Is Singapore ready for an AI outage?

The Straits Times, Is Singapore ready for an AI outage?

By Assistant Prof Roy Ka-Wei Lee, Information Systems Technology and Design (ISTD) & Design and Artificial Intelligence (DAI)

Anthropic suspended access to its most advanced models in response to a US government directive. It’s a warning sign.

Imagine you are overseas and trying to verify a suspicious bank transaction with an AI customer service assistant. Just as it is checking your identity, retrieving the transaction details and routing you to the right support channel, the system stops because the backend artificial intelligence model is no longer available.

For a bank customer, this is frustrating; for healthcare, public services or other critical services, the stakes can be even higher. This may sound hypothetical, but it is not far-fetched as AI becomes embedded into everyday digital services.

On June 12, Anthropic announced that it had suspended access to its Fable 5 and Mythos 5 models after receiving a United States government directive. According to the company, the order, triggered by national security concerns, applies to foreign nationals, whether inside or outside the US. To comply, Anthropic said it had to disable the models for all customers.

For many users, this may appear to be a dispute between one AI company and a government. For Singapore, it should be read more seriously.

It is a reminder that access to frontier AI can be abruptly cut off because a foreign government, regulator or model provider changes the rules.

Nevertheless, this is not a reason for Singapore to retreat from global AI; we will always remain an open economy that welcomes frontier technology firms and works with global partners. But it is a reason to be more clear-eyed about dependence.

Singapore has moved quickly on AI. Through our national AI strategies, we have invested in AI talent, research and adoption across sectors. More recently, Singapore’s Economic Strategy Review positions the nation as an AI-empowered economy and a global leader in deploying scalable AI solutions. But AI adoption is not the same as AI resilience.

Europe offers a useful warning. After years of dependence on foreign cloud and digital providers, countries such as France and Germany have moved towards trusted or sovereign cloud arrangements for public agencies and regulated sectors. The concern is not simply who provides the technology, but what happens when access, control or continuity becomes uncertain.

A country can use AI widely and still be vulnerable if critical systems depend too heavily on a small number of foreign models, cloud providers or vendors. The question is not only how fast we can adopt AI, but whether we can still stand, adapt and build if access changes tomorrow.

The answer to that is not technological isolationism. Singapore does not need to build every frontier model itself. What we need to do is to pursue a more flexible strategy: own the application layer, strengthen open and regional models, and turn global AI partnerships into deeper local capability.

Own it
Today, adopting AI can feel deceptively simple. A bank, insurer, clinic or government agency can connect a powerful model to its website or internal documents, and let it answer questions, summarise forms or route cases to staff.

To the user, it looks like a single seamless chatbot; behind the scenes, the service may rely on a model hosted and governed elsewhere. This works well – until the model changes, the terms change or access is withdrawn.

We may feel that we “have” AI because we can use it. But access is not the same as control.

While Singapore may not own any frontier model, we can own the way AI is applied. The value of AI lies in the workflows, safeguards, evaluation methods and human review processes surrounding the model.

Important systems should be model-agnostic. Organisations should know which model is best for each task, and which alternatives can take over if access changes.

A public service assistant, for example, could use a smaller local or open model for routine questions, a frontier model for complex cases, and human officers for sensitive decisions.

If one model becomes unavailable, a contingency model can step in, or the service may slow down or lose some features, but it should not collapse.

This is what resilience looks like in practice. Not every AI problem requires the most powerful model. Many practical uses can be handled by smaller, specialised or open models if the application layer is well designed.

Be open

Second, Singapore should support open-source and regional AI models. They are part of the resilience infrastructure, allowing researchers, companies and public agencies to inspect, adapt, fine-tune, deploy and evaluate AI systems with greater independence.

SEA-LION, Singapore’s open-source family of large language models for South-east Asian languages and contexts, is a good example.

Regional models need not beat the best frontier models on every benchmark. They matter because they address needs that global models may under-prioritise: local languages, cultural references, public-sector contexts, social norms and lower-resource communities.

For South-east Asia, this matters. Global models may not fully understand the multilingual reality of our societies, or local forms of misinformation, online harm, humour and cultural sensitivity.

Singapore should treat open and regional models as strategic public goods. Universities, research institutes, start-ups, public agencies and industry partners should contribute datasets, benchmarks, safety evaluations and fine-tuned models.

This is not about replacing frontier models completely, but knowing when transparent, adaptable and locally grounded alternatives are good enough.

When access to a frontier model changes, these alternatives can serve as fallback capacity, keeping routine services running while giving Singapore more time to adapt rather than wait for a vendor or a foreign regulator to restore access.

Make it win-win

Third, Singapore should not be merely a test bed for global deep-tech firms.

We should continue to welcome frontier AI companies. But the relationship must not be one-way. If Singapore is only a test bed, foreign firms gain deployment experience and regional market access while Singapore remains dependent on their platforms

Their presence should also become a training ground for Singapore. It should develop local engineers, AI researchers, product managers, safety evaluators and governance specialists who understand how to build, deploy and govern real AI systems.

Partnerships should create capability spillovers: joint labs, local hiring, internships, start-up formation, open-source contributions and benchmark development. If Singapore is good enough to be a test bed for frontier AI, it must also be a training ground for the people who build, evaluate and govern it.

That local capability is what allows Singapore to spot AI dependencies and prepare fallbacks before disruption happens.

Singapore should now map AI dependencies in critical sectors. High-impact AI systems should have continuity plans: fallback models, human escalation paths and clear procedures for what happens when access changes.

Every major AI partnership should also be judged not only by the technology it brings in, but by the capability it leaves behind: local talent, engineering depth, university and start-up links, open tools, evaluation standards and regional datasets.

These are what allow Singapore to adapt quickly when a model, vendor or regulator changes the rules.

The Anthropic episode should not make Singapore fearful of frontier AI. Instead, it should make us more clear-eyed. Singapore does not need AI isolation. It needs AI resilience.

The next stage of Singapore’s AI journey should be to ensure that when the AI plug is pulled elsewhere, Singapore can still stand, adapt and build.

• Roy Ka-Wei Lee is an associate professor at the Information Systems Technology and Design Pillar at Singapore University of Technology and Design.