SMS Tan Kiat How’s Opening Remarks at the FutureChina Global Forum

Distinguished guests, ladies and gentlemen,

1. Good afternoon. It is a pleasure to join all of you at the FutureChina Global Forum’s AI Track.

2. Let me start with the perspective I bring to today’s discussion. I have spent much of the last 25 years at the intersection of technology and public policy. I am fundamentally optimistic about technology — its potential to improve lives, strengthen businesses, and create new possibilities for society. I have also spent much of that time thinking about how we manage the risks and disruptions that come with technological change.

3. There is another question that is never far from my mind. How does it a small country like Singapore stay relevant, create value, and contribute? AI brings these three questions sharply together.

4. First, how do we fully realise the economic and societal value of AI?  Second, how do we manage the accompanying risks and disruptions? Not just the risks that may emerge in the future, but those that are already with us today. Third, what does all of this mean for Singapore?

5. These questions are connected because, ultimately, the challenge is not simply to build more capable AI. It is to build companies, workers, institutions, and societies that can turn increasingly capable AI into better outcomes for all — productively, safely and confidently.

6. Let me turn to the first question. How do we fully realise the value of AI? Much of today’s attention is focused on the current AI race – better models, more compute, more advanced chips, more data, and more energy.

7. That race will continue, but I think another race is becoming just as important – the adoption and transformation race. As advanced capabilities become more widely available, the bottleneck is shifting from the technology to the system around it.

8. There is a simple way to see this. Imagine giving two companies access to the same frontier AI model. Would they achieve the same results? Almost certainly not.

9. One company may give everyone access to an AI assistant – emails get drafted faster, documents get summarised more quickly, and coding becomes more efficient. These are useful improvements but, fundamentally, it is the same organisation doing the same work with a more powerful tool. They are incremental gains and not full business transformation.

10. I remember something similar when I first started work many decades ago. We were transitioning from typewriters to word processors. Before that, a supervisor might mark changes on a physical document, send it back to the typing pool, wait for a new version, make more changes, and send it back again. Word processing made all that much faster. But we were still using new technology to make an old workflow more efficient.

11. The bigger gains came when technology allowed us to redesign the workflow itself — people could work directly on documents, collaborate much more easily, and organise work differently. AI presents us with a similar choice, but on a much larger scale.

12. A second company may take the same AI and redesign the way work gets done. It combines the model with its own data, redesigns workflows, changes how people and machines divide the work. It also reorganises teams, retrains workers, changes how decisions are made, and performance is measured.

13. Same AI, with very different value. The model may be the same but the capability system around it is not, and that difference compounds. The second company learns from deployment and its people develop judgement about where AI works well and where it does not. Its managers become better at redesigning work and engineers become better at integrating each new generation of technology. So, when the next model arrives, it does not start again from zero but begins from a stronger base.

14. Capability is built by doing and capability compounds. This is why I believe the next bottleneck is increasingly not the intelligence itself. It is the capability of the system around it. And bottlenecks are often where economic value accrues. As intelligence becomes more abundant, the ability to absorb it, integrate it, and put it to work becomes more valuable.

15. That is at the enterprise level. If we think about other sectors, healthcare is a good example. At the Ministry of Health, I work on healthcare policy, particularly on using technology to transform our healthcare sector. The lesson there is very clear.

16. Imagine two health systems with access to AI with similar capabilities. Would they achieve the same health outcomes for its people? Almost certainly not. One may have an excellent model, but fragmented data, disconnected systems, and workflows that were never designed around AI.

17. Clinicians may be uncertain when to rely on its recommendations or when to challenge them, and the system may have no good way of measuring and learning whether deployment actually improves outcomes for the system and patients.

18. Another health system’s model may not be better, but it has much stronger foundations: trusted data, connected digital infrastructure, common platforms, healthcare workers equipped to know how and when to use the technology and improve it across its lifecycle, and governance that allows it to be developed and deployed with confidence.

19. The technology may be equally capable. The health system is not. That is why I sometimes think of it this way: Health impact = AI × health system capability.

20. AI can be a tremendous force multiplier, but its impact depends on the system surrounding it. This is also why Singapore's approach is not simply about developing individual AI healthcare applications. The application is visible. The capability system underneath it is what makes the application scalable.

21. Now if we extend the same thought experiment and imagine two economies with access to broadly similar AI technologies, will they derive the same economic value? Probably not.

22. One may have excellent connectivity and compute, but companies that struggle to redesign themselves. Or sophisticated companies, but workers without sufficient opportunities to build capability. Or strong research, but weak links between research and real-world deployment. Or innovative businesses, but institutions that cannot give sufficient confidence for consequential applications to scale.

