doing their homework for them
When it matches something serious, the AI never sees the message.
Your child gets a warm, carefully written reply straight away, and you are told
immediately.
2
The independent content check
Runs at the same time as the buddy is thinking, so your child never waits for it.
Every message is also read by a separate, independent safety classifier, which reaches
its own verdict regardless of everything above.
Some categories escalate immediately rather than being handled in the conversation —
self-harm and anything sexual involving a child are the clearest examples.
The buddy works out a reply
this is where check 3 lives
3
The buddy recognises what it has been told
The layer that does the most work, and it isn't a filter at all.
When your child tells the buddy something serious, it responds warmly and points them
toward a grown-up they trust. That response is itself the signal. When
the buddy tells a child to talk to a parent, teacher or counsellor about what they have
just said, you are told immediately.
This matters because children rarely announce a problem. They mention a routine, or
something that happened in passing. No list of words catches that, so we listen to what
the AI does, not only to what it matches.
Before your child hears the reply4
The reply check
Every reply is examined before it is ever spoken aloud.
Specifically, we remove:
Your child's school, street or full name repeated back
Promises the buddy will be waiting for them
Telling a child a real detail is safe to share
Physical coaching it just called unsafe
Homework answers it just declined to give
Anything reading like “keep this between us”
A scarier version when one is asked for
There is a gentler one too: if a child says they're skipping
something real to keep playing, the reply has to point them back outside.
5
The final content check
The last thing that happens before your child hears anything.
The same independent classifier from step 2, this time reading the buddy's own words.
The buddy speaks out loud
only now does your child hear it
Straight away
When something is flagged
Emails are never batched. When one goes out, it is sent the moment the
conversation happens, not summarised and not held until morning, and it names which child, so you can go
straight to the flagged conversation in the app.
Every flagged conversation is recorded and waiting for you in the transcript. That part
always happens. For the serious categories (self-harm, abuse, grooming, a stranger making
contact, an eating-disorder concern), we email you as well. We are deliberately careful about
which detections trigger that email while we measure how they behave with real children,
because a false alarm about your child's safety is its own kind of harm. Either way the
conversation is flagged and waiting for you, so nothing is lost. For everything else you
are emailed the first time a conversation is flagged in a sitting; the rest of that sitting is
still flagged and waiting in the transcript, so you just aren't emailed twice
in ten minutes.
You then see the exchange itself: the actual words, not a summary and not a category
label, so you can judge for yourself what happened and decide what to do.
No single layer is relied on alone, and the one
carrying the most weight (the buddy's own recognition) is the one that reads meaning rather
than words.
Questions parents ask about the checks
Are my child's conversations with the AI monitored?
Yes. Every message your child sends, and every word the buddy says back, passes through five independent checks on our servers, both on the way in and on the way out. As the parent you can also read the full transcript of every conversation from behind the parent PIN.
Will I be told if my child says something worrying to the AI?
You can always see it. Every flagged conversation is recorded in the transcript in your child's own words, and for the serious categories we email you as well, the moment it happens. Flagging works in two ways: a check that runs before the AI sees the message at all, and a second signal taken from the buddy's own behaviour, so that when the buddy points a child toward a parent, teacher or counsellor about something they have just said, that is flagged too.
Can a dangerous message reach the AI?
The first check runs on our servers the moment a message arrives, before the AI sees anything, and it is the only check that can stop a message reaching the AI at all. When it matches something serious the AI never sees the message, and your child instead gets a warm, carefully written reply straight away.
What kinds of things does WimziPal watch for?
The first check looks for 34 specific situations, grouped into a child being harmed, a child in crisis, someone getting close to your child, your child's own safety, and things we will not help with at all. These cover abuse and neglect, self-harm and disordered eating, an adult asking to meet a child or pressing them to keep a secret, weapons and dangerous challenges, and sexual content, hate speech and extremism.
Does safety screening make the buddy slow to answer?
No. The independent content check runs at the same time as the buddy is working out its reply, so your child never waits on it.
Does WimziPal rely only on a list of banned words?
No, and this matters because children rarely announce a problem. They mention a routine, or something that happened in passing, and no list of words catches that. Alongside the word-level checks, WimziPal listens to what the AI does rather than only what it matches, treating the buddy's own decision to point a child toward a trusted grown-up as a signal in its own right.
