SINGAPORE: The capacity for an artificial intelligence (AI) system to draw from multiple databases and act independently has introduced new risks, such as potential unauthorised payments and personal data theft.
Such risks, as well as guardrails to prevent the unintended outcomes, are outlined in a new framework, dubbed Model AI Governance Framework for Agentic AI, launched by the Infocomm Media Development Authority on Jan 22.
Unlike traditional AI, new agentic AI systems with the ability to understand natural language can reason and independently complete tasks on behalf of users.
AI agents include coding assistants that generate, test and debug code through prompts in natural language.
While these agents allow organisations to automate repetitive tasks and allow employees to take up “higher-value” activities, their ability to access sensitive data and make changes to the digital environment introduces new risks, said IMDA on Jan 22.
For instance, an agent could fix a medical appointment on the wrong date, negatively affecting a patient’s health outcomes. An agent could also take action outside its permitted scope of authority without human approval.
To reduce the chances of such incidents, the framework states that organisations should set limits on the number of tools and systems each agent has access to. For example, a coding assistant may not require access to a web search tool.
This framework is being introduced now to articulate the Government’s expectations while organisations are still in the midst of developing their own agent architecture, said Minister for Digital Development and Information Josephine Teo during a virtual doorstop with local media on Jan 22.
“The framework is still not as fully fleshed out as we would like it to be, but it’s still better to have articulated some expectations than none at all.”
Teo was speaking from Davos, Switzerland, where she participated in discussions about agentic AI and quantum technology at the World Economic Forum.
By launching the framework during the global forum, she added that the Government hopes to create awareness on what international providers should abide by when servicing clients in Singapore.
Despite the increasing capabilities of AI agents, the framework states that humans and organisations should still be accountable for their actions through clear allocation of responsibilities.
Organisations should also define significant checkpoints that require human intervention, which could include taking irreversible actions such as permanently deleting data, or when a delivery agent selects a route that is twice as long as its median distance.
The new framework was created with the help of feedback from government agencies and private sector organisations, said IMDA, adding that it welcomes more feedback and case studies on how AI agents can be safely deployed.
It builds upon an earlier version for traditional AI that was introduced in 2020, which set the foundation for how AI can be used ethically and responsibly.
By putting together a base of knowledge that organisations can refer to in the new framework, the Government also aims to make the know-how on implementing agentic AI more equitable among companies, including small and medium-sized enterprises, said Teo.
During a panel discussion about AI adoption and governance a day earlier on Jan 21, Teo said that AI agents could be potentially very helpful to seniors, who often need someone they can trust and help them navigate processes.
However, she added that people should still be allowed to interact with humans if they wish to, and companies should assess the risks of using AI agents in different contexts.
“There will be certain transactions where it may not be such a good idea to test out the use of AI agents,” said Teo.
“You want to start with things that people feel quite comfortable about, that are not very high risk, and build confidence over time.”
Ultimately, the aim of using AI agents is to bring the Government and its people closer together by reducing system-level friction, said Teo.
“It’s not always going to be at the interaction point with the citizen. It could very much be at the back end, making life easier... (so that) the interaction with the state is a more pleasant one.” - The Straits Times/ANN