23. Another economy may have built these capabilities together: technology and infrastructure, research and engineering, companies able to transform, workers able to adapt, trusted data, institutions able to govern and assure new technology, and strong connections between researchers, companies, investors, and real-world users, so that learning travels quickly through the system.

24. Same technology. Different capability system. Different outcome. This, to me, is a useful way to understand how Singapore is approaching AI. We know we will not lead every layer of the technology stack. We will not have the most compute. We will not build every frontier model. We will not manufacture every advanced chip. So, our strategy cannot simply be about possessing more AI. It has to be about becoming exceptionally good at putting AI to work.

25. This is why our national AI efforts increasingly focus not just on the technology itself, but on missions in areas such as healthcare, finance, advanced manufacturing, and connectivity — where AI can solve real problems, and where deployment itself builds capability. Because ultimately, AI does not transform an economy by itself. Companies transform. Workflows transform. Jobs transform. Institutions transform. And people learn to do things they could not do before.

26. And that brings us to the other side of the equation. The more deeply AI moves into real organisations and real systems, the more consequential it becomes. If we want to give AI greater access, autonomy, and responsibility, then our ability to manage it must grow alongside its capability.

27. There is a great deal of debate about the risks of AI, from technology displacing workers to much more extreme scenarios. Those debates matter. But we should not allow uncertainty about tomorrow to distract us from risks we can already see and steps we can take today.

28. Cybersecurity is one clear example. Capable AI models can already help bad actors identify vulnerabilities, write malicious code, automate reconnaissance, and execute attacks more quickly. Increasingly, agents can chain these actions together. That is a real change. AI changes the speed, scale, and economics of an attack, but many attacks still exploit vulnerabilities we already understand. An unpatched server. A weak credential. Excessive privileges. So, enterprises – all of us – are not powerless.

29. Basic discipline becomes more important when attacks can operate at machine speed. Patch quickly. Know your assets. Limit access. Use strong authentication. Monitor unusual behaviour. Keep logs. Test your controls. Be ready to respond.

30. AI does not make cyber hygiene obsolete, it makes it more important, and that is another side to the same problem. It is not only about defending against attackers using AI but also about securing AI and agents that we deploy. AI agents may access tools, call APIs, read databases, initiate transactions, or modify systems, which creates enormous possibilities for productivity.

31. But autonomy should not mean absence of control. When we deploy agents, we need to be clear about what they may access, what they may do on their own, what requires human approval, and whether their actions can be observed and stopped.

32. Autonomy needs appropriate safeguards. The same principle applies beyond individual enterprises. Sector regulators have an important role because the acceptable level of autonomy depends on the consequences. Regulation must be risk-based and sector-sensitive.

33. An agent booking a meeting is very different from one moving millions of dollars. An AI system helping with hospital administration is different from one making a clinical decision. Regulators do not need to treat every AI application the same.

34. In finance, that could mean authority limits, transaction thresholds, audit trails, and human escalation. In healthcare, it could mean clarity around accountability, evidence, human oversight, and the ability to intervene. Different sectors. Same principles.

35. Guardrails should be proportionate to the stakes. Singapore's experience has reinforced this practical approach. We started early with principles and governance frameworks. But very quickly, we confronted a practical question – how do you know whether the principles are being followed?

36. That led us from governance to testing — including AI Verify and the AI Verify Foundation — and now towards deeper technical evaluation and AI safety capabilities. Our progression experience is informative: Principles. Testing. Assurance. Monitoring. Learning.

37. Good governance is not simply about writing rules. It is about building the capability to understand what systems can do, test them, observe them in operation, and adapt safeguards as technology changes.

38. My simple message is this: we are not helpless. Enterprises can strengthen their defences and put boundaries around agents. Sector regulators can set proportionate safeguards. Governments can strengthen evaluation, incident reporting, and technical capability. Many of the tools we need are already within reach.

39. But there is also a second horizon. As AI systems become more capable, some will warrant stronger evaluation and assurance. We already understand this principle in healthcare. Take medicines – a promising molecule does not move straight from the laboratory into widespread use. We test it, gather evidence, assess safety and efficacy, and continue monitoring even after approval is given.

40. AI is different, and I am not suggesting that we simply copy pharmaceutical regulation, but the underlying discipline is useful. As capability and consequences increase, so should the strength of evaluation, evidence, and safeguards.

