October 14, 2026
Scam Detection: for now it's a call toggle on Pixel 9, not protection for your messages
Google promises to catch scammers right inside the chat, without sending anything to the cloud. The announcement gives no launch date and no list of apps.
Scams make up more than half of Singapore's recorded crimes, and in 81.8% of cases the victim transfers the money themselves.
On Oct 14, Google DeepMind and GovTech Singapore showed their answer to these numbers: a warning pops up before the person taps "send". A lightweight classifier cheaply scans incoming messages, and only a suspicious conversation goes on to a second stage, where a fine-tuned Gemma reads the thread for signs of manipulation.
Nothing to install yet. The announcement has no launch date, no list of apps, and no model version or size. The companies talk about a "carefully phased" rollout inside apps, and only with user consent. Gemma was fine-tuned in a GovTech-managed environment on anonymized complaints, and this deployment gives neither Google nor the government access to conversations.
What you can turn on today. For calls, there is Scam Detection in the Phone by Google app, under More → Settings → Scam Detection, off by default. It needs a Pixel 6 or newer in the US, or a Pixel 9 or newer in 13 countries, including Singapore. Audio and transcripts stay on the device and are not stored.
For messages, Spam protection already runs in Google Messages. It turns on by itself and runs models on the device, and the switch is under avatar → Messages settings → Protection & Safety → Spam protection.
Before this, Google and Singapore fought scammers at the app-store level: the previous round blocked 3.3 million attempts to install malicious apps and protected more than 770,000 devices. Now the phone checks the meaning of the conversation, not where a file came from.
Developers can build the same setup on open Gemma. Gemma 3 1B in int4 weighs 529 MB with a 2048-token context, the weights are in the litert-community organization on Hugging Face, and a ready-made demo APK is in the google-ai-edge/mediapipe-samples releases. It runs through the LLM Inference API and needs at least 4 GB of RAM.
The speed is enough for live chat: on a Galaxy S24 Ultra the same stack delivers up to 2585 tokens/s on prefill, and a page of text is processed in under a second (Google AI Edge measurement, Mar 12, 2025).
Until the phased rollout begins, there is no way to test the promise on your own phone.