Google DeepMind and GovTech Singapore used Singapore International Cyber Week to announce a different approach to scam warnings: analyse the conversation itself, on the phone, rather than scanning links and attachments in the cloud.
Google DeepMind cites more than S$4 billion in scam losses since 2019 and says conversational social engineering accounts for up to 80% of losses. The case for analysing chat context is that many scams unfold over several messages rather than through a single malicious link. Whether that protection justifies granting an app access to conversations depends on its effectiveness, consent controls and independently verifiable privacy safeguards.
How it is supposed to work
DeepMind says generic models can struggle with locally specific deception, which is why it fine-tuned a lightweight Gemma model on Singapore-related scam conversations. A low-power classifier first screens incoming messages; potentially suspicious exchanges are then evaluated by Gemma using a larger conversation snapshot. The second stage can issue a warning or dismiss the initial alert. The company has not published comparative results showing how much local fine-tuning improves performance.
The fine-tuning, the makers say, took place in GovTech's managed Google Cloud environment. Training excluded private or personally identifiable information and used anonymised datasets drawn from scam reports and intelligence feeds; testing used synthetic dialogues and local corpora.
The privacy claim is the product
DeepMind says the model will analyse conversations on-device, access them only with user consent under the host application's terms, and provide neither Google nor the Singapore Government access to those conversations through its deployment. The announcement does not describe an independent privacy audit, publish a third-party evaluation or explain how future app integrations will make these assurances verifiable. The model's local processing does not, by itself, establish what other data a host app might transmit.
The announcement does not explain how consent will be presented or how the model will access conversations across different messaging applications; those details will depend on future integrations.
On-device is an architecture claim, and architecture claims about software on your own phone should eventually come with evidence, not only with assurances from the two organisations that built it.
The numbers that are missing
DeepMind has not published a detection rate, false-positive rate or evaluation results broken down by scam type or language. Those measures matter because missed scams limit protection, while frequent false alarms can reduce trust in warnings. Independent testing will also need to assess performance against new tactics and ordinary conversations that mention money without being fraudulent.
Rollout
DeepMind says deployment will be phased through applications in partnership with the Singapore Government. It has not named the first integrating applications, supported devices, launch date or the detailed consent experience. The announcement therefore describes a developed model and intended deployment architecture, not a generally available scam-detection app.
The Straits Times reports that the project targets Android users, but the primary announcement names no supported devices, so that remains press reporting rather than an announced compatibility specification.
Sources
- [1] Google DeepMind blog (primary), "On-Device AI Against Scams: Google DeepMind & MDDI", 14 October 2026 — the announcement: model design, two-stage pipeline, consent and on-device claims, training governance, and the makers' loss figuresRead source
- [2] Unite.AI, "Google DeepMind and GovTech Unveil On-Device Scam Detection Model", 14 October 2026 — secondary coverage of the announcementRead source
- [3] TokenPost, "Google DeepMind and Singapore Build On-Device Tool to Flag Chat Scams", 14 October 2026 — secondary coverage; notes no release date, rates, devices or integrations were providedRead source
- [4] Crypto Briefing, "Google DeepMind and GovTech Singapore unveil on-device AI tool to fight chat scams", 14 October 2026 — secondary coverage of the SICW announcement; notes no performance figures were sharedRead source