# Pacing AI development is possible. How can Singapore contribute to this? > Source: > Published: 2026-09-21 03:00:00+00:00 # Pacing AI development is possible. How can Singapore contribute to this? The city-state plays a key role in verifiable AI governance THE [recent calls by US tech leaders](https://www.businesstimes.com.sg/opinion-features/lab-leaders-want-slow-down-ai-someone-needs-help-them) to pace the frontier are a sincere and important move for artificial intelligence governance. It reflects a recognition that AI companies’ commitments to safety and their commercial incentives are in serious conflict. [Anthropic’s Dario Amodei](https://www.businesstimes.com.sg/international/global/anthropic-ceo-urges-slower-ai-development-altman-musk-rally-behind-call) and [OpenAI’s Sam Altman](https://www.businesstimes.com.sg/companies-markets/gambling-our-lives-openai-open-slowing-cutting-edge-ai-ceo-sam-altman-tells-staff) have expressed support for embedding external auditors inside of their company to help verify safety claims. While we hope other companies follow suit in taking this important step, it remains insufficient as self-policing alone cannot resolve underlying commercial conflicts. Coordinated pacing of AI development at a global scale is possible, but will require the creation of mutually trusted mechanisms to make such agreements verifiable. [Singapore is in a unique position](https://www.businesstimes.com.sg/companies-markets/ai-firms-call-accountability-and-limits-opportunity-beckons-singapore) to contribute to these efforts. To understand the opportunity, we must examine what verifiable governance requires in practice. Pacing the frontier need not mean stopping all training: stronger verification can make more targeted restrictions credible by giving parties greater confidence that agreements are being followed. Agreements could require companies to conduct particular safety evaluations or implement safeguards before developing or deploying more capable models. In these cases, verification tools could help establish whether the required evaluations took place and whether the model serving users is the one that was evaluated. Furthermore, where agreements specify limits on model training, being able to externally verify whether computing infrastructure is being used to train new models or serve existing ones will thus be key. We need to develop new specialised software and data centre monitoring tools to reliably detect when this is happening. Work on this is already actively underway across various organisations. At Singapore AI Safety Hub (SASH), we are cautiously optimistic about solving the remaining technical challenges. ## Building global trust A diverse ecosystem of international organisations is already laying the technical groundwork for these monitoring capabilities. Serious contributions have been made by UK-based Amodo Design, Swedish organisation Lucid Computing and Israeli cryptography startup Attestable. However, technical tools developed purely in Western jurisdictions may not build international confidence on their own. Growing the number of Asian organisations working on developing these technologies will be important both to broaden the technical base and to increase confidence in the resulting tools. These tools will need to be developed in a manner that inspires trust internationally for them to be effective. This can involve researchers around the world developing open-source tools to monitor data centre activity in a robust and privacy-preserving way. Open sourcing these tools can be important because allowing public scrutiny has historically been a good way to foster trust and identify vulnerabilities. Protecting model weights, customer data and other commercially sensitive information will be important if companies and governments are to accept their use. External scrutiny of AI companies will be key to creating stronger industry incentives for safety. Organisations such as Model Evaluation and Threat Research in the US and Neo Research in Singapore can help provide this, but a wider international base of third-party AI evaluators will be important to create trust in safety assessments across countries and provide the capacity to scrutinise the AI industry as it grows. ## Singapore as convenor and testbed This is where Singapore’s unique institutional setting and policy ecosystem allows it to serve as a convenor and testbed. The country could play an important role in supporting such agreements. For example, hosting dialogues between technical experts developing verification mechanisms could help establish what evidence different parties would accept. Supporting joint pilots involving international researchers and local cloud providers would be a practical next step and aligns well with the growing AI assurance ecosystem in Singapore. A recent proof of concept aimed at facilitating third-party auditing – between Google DeepMind, Singapore AI Safety Institute and non-profits OpenMined and AI Verification and Evaluation Research Institute – is a great step in this direction. Businesses in Singapore may stand to benefit from growing demand for AI assurance and verification. There are commercial applications for verification technologies, including creating more privacy-preserving means of accessing AI services and checking that providers are using the models and safeguards they claim to be using. A bank adopting AI, for example, may want confidence that the model processing its requests is the same version that passed its security assessment. Agreements to pace the frontier of AI development will still leave significant room for economic growth stemming from AI. Major productivity gains could be reaped by developing more applications of existing AI systems and adopting them in a wider range of contexts. Our team at SASH has been working on prototyping these tools and attempting to catalyse research and policy interest in them. **The writers are from SASH. Zac Richardson is special projects manager and Miro Plueckebaum is founder and managing director**