Google DeepMind and Singapore’s Government Technology Agency (GovTech) have delivered an on-device scam-detection model that analyses conversations on the phone itself and warns users in real time — without sending private messages to a cloud server. Why it matters: it is a working example of AI safety protection that does not require giving up the privacy of encrypted chats.
October 14, 2026 — Singapore.
What happened
DeepMind announced the model on October 14 in a post presented at Singapore International Cyber Week, as part of its National AI Partnership with the Singapore Government. The system starts from Google’s lightweight Gemma model, fine-tuned for the kinds of deception used locally. When it spots manipulative patterns in a conversation, it puts a warning on the user’s screen, prompting them to pause before handing over money or sensitive details.
The rollout inside apps will be phased and developed with the Singapore Government. DeepMind says the model will only be able to access on-device conversations with the user’s consent, under the terms of the app it is integrated into, and that neither Google nor the Singapore Government will be able to access users’ conversations through its deployment.
Why scams in Singapore are the test case
According to DeepMind, scams make up more than half of Singapore’s reported crimes and have accounted for more than S$4 billion in losses since 2019. The tactics have also changed: instead of simple phishing links, attackers now run long, convincing conversations that build trust before asking for money, and DeepMind says those conversational scams now account for up to 80% of all losses.
That shift is exactly why server-side scanning struggles. Many messaging apps are end-to-end encrypted, so a cloud service cannot inspect the conversation at all. Moving the analysis onto the device is the workaround the two partners chose.
How the on-device model works
Generic models often miss local deception techniques, so Gemma was fine-tuned on local scam conversations — including impersonations of public institutions, banks and local figures, and trending regional schemes — so it can tell ordinary social chat from genuine manipulation.
To keep battery drain down, the system uses two stages. A lightweight classifier first screens inbound messages using minimal power. Only if something looks risky does the fine-tuned Gemma model examine a larger snapshot of the conversation to confirm a threat or discard a false alarm.
Development followed three governance rules the partners call Sovereign Sandbox, No Citizen Data and Stress-Testing. Fine-tuning happened inside a GovTech-managed cloud environment with training data kept isolated; training used anonymised datasets drawn from scam reports and intelligence feeds, with no private or personally identifiable information; and the model was tested against synthetic dialogues and local material covering diverse scam categories and evolving impersonation tactics.
Why it matters
For readers everywhere — not just in Singapore — this is the practical version of a debate that usually stays abstract. Scam detection normally means someone scanning your messages; encryption normally means nobody can. An on-device model is a third option: the warning happens on your phone, and the conversation never leaves it. DeepMind says it hopes to share more soon and to explore similar deployments in other countries. No performance figures, such as detection or false-positive rates, were published with the announcement, so the real test is still ahead — how often it catches a scam early, and how often it cries wolf on an innocent chat. That, in the “everything AI, tested” spirit, is what will decide whether this becomes a template or a one-off.
Frequently asked questions
Does the model read my messages in the cloud?
No. The analysis runs entirely on the device. DeepMind says neither Google nor the Singapore Government can access users’ conversations through the model’s deployment.
Which model is it based on?
It is a fine-tuned version of Gemma, Google’s lightweight model family, trained on anonymised local scam material from Singapore.
When can I use it?
No consumer launch date was announced. DeepMind says rollout within applications will be carefully phased, in partnership with the Singapore Government, and access will be consent-based under each app’s terms.
Sources: Google DeepMind, Unite.AI, Crypto Briefing