Good morning everyone, it’s always a pleasure to be among practitioners.
I want to begin by commending the Law Society for convening this gathering as part of the Singapore Data Festival's partner events. I’d also like to acknowledge and thank the Law Society and acknowledge your partnership as one of our communities of practice. We really value the contribution and conversions you’ve made for our policy thinking.
The themes of today's programme — artificial intelligence, data sharing and managing cybersecurity incidents — they are not merely technical matters. They give rise to legal questions that your clients are already grappling with, and the quality of the counsel you can offer depends on how well-equipped you are to navigate them.
So, to that end, I want to use today's time to walk you through guidance that the Personal Data Protection Commission has recently released. Each piece speaks directly to a theme on today's programme, and taken together, they represent our most comprehensive effort yet to equip organisations — and the lawyers who advise them — to navigate this rapidly evolving landscape.
Many of you will be advising clients who are either deploying Generative AI tools, products, or simply using them in day-to-day operations. The question we hear most often is: "What does the PDPA actually require when it comes to Gen AI?" Up till now, we did not have specific guidelines on how the PDPA applies to Gen AI, although as you know, the general legal obligations and principles would of course still apply.
The Commission has this week issued the Advisory Guidelines on Use of Personal Data in Generative AI. We consulted publicly and received feedback across a wide range of stakeholders. So, what does GenAI change, and why did we need new guidance? Let me highlight three new personal data issues that GenAI introduces, which the Guidelines directly address.
First, at the development stage, there is the matter of large-scale web scraping. Developers today collect vast amounts of online data, which may include personal data, to train their models. The first thing that advisory guidelines does is confirm that the Publicly Available Exception generally can be applied to web-scraped data. However, there is a legal assessment to be made as to whether the Exception can be relied on.
The Guidelines now clarify how model providers can rely on the Publicly Available Exception in different web scraping scenarios, including where personal data sits behind paywalls. Where one of the exceptions under the Act does not apply, there is also the issue of what organisations need to do to obtain meaningful consent to process personal data for Gen AI purposes.
It also explains what information organisations must provide when notifying about and seeking consent for personal data use in GenAI model training. This is a practical point for your clients. The status quo — where a broad privacy policy mentioning "personalisation of services" or "product development" will no longer pass muster in Gen AI development contexts. Where exceptions to consent do not apply, organisations must obtain consent by providing what we call an AI-Specific Notification: which is shorthand for having a statement that personal data will be used to train a GenAI model, along with information on what data is involved and how individuals can meaningfully consent, decline, or withdraw. We have been intentional about making clear that organisations still retain discretion to determine the appropriate manner and form of their notifications.
There are other aspects to the Guidelines which I won’t have time to go through. But as you engage with the guidelines, I hope you agree with me that these Guidelines are not merely a compliance checklist. They are a framework for advising clients on how to structure their GenAI development and deployment responsibly — from the contracts they sign with model providers, to the consent mechanisms they put in place, to the governance structures that they build internally. I invite all of you to carefully review them, and to encourage your clients to do the same.
My second theme is cybersecurity readiness. The threat landscape has evolved significantly. Past data breaches have seen a marked increase in ransomware attacks, web application attacks, and personal data exfiltration. In many cases, the root cause was not a sophisticated or unforeseeable attack, but caused by system misconfigurations and inadequate systems testing before deployment. These are preventable failures.
Having studied the range of data breaches that come to our attention, a less visible but important piece of work we have done is to update the Guide to Data Protection Practices for ICT Systems. This Guide is a single, practical compendium of data protection best practices for ICT systems. It is organised around five categories — Assets, Secure and Protect, Update, Backup, and Respond. It is designed to closely model CSA’s Cyber Essentials Mark (as part of Singapore Standards SS 712), providing organisations with a sound cybersecurity baseline for data protection. These elevated standards now define the baseline for responsible data protection practice.
For ease of adoption, the Guide builds on a mnemonic that PDPC has long advocated to Data Protection Officers — B.E.S.T. The B stands for Backup: back up data regularly and securely offsite for restoration purposes. The E stands for Encrypt: encrypt sensitive data, both at rest and in transit. The S stands for Strengthen: strengthen access controls, for example through multi-factor authentication, to reduce the risk of unauthorised access. And the T stands for Track: track data assets and ensure timely maintenance of systems with upgrades and patches.
