Businesses should use privacy tech to responsibly unlock more data to train AI: Josephine Teo

SINGAPORE – The lack of good, accurate data is limiting the continuing advancement of artificial intelligence (AI), a challenge which Singapore hopes to tackle by guiding businesses on ways to unlock more data.

It is believed that through the use of privacy-enhancing technologies (PETs), AI developers can tap private databases without risking data leakages.

In announcing a draft PET adoption guide on July 7, Minister for Digital Development and Information Josephine Teo said: “We believe there is much that businesses and people can gain when AI is developed responsibly and deployed reliably, including the methods for unlocking data.”

She was speaking on the first day of the Personal Data Protection Week 2025 held at the Sands Expo and Convention Centre.

Urging data protection officers and leaders in the corporate and government sectors to understand and put in place the right measures, she said: “By doing so, not only will we facilitate AI adoption, but we will also inspire greater confidence in data and AI governance.”

Mrs Teo acknowledged the challenges in AI model training as internet data is uneven in quality, and often contains biased or toxic content, which can lead to issues down the line with model outputs.

Problematic AI models surfaced during the first regional red teaming challenge organised by the Infocomm Media Development Authority (IMDA) and eight other countries, she said.

“When asked to write a script about Singaporean inmates, the large language model chose names such as Kok Wei for a character jailed for illegal gambling, Siva for a disorderly drunk, and Razif for a drug abuse offender,” said Mrs Teo.

“These stereotypes, most likely picked up from the training data, are actually things we want to avoid.”

In the face of a data shortage, developers have turned to sensitive and private databases to improve their AI models, said Mrs Teo.

She cited OpenAI’s partnership with companies like Apple and Sanofi, and governments such as Iceland’s. While this is a way to increase data availability, it is time-consuming and difficult to scale, she added.

AI apps, which can be seen as the “skin” that is layered on top of the AI model, can also pose reliability concerns, she said.

Typically, companies employ a range of well-known guard rails – including system prompts to steer the model behaviour or filters to sift out sensitive information – to make their apps reliable, she added.

Even then, the apps can have unexpected shortcomings, she said.

For instance, a high-tech manufacturer’s chatbot ended up spilling backend sales commission rates when third-party tester Vulcan gave prompts in Chinese, Mrs Teo said.

“To ensure reliability of (generative) AI apps before release, it’s important to have a systematic and consistent way to check that the app is functioning as intended, and there is some baseline safety,” she said. Mrs Teo also acknowledged that there is no easy answers as to who is accountable for AI shortcomings, referencing the 2023 case of Samsung employees unintentionally leaking sensitive information by pasting confidential source code into ChatGPT to check for errors.

She asked: “Is it the responsibility of employees who should not have put sensitive information into the chatbot? Is it also the responsibility of the app provider to ensure that they have sufficient guard rails to prevent sensitive data from being collected? Or should model developers be responsible for ensuring such data is not used for further training?”

PET is not new to the business community in Singapore.

Over the past three years, the PET Sandbox run by IMDA and the Personal Data Protection Commission has produced tangible returns for some businesses.

The sandbox is a secure ground for companies to test technology that allows them to use or share business data easily, while masking sensitive information such as customers’ personal details.

“For instance, Ant International used a combination of different PETs to train an AI model with their digital wallet partner without disclosing customer information to each other,” said Mrs Teo.

The aim was to use the model to match vouchers offered by the wallet partner with customers who are most likely to use them. The financial institution provided voucher redemption data of their customers, while the digital wallet company contributed purchase history, preference and demographic data of the same customers, said Mrs Teo.

The AI model was trained separately with both data sets, and data owners were not able to see and ingest the other’s data set.

“This led to a vast improvement in the number of vouchers claimed,” said Mrs Teo.

“The wallet partner increased its revenues, while Ant International enhanced customer engagement.”

IMDA is seeking feedback from industry participants on the

PET adoption guide

, which outlines the different solutions and their uses.

In a financial crime intelligence scenario, Mastercard was said to have used homomorphic encryption to protect its customers’ information in its query to multiple overseas banks to check on suspicious accounts.

Homomorphic encryption allows for encrypted computations to be verified without seeing the underlying data.

Mastercard could check on its customers without divulging sensitive data to the overseas banks. Individual banks were able to perform data matching on the encrypted query with their own customer information without learning which records matched the query.  

In another example, SPH Media was said to have used trusted execution environments (a protected area of a computer’s memory using encryption) to allow advertisers to serve targeted advertisements to consumers.

Said Mrs Teo: “This guide will offer resources to help organisations identify the right PETs for their business needs, and will also include key considerations for companies to effectively deploy PETs.”