Imagine a message from a recruiter offering a remote job. Over several days, the conversation feels friendly and professional. Then the “recruiter” asks for a bank login to set up payroll. Or a caller claiming to represent your bank sounds calm and knowledgeable, then asks you to read out a one-time code. Scams like these rarely fail a basic spam filter, because they contain no obvious malicious link or suspicious keyword. They succeed by building trust gradually.
Google’s on-device AI scam detection is the company’s answer to this kind of social engineering. Instead of only checking for known bad numbers or links, supported features analyze conversational patterns on the phone itself and warn the user before they share information or send money. This article explains how these features work, which scams they are designed to address, how they handle privacy, which devices and countries support them, and where their limits lie.
Quick Summary: What Google’s AI Scam Detection Does
- What it is: A set of on-device AI features that flag suspicious patterns in supported calls and messages.
- What it looks for: Conversational patterns Google describes as commonly associated with scams, such as pressure to act quickly, requests for sensitive information, and impersonation.
- Where it runs: Call Scam Detection processes audio on the device. Google states that no conversation audio or transcription is stored on the device, sent to Google servers, or retrievable after the call.
- Which models: Gemini Nano powers call Scam Detection on Pixel 9 and later. Earlier supported Pixels use other on-device machine learning models.
- Who can use it: Not every Android phone. Call detection is limited to supported Pixel phones, and availability varies by country and language.
- What it cannot do: It does not guarantee that every scam is caught. Google’s own documentation says it is not 100% accurate.
Google’s separate Singapore project, developed with GovTech and Google DeepMind, uses a fine-tuned Gemma model and a two-stage pipeline. It was announced on October 14, 2026, but no launch date or public performance figures have been given.
What Is Google’s On-Device AI Scam Detection?
On-device AI means that the machine-learning model runs on the phone’s own processor rather than on a remote server. The phone analyzes the information locally and produces a result without sending the underlying content to the cloud.
Google’s features use these local models to look for signals that a conversation may be a scam. When a pattern appears suspicious, the feature warns the user, giving them a chance to hang up, stop replying, or verify the request through a trusted channel. The warning is a prompt to pause, not a verdict.
This distinction matters. A model detecting suspicious behavior is not the same as proving criminal intent. A legitimate bank might ask a customer to call back, and an unusual but genuine job recruiter might use urgent language. Scam detection estimates risk from patterns. It does not know who is on the other end.
Google currently offers several related implementations, and they should not be treated as one system:
- Call Scam Detection in the Phone by Google app, which analyzes supported calls.
- Message Scam Detection in Google Messages and in supported chat notifications.
- The Singapore collaboration, which describes a fine-tuned Gemma model for multi-turn message analysis. It is a research and deployment initiative, not a feature available on phones today.
How Google’s On-Device AI Scam Detection Works
Step 1: Analyze supported calls or messages
Each implementation examines different inputs. Call Scam Detection listens to the conversation during a supported call, processing the audio on the device. Message detection analyzes supported incoming messages, including SMS, MMS, and RCS messages from unknown numbers in Google Messages, as well as new message notifications from supported chat apps.
The Singapore model is designed for conversations, analyzing a sequence of messages rather than a single one. Not every messaging app or call type is covered, and each feature’s scope depends on its documentation.
Step 2: Identify suspicious patterns
The models look for patterns associated with fraud. Google describes call detection as listening for conversation patterns commonly associated with scams. Examples of suspicious signals include:
- Requests for passwords, one-time codes, or banking details.
- Pressure to transfer money or act immediately.
- Impersonation of a bank, government agency, or trusted person.
- Attempts to build rapport before introducing a financial request.
These are examples of the kinds of signals a scam-detection model might weigh. Google has not published a complete list of the signals used by each model.
Step 3: Explain the two-stage architecture
The Singapore collaboration, announced by Google DeepMind and GovTech, uses a two-stage pipeline:
- A lightweight, low-power classifier screens incoming messages for potential threats.
- Messages that look suspicious move to the fine-tuned Gemma model, which evaluates a larger snapshot of the conversation.
- The second stage either confirms a threat or discards a false alarm, producing a scam alert or a benign verdict.
This staged design addresses a practical constraint. Running a large model on every incoming message would consume significant battery and processing power. A fast first filter means the heavier model runs only when a message warrants closer review. Google has not published latency or battery figures for this pipeline.
Step 4: Deliver a warning
Supported features alert the user in different ways. Call Scam Detection, for example, plays an audible alert and a haptic vibration, and displays a “Likely scam” visual warning. Message warnings appear as a notification or in the conversation, with options such as Done, Not a scam, or reporting the message. The exact alert design depends on the product.
Simplified conceptual flow (not a universal architecture):
Incoming call or message → On-device analysis → Suspicious-pattern assessment → Warning or benign result
What Types of Scams Can Google’s AI Detect?
