AI That Detects Scams Without Sending Your Messages to the Cloud
Imagine receiving a message from someone claiming to be a bank employee. The conversation sounds professional, the person knows how to build trust, and eventually they ask you to transfer money to a supposedly safe account.
Traditional spam filters may not immediately recognize the danger. The suspicious activity becomes clear only after several messages reveal the sender's real intentions.
Google DeepMind is working on a new approach to this problem: using artificial intelligence directly on a user's device to identify potentially fraudulent conversations while protecting personal privacy.
On October 14, 2026, Google DeepMind announced a collaboration with Singapore's Government Technology Agency (GovTech Singapore) to develop a privacy-preserving AI model for scam detection. The project focuses on recognizing local scam tactics and warning users before they share money or sensitive information.
According to Google DeepMind's official announcement, the collaboration aims to make scam detection more effective without requiring private conversations to be sent to Google or the Singapore government for analysis.
How Does On-Device AI Scam Detection Work?
Many online services analyze suspicious activity using cloud-based systems. While these systems can be powerful, sending private conversations to external servers raises understandable privacy concerns.
On-device AI takes a different approach. Instead of sending conversation data to a remote server for analysis, the AI model runs directly on a compatible device.
Google DeepMind and GovTech have fine-tuned a lightweight version of Google's Gemma AI model using examples of scam conversations relevant to Singapore. These include impersonation attempts involving public institutions, banks, local figures, and other region-specific fraud schemes.
The model is designed to recognize manipulative patterns across conversations rather than simply searching for suspicious keywords.
For example, a conversation may begin with an ordinary greeting before gradually introducing urgency, financial pressure, or requests for confidential information. Examining the broader context can help an AI system identify warning signs that a basic spam filter might miss.
A Two-Stage System for Faster Detection
Running a capable AI model continuously on a smartphone can consume processing power and battery life. To address this challenge, the project uses a two-stage detection approach.
Stage 1: Lightweight screening
A smaller, efficient classifier quickly checks incoming messages for potential warning signs. Messages that appear harmless can pass through without requiring deeper analysis.
Stage 2: AI conversation analysis
When the initial screening identifies something suspicious, the fine-tuned Gemma model examines a broader portion of the conversation to assess whether it shows signs of manipulation.
If the system detects a likely scam, it can trigger a warning directly on the device, encouraging the user to pause before transferring money or sharing sensitive information.
This layered approach is intended to balance detection quality, speed, and battery efficiency. It also helps avoid running the more computationally demanding model on every incoming message.
Why Privacy Is a Major Part of This Development
Scam protection becomes less useful if users must sacrifice their privacy to receive it. Personal messages can contain financial details, family conversations, verification codes, and other sensitive information.
The on-device approach is designed to reduce that trade-off.
According to Google DeepMind, the model processes conversations locally. When deployed in applications, it will access on-device conversations only with the user's consent and in accordance with the application's terms of use.
Because the model runs locally, Google and the Singapore government will not receive users' private conversations through this deployment.
The development process also includes privacy safeguards. Google DeepMind says model training took place within a managed Google Cloud environment controlled by GovTech, using anonymized datasets derived from scam reports and intelligence sources rather than private, personally identifiable citizen data.
These safeguards are important because effective security tools should protect people without unnecessarily collecting their personal information.
What Types of Scams Could It Help Detect?
The model is being developed to recognize sophisticated social engineering, where scammers manipulate people through convincing interactions instead of relying only on suspicious links or obvious spam.
Potential use cases include:
Bank impersonation: Someone pretending to represent a financial institution and requesting a transfer.
Government impersonation: Messages claiming that a recipient must urgently pay a fee or provide confidential information.
Investment fraud: Conversations designed to build trust before persuading victims to invest money.
Fake opportunities: Fraudulent job offers or other schemes that gradually pressure people into paying fees or revealing personal details.
These are target use cases, not a guarantee that every scam in these categories will be detected. Scammers continually change their methods, so the model will need ongoing evaluation and improvement.
When Will Users Be Able to Access This Technology?
The collaboration was announced in Singapore as part of the country's wider efforts to use AI for public benefit and cybersecurity.
However, the announcement describes a technology being developed for phased integration into applications in partnership with the Singapore government. It does not mean the specific model is already available to every smartphone user.
Its initial development is tailored to local scam patterns in Singapore. Expanding the approach to other countries would require adapting it to different languages, fraud tactics, and local contexts.
This distinction matters because scams that are common in one country may look very different elsewhere. A model trained on Singapore-specific examples may need additional training and testing before it can reliably recognize scams in other regions.
The Future of AI-Powered Digital Safety
Google DeepMind's project highlights a growing role for smaller, specialized AI models that operate directly on personal devices.
Instead of relying entirely on large cloud-based systems, developers can use on-device AI for tasks that benefit from quick responses and limited sharing of personal data.
The same principle could support future applications in fraud prevention, suspicious-message detection, and other forms of digital security. Success will depend on how accurately these systems identify threats while avoiding unnecessary warnings and protecting users' information.
For users, the goal is simple: receive a timely warning when a conversation appears dangerous without having to give up control of private messages.
Final Thoughts
Google DeepMind and GovTech Singapore are exploring how AI can help detect increasingly sophisticated scams while keeping personal conversations private.
Their two-stage approach combines fast screening with deeper conversation analysis, aiming to identify manipulative behavior before users expose their money or sensitive information.
The technology is still being prepared for phased deployment, so its effectiveness and availability will become clearer over time. Nevertheless, it demonstrates an important direction for AI security: smarter protection that works closer to the user and treats privacy as a core requirement.
