A message that appears to come from a government official or a bank requests an urgent transfer, and the phone warns you before you hit send. That is the proposal of a new artificial intelligence model that Google DeepMind has developed together with the Government of Singapore to detect scams without accessing the content of conversations.
The project is gaining relevance at a time when digital scams, especially those based on social engineering, have become one of the main security concerns in several countries. The proposal seeks to anticipate fraud without compromising user privacy, a balance that until now was difficult to achieve.
A growing problem in Singapore
The system was presented during the Singapore International Cyber Week (SICW 2026), the result of a collaboration between Google DeepMind and GovTech Singapore, the technology agency of the Asian country's government. For now, it is a joint announcement: integration into applications will be progressive and does not yet have a confirmed date.
According to data from Google DeepMind, scams—which include investment fraud, fake job offers, and impersonation of officials—account for more than half of the crimes reported in Singapore. Since 2019, cumulative losses have exceeded 4 billion Singapore dollars.
The nature of fraud has also changed. Attackers no longer limit themselves to malicious links that are easy to block, but instead resort to long, elaborate conversations that prepare the victim before asking for money. This type of deception, according to the company, accounts for up to 80% of total losses.
Gemma, trained with local scam cases
The model used as a base is Gemma, Google's lightweight artificial intelligence version. Technical teams fine-tuned it with real scam conversations from the Singaporean context: impersonations of public institutions, banking entities, and well-known figures in the country, in addition to the most common fraud schemes in the region.
Google DeepMind explains on its blog that generic models often fail to identify deception tactics adapted to local contexts. For this reason, the specific training seeks to ensure the system distinguishes between an innocuous social conversation and a real attempt at manipulation when operating within everyday applications.
How the system works in practice
An example cited by the company illustrates how it works: a person receives messages from someone claiming to represent an official body, with a cordial tone that gradually increases the pressure to transfer money to a supposedly secure account. The model analyzes the pattern of the conversation and, if it detects signs of manipulation, displays a warning recommending that the user stop before sharing data or funds.
Two-phase analysis to save battery
Running a complex detection system on a phone involves a technical challenge, and battery consumption is one of the critical points noted by the project's authors, JJ Geewax and Miguel de Andrés. To solve this, the system works in two stages.
In the first phase, a lightweight classifier reviews incoming messages with minimal energy expenditure, looking for potential threats. Only when it detects something suspicious does the second phase kick in, in which the fine-tuned Gemma model examines a larger portion of the conversation to confirm whether there is actually an attempt at fraud or if it is a false alarm.
This cascading design means that the heavier model is only activated in the face of real signs of risk, so most everyday conversations do not require that level of analysis.
What data Google can see and what it cannot
Unlike traditional detection systems that scan messages on external servers, this model processes information directly on the user's device. According to Google DeepMind, by operating entirely locally, neither the Government of Singapore nor the company itself can access conversations through this deployment.
Furthermore, the model's access to chats depends on the user's explicit consent, in accordance with the terms of use of each application. Mark Chen, Director of Trust and Safety at GovTech Singapore, summarized the project's philosophy by noting that fighting scams should not mean sacrificing personal privacy.
Google DeepMind presents this architecture as a potential reference model for messaging platforms that seek to protect their users without giving up end-to-end encryption.
Isolated training without personal data
The model's fine-tuning was carried out within a Google Cloud environment managed by GovTech Singapore, which allowed the data used during training to be kept isolated. According to the information released, the process excluded any private or identifiable citizen data.
Instead, anonymized datasets from scam reports and specialized intelligence sources were used. Subsequently, the model was subjected to tests with synthetic dialogues and local corpora to cover different categories of fraud and impersonation tactics.
One aspect that has not yet been clarified is the actual performance of the system once it is in operation. The deployment phase will be carried out in stages together with the Singaporean government, and key data such as the false alarm rate have not yet been published.
Lessons applicable beyond Singapore
Although the pilot test focuses on Singapore, the approach leaves some useful recommendations for any user concerned about these types of fraud.
It is advisable to be wary of sustained pressure in a conversation. The data cited by the company indicate that most financial losses come from persuasive and prolonged interactions, not isolated links. If someone insists on a long chat to get money moved, it is recommended to stop and verify the information through another channel.
It is also important to review the permissions granted to messaging applications. Any such model's access to conversations depends on user consent, according to each app's terms. If an application offers on-device conversation analysis in the future, it is worth checking what data it uses and where it is processed.
Distinguishing between local processing and cloud processing is key to understanding the actual level of privacy offered by each tool. Having the analysis occur on the phone itself is, according to those responsible for the project, what prevents third parties from accessing the content of the messages.
What this announcement could mean for the future
For now, those residing in Singapore will be able to benefit from this protection only when applications integrate the model, and its activation will depend on each user's consent. Outside the country, the situation does not change for the moment, although Google DeepMind has indicated that it is evaluating bringing similar technology to other markets.
If at any point a messaging application incorporates an equivalent system, the user would receive a warning before sending money or sensitive data, without the conversation leaving the phone itself. Until that happens, the security recommendation remains the same: verify any request for money through another channel, no matter who appears to be behind the message.
Frequently Asked Questions
What is the AI model created by Google DeepMind and Singapore?
It is a system based on Gemma, Google's lightweight model, fine-tuned to detect scam conversations and processed entirely on the user's device.
Does the model read users’ messages?
According to Google DeepMind, the analysis is performed locally on the phone, so neither the company nor the Government of Singapore can access the content of the conversations through this system.
When will this technology be available?
The rollout will be gradual across applications and does not yet have a defined date. For now, it is a joint announcement between Google DeepMind and GovTech Singapore.
Why was this project developed in Singapore?
Because scams account for more than half of the crimes reported in the country, with cumulative losses exceeding 4 billion Singapore dollars since 2019.
Does the user need to give permission for the model to analyze their chats?
Yes. The model's access to conversations depends on the user's consent, in accordance with the terms of use of each application.
Will this technology reach other countries besides Singapore?
Google DeepMind has indicated that it is studying the possibility of extending similar technology to other markets, although there are no confirmed plans for now.