The following categories describe documented use cases where available, along with illustrations of how these scams typically unfold. Google does not guarantee detection for every scam in each category.
Bank and financial institution impersonation
A caller says they work for your bank and reports suspicious activity on your account. They urge you to act quickly, asking you to confirm your identity or read out a code sent to your phone. Call Scam Detection is designed to notice pressure tactics and requests of this kind during a call. It cannot verify who the caller really is, so you should still end the call and contact your bank through its official number.
Fake job offers and recruitment scams
A supposed recruiter contacts you with an attractive offer, builds credibility over several messages, and then asks for personal documents or a payment for “training” or equipment. Message-based detection analyzes conversations rather than single messages, which can help when each message seems harmless on its own. Google’s public descriptions focus on patterns rather than listing job scams specifically, so treat this as an illustrative scenario rather than a guaranteed detection category.
Romance-investment scams and “pig butchering”
In these scams, a fraudster cultivates a relationship over weeks or months, then introduces a supposedly profitable investment platform. A single message may look like friendly conversation. Analyzing the broader exchange is what makes this pattern visible. Google’s documentation does not specifically list romance-investment scams as a covered category, so readers should not assume detection for this type.
Government and trusted-person impersonation
The Singapore project is trained on local scam patterns, including impersonation of public institutions, banks, and local figures, as well as trending regional schemes. These are documented training categories for that collaboration. They do not imply that the same model is available on phones in other countries.
Phishing and requests for sensitive information
Message detection can flag suspicious requests for personal information. This is a different function from malware detection, malicious-link blocking, or transaction verification. A message can be flagged as suspicious without any link being checked, and a safe-looking link does not guarantee a message is genuine.
What it may not reliably detect:
- Scams in languages, apps, or countries that the feature does not support.
- New scam tactics that differ substantially from patterns the models were trained on.
- Scams delivered through channels outside the supported call or message features.
- Scams that never involve an unusual conversation, such as a simple fake payment page reached through a search result.
Google’s Scam Detection Features Compared
| Feature | Main purpose | Important distinction |
| Android call Scam Detection | Identify suspicious speech patterns during supported calls | Uses on-device models; Pixel-only and country-dependent |
| Google Messages and supported message detection | Flag potentially fraudulent messages | Availability and analysis depend on the implementation |
| Android fake-call impersonation protection | Help detect suspected AI-generated impersonation calls under documented conditions | Separate from general conversational Scam Detection |
| Singapore’s Gemma-based collaboration | Analyze message conversations for local scam patterns | Specific collaboration with a two-stage architecture; not yet a public feature |
These systems complement one another but are not interchangeable. The Singapore Gemma model is not deployed universally across Android, and no single phone offers all of these protections in every country.
Read More: OpenAI Breached Using Anthropic AI Tools: What Happened?
How Google Protects Privacy With On-Device AI
Why local processing matters
Most messaging platforms use end-to-end encryption, which means a provider cannot read message content in transit. Traditional spam and fraud filtering often depends on servers scanning messages, which encryption can make impossible. Running detection on the phone lets the analysis take place within the same device that already holds the decrypted conversation, reducing the need to send that content elsewhere.
What Google says about call privacy
According to Google’s Phone app documentation, Scam Detection processing happens on the device. Google states that no conversation audio or transcription is stored on the device, sent to Google servers or anywhere else, or retrievable after the call. The feature is off by default, and when it is on, participants hear a beep at the start of the call and periodically during it.
What happens to flagged messages?
Message detection works differently from call detection. Google states that suspicious message content is analyzed to produce the warning, and users can choose to report a flagged message to Google when that option is available. Reporting is a voluntary step, not an automatic upload. Readers should check the Google Messages settings and help pages on their own device for the exact behavior.
Chrome’s on-device protections work differently again. Google has described a system in which Gemini Nano evaluates page content locally to extract security signals, which are then sent to Safe Browsing for a final determination. This is a web-protection system, not a call or message feature, and it should not be described as fully offline.
What about the Singapore project?
Google DeepMind and GovTech describe several training and testing safeguards. Training relied on anonymized material derived from scam reports and intelligence feeds. Private or personally identifiable citizen data was excluded. The model was stress-tested against synthetic conversations and local language patterns. These statements describe that project only, and they do not apply automatically to other Google features.
Which Phones and Countries Support Google’s Scam Detection?
Availability changes frequently, and the official support pages are the authority. The following reflects the Phone by Google documentation as of October 14, 2026, but the country list has varied between versions of the page, so confirm it on Google’s live support page before relying on it.
Call Scam Detection (Phone app):
- Pixel 6 and later in the United States.
- Pixel 9 and later in Australia, Canada, France, Germany, India, Ireland, Italy, Japan, Mexico, Singapore, Spain, and the United Kingdom.
