- What it asks
- What it does not prove
Next
Readback · invented message · practice only
Before you tap, read it back.
Paste a suspicious message. Readback shows the exact requests inside it, and what to do next, in English or Arabic. Nothing leaves this browser.
Your text stays in this tab. No account, no upload, no cloud key.
Read one message
The message goes here.
Nothing read yet
Paste a message, or start with a practice one. You will see what it asks of you, and what it cannot prove.
What it asks of you
Local model suggestion
A practice tool. It can miss a scam or flag a genuine message, and it cannot confirm who sent it.
Three messages, three different asks.
Every message here is invented. The readback is produced by the same engine that runs in the checker.
It reads for the ask, not the tone.
First, it splits the claim from the request.
“Your bank account is suspended.”
A claim
Mentioning a bank is not evidence of anything. The line is kept separate and marked unverified.
Then it looks for what the message wants you to do.
“Send the six-digit code to restore access.”
An ask
A request for a code, a password, a card detail or money is the part that can actually cost you something.
Last, it says what it cannot know.
“Is this really my bank?”
Not proven by any message
Readback never checks a sender, a link destination or a website. It reads text. The verification stays with you, through a channel you already trust.
The model, tested.
Sixteen invented messages are held out of training. This is what the classifier gets right, and where the warning rules still speak for themselves.
| Language | Label accuracy | False positives | False negatives | Uncertain |
|---|---|---|---|---|
| English | · | · | · | · |
| العربية | · | · | · | · |
· Reading the held-out set…
Sixteen synthetic examples are a test of the pipeline, not a measure of real-world accuracy.
Questions worth asking.
Does my message leave this browser?
No. Standard Check is plain JavaScript running in the page; there is no server, no account and no API key. AI-assisted Check downloads its model files, then runs the model on this device. In both modes the message text stays in the tab.
Can it tell me who really sent the message?
No. Text cannot verify a sender, and Readback does not try. A bank name, a logo or an official-looking reference is shown as an unverified claim, never as proof of identity.
Why can the result be wrong?
The classifier learns from invented training messages, so it knows their patterns rather than the real world. The warning rules are deliberately narrow and can miss a scam written in an unfamiliar way, or flag a genuine message that asks for money. Treat the result as a reading, not a verdict.
What does AI-assisted add?
A multilingual zero-shot model that suggests a broad topic for the message. It never raises the caution level and never overrides the warning signs. The first run downloads about 268 MB, so standard mode is the default and needs no network at all.
Does it work properly in Arabic?
The interface, the evidence text and the next step are written in Arabic, and the layout mirrors right to left. The rule checks read Arabic as well as English, including messages that mix the two. The training examples are synthetic in both languages.
What should I do when it says high caution?
Do not reply, do not open the link, and do not share a code, a password or a card detail. Find the organisation yourself: the number on your card, or the app you already use. Ask them directly. If money has already gone, contact your bank immediately.
Start with one message.
Paste it below. You will see the asks, the evidence and the next step, before you tap anything.