ModelsArchitectures & capability
Google DeepMind and GovTech Singapore fine-tune Gemma model to flag scam conversations on users' phones
Google DeepMind and Singapore's GovTech built an on-device scam detector: a lightweight screening classifier passes suspicious threads to a fine-tuned Gemma model trained on anonymized local scam data, aiming to warn users mid-conversation before they transfer money.
From the Google DeepMind article of October 14, 2026 and cited sources: The system works in two stages on the phone. [1] [7]
SOURCES[1] Google DeepMind: on-device AI scam detection (Oct 14, 2026) deepmind.google[7] CXO Digital Pulse: DeepMind and GovTech build on-device AI cxodigitalpulse.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Google DeepMind article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 14, 2026 article published
- Covered
- Oct 14 edition
- Evidence
- 9 sources across 8 sites
- 1Google DeepMind: on-device AI scam detection (Oct 14, 2026)deepmind.google
- 2e27: Singapore wants Google's AI to catch scammerse27.co
- 3Google: New AI-powered scam detection features (Mar 4, 2025)blog.google
- 4MDDI: Google and Singapore expand partnershipmddi.gov.sg
- 5Google DeepMind: Strengthening Singapore's AI future…deepmind.google
- 6Unite.AI: DeepMind and GovTech unveil on-device scam…unite.ai
QuietHow it’s measured
Hacker News · 2 likes · 1 comment
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
The system works in two stages on the phone. A low-power classifier screens incoming messages, and only flagged threads go to the fine-tuned Gemma model, which reads more of the conversation for manipulative intent. Messages stay on the device, and DeepMind says neither Google nor the Singapore government can read them. [1] [7]
Fine-tuning ran inside GovTech's own Google Cloud environment. The training data was anonymized material from scam reports and intelligence feeds, focused on impersonation of public bodies, banks and local figures, plus scams trending in the region. [1] [6]
the design and the data pipeline are described; effectiveness is not.
Read the full assessment
Implication: banks, telcos and app makers may be offered integrations, and the sandbox-trained, on-device pattern could serve as a template for other privacy-sensitive government–lab projects.
Sources are in hand; writing the dossier now.
Executive brief
Google DeepMind and GovTech Singapore have fine-tuned a lightweight Gemma model to spot scam conversations on users' phones. Messages stay on the device, and DeepMind says neither Google nor the Singapore government can read them (DeepMind). The goal is to warn people while a scam is still under way, before they hand over money or credentials. That matters in Singapore because in 81.8% of 2025 scam cases, victims made the transfers themselves (e27). There is no launch date, no list of partner apps and no official accuracy figure.
What changed and event timeline
Google launches on-device scam detection
Google Messages started flagging likely scam texts as they arrive, and Pixel 9+ call alerts ran on Gemini Nano. Both launched in English first, in the US, UK and Canada ().
National AI Partnership signed
Google and Singapore's Ministry of Digital Development and Information (MDDI) signed a memorandum of understanding that expands a 2022 agreement. It covers health, science, education and a secure AI ecosystem (;).
Scam model unveiled at Singapore International Cyber Week (SICW)
DeepMind and GovTech described a two-stage on-device pipeline built around a fine-tuned Gemma model. It was trained on anonymized local scam material (;).
Rollout still open
DeepMind describes deployment as "carefully phased" through partner apps but names none. e27 reports no launch date and Android-only availability (;).
Capabilities and access
- Model: a fine-tuned version of "Google's lightweight Gemma model." No variant, parameter count or release version is named (DeepMind).
- Function: reads multi-turn conversations for manipulative intent and shows a real-time warning.
- Access: through partner apps that have not been named, with user consent. e27 says it is Android-only (e27).
Read the full section
- Model: a fine-tuned version of "Google's lightweight Gemma model." No variant, parameter count or release version is named (DeepMind).
- Function: reads multi-turn conversations for manipulative intent and shows a real-time warning. It was trained on impersonation of public bodies, banks and local figures, plus scams trending in the region (DeepMind).
- Access: through partner apps that have not been named, with user consent. e27 says it is Android-only (e27). No public weights or API have been announced.
Technical analysis for researchers and developers
- Cascade design: a low-power classifier screens incoming messages first. This keeps battery and processing costs down (DeepMind).
- Training setup: fine-tuning ran inside GovTech's own Google Cloud environment, which coverage calls a "sovereign sandbox."
