NewsGoogle DeepMind2 min read
Google DeepMind and Singapore Build an On-Phone Scam Detector
Google DeepMind and GovTech Singapore present a Gemma model that spots scams in chats on the phone itself, without sending conversations to the cloud.
Key points
- Google DeepMind and GovTech, the Singapore government’s technology agency, presented it on 14 October.
- It is a lightweight Gemma model, fine-tuned on real local scam conversations, that runs on the phone itself.
- It checks messages in two stages: a quick filter and, if something looks suspicious, a fuller review of the conversation.
- Messaging apps and operating system makers can integrate it; no date has been given for users.
⚠ Pending verification. We have not yet confirmed these details with an official source. Check them before relying on them:
- Which Gemma version it uses and which apps will carry it: the announcement does not say.
- Whether it will reach countries other than Singapore.
Message scams, from the relative in trouble to the fake bank, are no longer spotted just by an odd link: they slip into conversations that look normal. On 14 October Google DeepMind and GovTech, the Singapore government’s technology agency, presented a model that watches for them on the phone itself.
How it works
The system uses a lightweight Gemma model, Google’s family of open models, fine-tuned on real scam conversations from Singapore. It works in two stages:
- A quick filter screens each incoming message.
- If something looks suspicious, a second analysis looks at a larger slice of the conversation to confirm or dismiss the threat.
It all happens on the device: the model is optimised for typical smartphones in Singapore and processes conversations locally. That is the big difference from a cloud filter: your private chats do not have to leave the phone to be analysed.
Why Singapore
| Figure | Value |
|---|---|
| Scams as a share of reported crime in Singapore | More than half |
| Losses since 2019 | More than S$4 billion |
| Share of losses from social engineering | Up to 80% |
The figures are those given in the announcement. Social engineering, persuading victims to make the transfer themselves, is exactly what a link filter misses and a language model can recognise from the tone and the conversation.
When it will arrive
There is no date and no specific app. Google DeepMind says messaging platforms and operating system makers can integrate these models directly. For now it is a project with the Singapore government, presented at Singapore International Cyber Week.
What it means for you
- If you live outside Singapore, nothing changes yet, but it shows a direction: small models that protect you without reading your chats on a server.
- Meanwhile, the rule stays the same: distrust any message that rushes you to pay or share codes, even if it seems to come from someone you know. We cover AI risks in hallucinations and privacy.
- If small models interest you, this week also saw LFM2.5-24B-A2B, another example of AI that works without the cloud.