- Pixel 9a and Pixel 10a in the United States only.
The feature works only with a SIM card associated with a listed country, while the Pixel is in that country. This is call-specific eligibility and does not apply to every Android security feature.
Message Scam Detection:
Google’s September 2026 Pixel Drop extended notification-based chat scam alerts to Pixel 6 and newer phones in the United States, Australia, Canada, France, Germany, India, Japan, Mexico, Singapore, and the United Kingdom, in seven languages: English, Arabic, French, German, Japanese, Portuguese, and Spanish. The Pixel Tablet is not included. The mid-reply Gboard warning is limited to the United States. Do not reuse the call eligibility list for message detection.
Samsung and other Android devices: Google has said it is working to bring scam detection to non-Pixel Android phones, without announcing a timeline. Check the Android Help page for current device and region support.
Singapore Gemma model: Announced October 14, 2026, with no launch date, participating app, or public performance data.
How to Enable and Use Scam Detection
Enable call Scam Detection
- Open the Phone by Google app.
- Open Settings, then select Scam Detection.
- Turn the feature on if your device and region support it.
The feature is off by default. You can turn it off in Settings at any time, or turn it off for a single call from the in-call menu by tapping More, then Scam Detection.
Review suspicious messages
When a message is flagged, check the warning and read the explanation if one is shown. Use the available options to mark the message as Done, mark it as Not a scam if you believe the flag was wrong, or report it. Settings and defaults vary by implementation, so review your Google Messages settings for the available controls.
If the feature is unavailable
Confirm that your phone is a supported Pixel model, that your SIM and location match a listed country, and that your software and Phone or Messages apps are up to date. Third-party apps cannot activate Google’s built-in scam detection on an unsupported phone.
Limitations and Best Practices
No scam detection system is perfect. Warnings can be false positives, flagging a legitimate call or message. They can also be false negatives, missing a real scam. Scammers constantly change their language, so models trained on past patterns may miss new tactics.
A warning shows that a conversation contains patterns associated with scams. It does not prove that the other person is a scammer. Equally, the absence of an alert does not confirm that a call or message is safe.
Practical habits matter more than any single feature:
- Never share one-time passwords or banking credentials with someone who contacted you unexpectedly.
- Verify financial requests through your institution’s official app or a phone number you find independently.
- Do not transfer money because a caller sounds familiar or authoritative.
- Pause and verify any unexpected request for urgent payment.
- Keep your phone’s software and supported apps updated, and enable built-in protections.
Conclusion
Google’s on-device AI scam detection is a useful layer of defense. It can analyze conversational patterns on the phone, which helps with scams that unfold gradually rather than through a single suspicious link. Its privacy design, in which call audio is not stored or sent to Google, is a meaningful advantage for personal conversations.
These features are not interchangeable. Call detection, message detection, impersonation protection, and the Singapore Gemma collaboration differ in capability, eligibility, and deployment status. Check whether your device and country are supported, enable the protections you have, and verify unexpected requests through a channel you trust.
Frequently Asked Questions
What is Google’s on-device AI scam detection?
It is a group of features that run machine-learning models on your phone to spot suspicious patterns in supported calls and messages, then warn you before you act on a risky request.
How does Google’s AI detect scam calls?
During supported calls, Scam Detection analyzes conversation patterns on the device and alerts you with a sound, vibration, and visual warning. Google states that call audio is not stored or sent to its servers.
Can Google’s AI detect fake job offers?
Message analysis can flag suspicious conversations, which may include recruitment scams. Google has not specifically listed job scams as a guaranteed category, so treat detection as possible rather than certain.
Can Google detect romance-investment scams?
Google has not specifically listed romance-investment scams as a covered category. Conversational analysis may help identify such patterns, but detection cannot be assumed for this scam type.
Does Google’s scam detection send private messages to its servers?
For call detection, Google states that audio and transcriptions are not sent to its servers. Message reporting is voluntary, and users can choose to share a flagged message when that option is available.
Which Android phones support Scam Detection?
Call Scam Detection requires a supported Pixel phone, such as Pixel 9 or later in listed countries. Message alerts extend to Pixel 6 and newer in listed regions. Other Android phones are not yet supported.
Does Google use Gemini Nano for scam detection?
On Pixel 9 and later, Google says call Scam Detection is powered by Gemini Nano. Earlier supported Pixels use other on-device machine learning models instead.
Can Google’s AI detect deepfake voices?
Scam Detection analyzes conversation patterns, not voice authenticity. Google’s separate fake-call protection addresses suspected AI-generated impersonation under documented conditions, which should be checked on Google’s official pages. Neither should be treated as universal deepfake detection.
Is Google’s scam detection completely accurate?
No. Google states that Scam Detection is not 100% accurate and that scammers change tactics. Use it as an extra warning, and verify unexpected financial requests independently.