41. Not every AI system should face the same scrutiny. A simple productivity tool is different from a highly autonomous system capable of operating critical infrastructure, exploiting cyber vulnerabilities, or assisting with biological research. So, the approach should be proportionate: test what systems can do, strengthen safeguards as capabilities increase, and continue monitoring after deployment. Not drug-style regulation for AI. But drug-style discipline about evidence.

42. We are beginning to see new institutional approaches emerge. Frontier AI companies are experimenting with giving external evaluators deeper access to systems during development — for red-teaming, alignment assessments, and safeguard testing.

43. The interesting point is not the specific institutional model. It is the direction of travel. Evaluation is moving closer to where the technology is being built. And this raises an important question: Who has the capability, independence, and access to evaluate increasingly powerful AI credibly? That is an emerging capability.

44. Singapore has been building capabilities in AI testing and assurance, including through AI Verify, and we are working with international partners on evaluation and AI safety. We do not have all the answers, but we can contribute to developing the people, methods, and institutions needed to evaluate AI credibly. That can be particularly valuable in high-consequence areas such as healthcare, finance, critical infrastructure, and increasingly, autonomous systems.

45. And it is noteworthy that different ecosystems are beginning to confront similar questions. China released its updated AI Security Governance Framework earlier this month, updating its approach to emerging risks, including those created by increasingly capable AI agents.

46. Different systems may take different approaches, but there is a common underlying question. As AI becomes more capable and consequential, how do evaluation and safeguards keep pace? That is no longer something any single company, regulator, or country can solve on its own.

47. Which brings me to the international dimension. Competition will remain intense. That is particularly evident today. President Trump and President Xi are due to meet later today, with AI expected to be part of the broader conversation between the two countries. I do not want to anticipate the outcomes of those discussions. But the broader point is clear. Major technology companies will compete. Major economies will compete. Different technology ecosystems will develop along different paths. Realistically, we should not expect one global rulebook.

48. Competition does not eliminate shared interests, because as AI capabilities advance, there are some outcomes that nobody benefits from – major attacks on critical infrastructure, AI-enabled biological catastrophe, industrial-scale fraud and scams, and highly autonomous systems behaving in ways that cannot be effectively controlled. Competition at the frontier may be inevitable. Catastrophe is not.

49. So even amid competition, there is a mutual interest in establishing a floor beneath it. Just this week, Singapore joined countries from around the world in a call for stronger safeguards and international cooperation on frontier AI.  The point is not to make every country regulate AI in the same way. It is to build enough common ground around a few practical things.

50. First, credible approaches to evaluation. If increasingly capable models can assist with advanced cyber operations, biological research, or autonomous actions, we need credible ways to test those capabilities and greater consistency around what is being measured.

51. Second, meaningful thresholds and safeguards. If evaluations show that a system has crossed an important capability threshold, there should be clearer expectations about what additional safeguards follow. And for most capable frontier systems, guardrails should not simply be bolted on after the fact. Where serious risks are foreseeable, safeguards should increasingly be part of the design.

52. That can mean controlling access to dangerous capabilities, limiting what tools or systems a model can reach, building in monitoring and logging, testing attempts to bypass safeguards, and ensuring operators can intervene when necessary. Evaluation tells us what a system can do. Guardrails by design help determine what it is allowed to do.

53. Third, incident reporting and information sharing. When serious failures or unexpected behaviors occur, we should learn from them. If a vulnerability or failure mode could affect others, there is a shared interest in getting actionable information to the right people quickly. And importantly, we need communication channels that remain open even when competition is intense.

54. The floor beneath competition is therefore not about making everyone regulate in the same way. It is about building enough common ground around evaluation, safeguards, incident reporting, and communication so that competition does not become recklessness. Not uniformity. Not one rulebook. Enough common ground to prevent the worst outcomes.

55. So, we need to work on two horizons at once. For risks already here: strengthen basic defences, control access, and set boundaries around agents and build resilience. For future capabilities: strengthen evaluation, assurance and monitoring, and preserve enough cooperation across competing ecosystems to manage the most serious shared risks.

56. That points towards the role Singapore can play. If value increasingly comes from the systems around AI — and if trust and assurance are becoming essential parts of that system — what should a small and open economy choose to become exceptionally good at?

57. For Singapore, we must be clear-eyed on the race we are in. We cannot compete on scale and outspend the largest economies. But Singapore has never stayed relevant by simply trying to do what much larger countries do. We have always had to understand where the world is moving, understand our own strengths, and carve out a racetrack where those strengths matter.

58. Finance is a good example. Singapore did not become an international financial centre because we had the largest domestic market or the most domestic capital. We became relevant because companies, investors, and financial institutions from different parts of the world could operate here with confidence.