The Guide now includes case studies drawn from lessons learnt from past data breaches — real-world examples that illustrate what went wrong and how it could have been prevented. For instance, the updated Guide now recommends that organisations deploying websites or web applications use a web application firewall as a first layer of defence against common web application attacks. A web application firewall is not a replacement for secure application development, but it intercepts attacks that may slip through application-level controls, or that exploit vulnerabilities in third-party components beyond the organisation's control.
The Guide also, for the first time, includes specific practices for organisations that adopt and use AI systems. Given the rapid pace of AI development, system owners must instil data protection practices from the outset and remain vigilant throughout the AI lifecycle. For those of you advising clients on their data protection obligations, or who are involved in regulatory investigations following a data breach, the Guide serves as a benchmark. It answers the question: "What is the minimum standard of data protection for ICT systems that a responsible organisation should have in place?" That is a question that will matter enormously in any enforcement or litigation context. I want to take a moment to thank the Law Society for the meaningful feedback that the legal community provided during our consultation process.
My third and final theme is about how we can share data more safely, more freely, and more productively, using Privacy Enhancing Technologies, or PETs. The challenge of data sharing is one that your clients in healthcare, finance, and technology face acutely. Organisations hold valuable data but face real barriers to sharing it — whether due to data protection regulations, cross-border considerations, commercial sensitivity, or simply the risk of losing control of proprietary information. The result is data silos: islands of information that could, if combined, generate enormous value, but which remain locked away.
There are legal and regulatory solutions, which you are familiar with. But PETs offer a technical solution. They are a family of technologies that can ease the flow of data, unlock data silos, and enable cross-border data flows — all while preserving privacy, provided they are implemented with care. They fall into two broad categories: those that allow the flow of insights instead of original data, and those that obfuscate or hide the original data.
IMDA launched the PETs Sandbox in 2022, and we have since completed nine real-world use cases across finance, healthcare, advertising, and technology — with organisations including Mastercard, Meta, Zuellig Pharma, Singapore General Hospital, and Grab, among others. The insights from these use cases have directly shaped our policy thinking and resulted in practical guides for industry.
Today, I am pleased to highlight two of these guides. The first is our new Guide to Federated Learning. Federated Learning is a machine learning approach that trains AI models across multiple data sources, without the data ever leaving its original location. Think of it as training AI models without moving the data. Each data source trains the model locally and shares only the resulting model updates, keeping raw data where it originates.
Why does this matter? Because it is privacy-preserving by design. It eliminates the need to move or pool sensitive data, reducing the risk of data exposure. It enables better AI development by allowing valuable but siloed data to contribute to model training without ever being centralised. And it enables secure collaboration — potentially allowing competitors, partners, or agencies to work together on AI development without surrendering control of their own data.
The applications are especially relevant to sectors your clients may be operating in. For example, in healthcare, Federated Learning can unlock larger, more diverse patient and clinical datasets while patient data stays local. This enables better predictive diagnostics and treatment recommendations. In finance, it enables cross-institutional collaboration for joint assessments, such as anti-money laundering and fraud detection, without institutions having to share sensitive raw transaction data with one another.
The second guide I want to highlight is our revised Guide to Synthetic Data Generation. Synthetic data is artificially generated data that mimics the statistical properties of real data, without retaining one-to-one links to real individuals. We have updated this guide as the field continues to develop, expanding our case studies to reflect how organisations are putting synthetic data to use in practice. This is an area which we are committed to tracking closely, and our guidance will continue to be updated to reflect that.
For lawyers, these guides introduce a set of technologies that your clients can practically adopt to enable data collaboration — safely and responsibly. They are tools for enabling innovation, not just managing risk.
Let me close with a word of acknowledgment and thanks to the legal community. I am conscious that many of you in this room have been navigating these very issues with your clients long before regulators like me have had the chance to catch up with formal guidance. You understand, perhaps better than most, how the gap between legal principle and operational reality plays out in practice. As a result, a number of you have told me that your role has evolved from purely legal advice to being a solutions architect or business advisor. I believe you see this as an opportunity, and I hope to continue to partner with you as we navigate these issues together.
Thank you.