- Evaluation: stress-tested on synthetic dialogues and local scam data.
Read the full section
- Cascade design: a low-power classifier screens incoming messages first. Only threads it flags go to the fine-tuned Gemma model, which looks at a longer stretch of the conversation. This keeps battery and processing costs down (DeepMind).
- Training setup: fine-tuning ran inside GovTech's own Google Cloud environment, which coverage calls a "sovereign sandbox." The data was anonymized material from scam reports and intelligence feeds, with no personal data (Unite.AI).
- Evaluation: stress-tested on synthetic dialogues and local scam data. DeepMind has published no benchmark, held-out test set, model card or error analysis, so outsiders cannot reproduce the results.
- On-device inference means it can work alongside end-to-end encrypted chats (CXO Digital Pulse). Exactly how it gets access to other apps' messages is not documented.
Claims and evidence
- Scale of the problem: scams make up more than half of Singapore's reported crimes, with over S$4B lost since 2019 and social engineering behind up to 80% of losses.
- Contested performance numbers: TechBytes reports 94.6% precision, a false-positive rate under 0.08%, a 100,000-interaction test set, a 4-bit 2B-parameter model, latency under 120 ms and 3.5M devices (TechBytes).
- Privacy: the partners say no server-side scanning takes place. No independent audit has been reported.
Read the full section
- Scale of the problem: scams make up more than half of Singapore's reported crimes, with over S$4B lost since 2019 and social engineering behind up to 80% of losses. These figures come from the partners (DeepMind). Separately, e27 reports S$913.1M in 2025 losses, down 19% year on year (e27).
- Contested performance numbers: TechBytes reports 94.6% precision, a false-positive rate under 0.08%, a 100,000-interaction test set, a 4-bit 2B-parameter model, latency under 120 ms and 3.5M devices (TechBytes). None of these figures appear in DeepMind's post, and e27 says no accuracy or false-positive data was released. Treat them as unverified.
- Privacy: the partners say no server-side scanning takes place. No independent audit has been reported.
Context and prior work
- Google already runs on-device scam detection with Gemini Nano. The Singapore project mainly adds an open-weight Gemma model fine-tuned on local, government-sourced scam data.
- In Singapore, the government's ScamShield app and third-party apps like Whoscall and Truecaller already filter calls and messages (e27).
Read the full section
- Google already runs on-device scam detection with Gemini Nano. It started in March 2025 (Google), and Google Messages now also flags romance-baiting and job-offer scams on flagship phones (8bitToast). The Singapore project mainly adds an open-weight Gemma model fine-tuned on local, government-sourced scam data.
- In Singapore, the government's ScamShield app and third-party apps like Whoscall and Truecaller already filter calls and messages (e27).
Limitations, safety and contested findings
- No accuracy, false-positive or latency figures are published. Third-party numbers conflict with the primary source (see above).
- e27 reports the model is Android-only, which leaves out iPhone users in a market where iPhones are widely used.
- Testing relied partly on synthetic dialogues. There is no published evidence on how well it handles new scam scripts or Singapore's mix of languages.
Read the full section
- No accuracy, false-positive or latency figures are published. Third-party numbers conflict with the primary source (see above).
- e27 reports the model is Android-only, which leaves out iPhone users in a market where iPhones are widely used. It also notes it is unclear how the model reads encrypted apps like WhatsApp and Telegram, and that relying on a US vendor carries geopolitical risk (e27).
- Testing relied partly on synthetic dialogues. There is no published evidence on how well it handles new scam scripts or Singapore's mix of languages.
Business and practitioner implications
- Banks, telcos and app makers: expect to be offered partner integrations. Plan for user-consent flows and for how warnings appear in your product.
- ML teams: the pattern to copy is a cheap screening classifier, then a small fine-tuned model for flagged cases, trained in the government's own cloud on anonymized data.
- Buyers: ask for false-positive rates by language and scam type before deploying. None have been published yet.
Read the full section
- Banks, telcos and app makers: expect to be offered partner integrations. Plan for user-consent flows and for how warnings appear in your product.
- ML teams: the pattern to copy is a cheap screening classifier, then a small fine-tuned model for flagged cases, trained in the government's own cloud on anonymized data. It fits other privacy-sensitive classification tasks.
- Buyers: ask for false-positive rates by language and scam type before deploying. None have been published yet.
- Policy: the project is a template for government–lab partnerships where the data stays under national control.