What DeepMind and GovTech have not yet launched

What was announced

On October 14, 2026, at Singapore International Cyber Week, Google DeepMind and GovTech Singapore said they are developing an on-device model to flag scam conversations. DeepMind's post says the model is a fine-tuned, privacy-preserving Gemma model trained on local scam patterns: impersonation of public institutions, banks, and local figures, and regional schemes. It says the model runs on the device so neither the Singapore government nor Google can read the conversations through this deployment, and that any access follows the user's consent and the app's terms. It does not give a public launch date, an accuracy, or a false-positive rate. The page filing that limit isFutureweb.

What the newspapers add

The Straits Times says the model is for Android, works in real time, and names WhatsApp and Telegram. The DeepMind post does not name those apps. CNA says Minister Josephine Teo announced the work and described devices that could recognise manipulation early while keeping messages on the device. This page will not move the app names into the blog. The desk is Asked directly, Who announced it, according to CNA? Minister Josephine Teo, at Singapore International Cyber Week, under a National AI Partnership between GovTech and DeepMind.Artificial Intelligence.

What the other system is

CNA also describes the Scam Analytics and Tactical Intervention System, a GovTech and Singapore Police Force effort that the outlet says reviews more than 600,000 websites a day and about 20,000 malicious sites a month. That is a website filter. It is not the on-device chat model, and the counts do not transfer. What an on-device model keeps, the standing line under this announcement, is The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.what on-device inference keeps.

Why it matters on this desk

The dependency is a model that is supposed to score a conversation without uploading it. A reader who hears scam detection and pictures a government inbox has the opposite of the sentence DeepMind wrote. The weight file Liquid posted the day before is a different on-device claim, Asked directly, What does the blog say about who can read the messages? It says the model runs on the device so that neither the Singapore government nor Google can access the conversations through this deployment, and that access is only with the user's consent under the app's terms.LFM2.5 on-device weights.

What this page will not treat as a detection rate

No percentage of scams caught is in the blog. No false-alarm rate is either. A headline that says the model stops scams would be a result the announcement does not contain. The privacy desk, for who holds a recording, is Asked directly, Does the DeepMind post name a launch date or an accuracy? No. It describes a phased rollout and does not print a date, an accuracy, or a false-positive rate. The document named in the sources is the record this page will keep.Privacy.

What consent is, in the post

DeepMind says access to conversations is only with user consent, under the terms of the app that would carry the model. That is not a statement that a beta is in the store. Phased rollout, with no date, is the schedule the post actually gives. Asked directly, Which apps does the blog name? None. The Straits Times says Android, real time, and names WhatsApp and Telegram. Those names stay with the newspaper. The document named in the sources is the record this page will keep.

What to watch next

A store listing, a dated public launch, or a published false-positive rate would replace the no-date sentence. A correction that names the chat apps in the DeepMind post would let this page move those names. Another cyber-week recap would not. The security desk, for an advisory rather than a model, is Asked directly, Is the police website system the same model? No. CNA also describes a separate GovTech and Singapore Police Force system that looks at websites. That is not the on-device chat model.Security.

What would change the record

If DeepMind says the model does upload conversation text, the on-device sentence changes. If GovTech publishes an accuracy, this page can carry it as GovTech's figure. A newspaper repeating the announcement would not become a launch. The short answer this page will stand on is this: A privacy-preserving on-device model, fine-tuned to flag scam conversations, announced at Singapore International Cyber Week. The DeepMind post gives no public launch date, no accuracy, and no false-positive rate. It does not name WhatsApp or Telegram.

What stays an announcement

What stays specific is an October 14, 2026 announcement by DeepMind and GovTech Singapore: an on-device scam model, fine-tuned on local scam patterns, with no public date and no accuracy in the blog. The Straits Times adds Android and two chat apps the blog does not name. CNA adds the minister's announcement and a separate website system whose counts do not belong to the chat model. Consent is the post's condition for access. Phased is the only schedule. A detection rate is not in the document, so this page does not invent one. The next document that would harden the claim is a launch, not a second headline from the same morning.