59. We built trusted institutions, clear rules, deep professional capability, and strong connections to markets much larger than our own. In other words, we became good at converting capital into economic activity and connecting it with opportunities across markets. I think there is a useful lesson here.

60. The race Singapore should run is not simply to accumulate the most AI capacity. It is to become exceptionally good at converting AI capability into economic and social value. Helping companies move from adopting AI to meaningful applications. From an application to a redesigned workflow. From pilot to production. And from deployment to measurable outcomes. If access to capable intelligence becomes more widespread, then the ability to absorb it, integrate it, and put it to work becomes more valuable.

61. Singapore should aim to become very good at that conversion. A place where technology companies work closely with demanding users. Where engineers are close to real operational problems. Where companies can test not only whether a solution performs well, but whether the whole system produces a better outcome. And where each deployment leaves our companies, workers, and institutions more capable for the next one. In much the same way that we built the ecosystem around our financial sector, we should build an ecosystem that makes conversion in AI one of Singapore's strengths.

62. But conversion alone is not enough. As systems become more autonomous, they can act, transact, and make decisions in the real world. Trust itself becomes an economic infrastructure.

63. Finance again gives us a useful illustration. Financial markets work because participants rely on identity, rules, audit, supervision, and infrastructure that allow people and institutions that may never have met to transact with confidence.

64. We are entering a similar phase with AI. Today, an AI system may recommend what I should buy. Tomorrow, an AI agent may buy it for me. Some very practical questions arise. Who authorised the agent? Whose behalf is it acting on? What is it allowed to do? Can its actions be traced? And who is responsible if something goes wrong? These are no longer abstract questions about AI governance. They are becoming requirements for commerce.

65. That is why I find one recent development in payments particularly interesting. Ant International, Mastercard, and Visa are working together on a Know-Your-Agent interoperability framework as part of a sandbox by the Monetary Authority of Singapore. They operate different networks and trust systems. The aim is not to make them identical but to develop common ways for an AI agent to be identified and trusted across those networks, while each retains its own controls.

66. That illustrates a larger principle: Interoperability does not require uniformity. Different systems can remain different. What matters is establishing enough common language, identity and trust for them to work together.

67. Singapore is often described as a bridge. That remains useful. But perhaps in the next phase of the digital economy, we should think one step further. A bridge connects two places, but interoperability allows many different systems to work together.

68. The future will contain different models, clouds, payment systems, identity regimes, data frameworks, and approaches to governance. We should not assume those differences will disappear, nor should they need to. What matters is preventing difference from becoming fragmentation, because that comes at a cost.

69. If a company must rebuild its technology, identity, assurance, and compliance architecture each time it enters a new market, scaling becomes more expensive. Large companies may absorb that cost. Smaller companies may simply stay home. Fragmentation becomes an economic tax. So, interoperability is not a technical or diplomatic issue, it is an economic capability.

70. This is also the instinct behind the Singapore-China Digital Policy Dialogue, which I co-chair. Much of the work is practical — trusted data flows, digital documents and credentials, and helping companies navigate different regulatory systems. These may sound technical. But they reduce friction for businesses operating across borders. They impact the real-world experience of the digital economy. Interoperability is built — standard by standard, process by process, and transaction by transaction.

71. So, when I think about Singapore's role in the next phase of the digital economy, I come back to three capabilities. Conversion. Trust. Interoperability. Conversion — turning technological capability into real economic and societal values. Trust — building the evaluation, assurance, and governance capabilities that allow increasingly powerful technologies to be deployed confidently. Interoperability — helping different technological, commercial, and regulatory ecosystems work together without requiring them to become identical.

72. These are not entirely new capabilities for Singapore. They build on things we have spent decades learning to do. Our port connects supply chains. Our financial centre connects capital and opportunities. Our digital economy connects businesses and markets. The AI era gives us new things to connect — intelligence, data, agents, and digital services. For a small country, relevance does not have to come from scale. It can come from being useful.

73. So, ladies and gentlemen, I began with three questions. How do we fully realise the value of AI? For me, it is to build the capability around it. How do we minimise harms, build assurance, and have enough common ground to put a floor beneath competition? And what race should Singapore run? Convert. Assure. Connect.

74. In some ways, AI is ushering in a new technological era. But the strategic challenge for Singapore is familiar. The first AI race was about making AI more capable. The next will increasingly be about how much more capable AI can make us.

75. For Singapore, relevance will not come from being the biggest. It will come from being useful in the next phase of an AI-powered global digital economy.

76. Thank you.