June 5, 2024
SINGAPORE – Singapore wants to be at the front and centre of testing the safety of generative artificial intelligence (gen AI) models and has launched its most ambitious project yet.
Dubbed Project Moonshot, the new toolkit targeted at AI app developers aims to address safety and security challenges associated with large language models (LLMs) – the systems that form the cornerstone of many gen AI-driven solutions.
“Our objective is to provide a safer environment for businesses and citizens to use AI and enable AI innovations,” Minister for Communications and Information Josephine Teo said on May 31 at the ATxAI conference, part of the Asia Tech x Singapore event, held at Capella Singapore hotel.
She said, as its name suggests, Project Moonshot is “an undertaking to challenge ourselves”.
The increasingly widespread adoption of AI has resulted in people wanting more protections against risks caused by the technology, even as businesses worry that such protections could dampen innovation in the field, she added.
In Singapore, the authorities believe that in AI, governance is just as important as innovation, Mrs Teo said, describing Project Moonshot as the next step in the pursuit of good governance in AI.
“Good governance is not the enemy of innovation,” she said.
Project Moonshot was developed by the AI Verify Foundation and the Infocomm Media Development Authority (IMDA), through working with partners such as AI company DataRobot and investment firm Temasek to ensure its alignment with industry needs.
The open-source toolkit will allow businesses to assess their applications against specific benchmarks, such as whether they are able to understand local languages or cultural contexts.
Project Moonshot also employs attack modules to test whether applications can be manipulated into “misbehaving”, such as saying something inappropriate or producing violent content – a practice known as “red-teaming”.
IMDA BizTech group assistant director Thomas Tay noted that the Government Technology Agency had also contributed a tool that allows Project Moonshot to identify and evaluate vulgar Singlish terms.
“Together, benchmarking and red-teaming complement each other to help make AI better and safer,” Mr Tay said.
He added that Project Moonshot required organisations to only connect the toolkit to their applications and choose pre-established criteria to evaluate their applications against.
The authorities had in 2022 introduced the AI Verify toolkit, which allows companies to test their traditional AI systems against key ethics principles such as safety and fairness.
Mrs Teo noted additionally that two leading AI organisations, MLCommons and the AI Verify Foundation – a not-for-profit subsidiary of IMDA, whose 120 members include tech giants such as IBM and Google – had signed an agreement to develop a common testing benchmark for LLMs.
“This benchmark can be used to test their basic safety and trustworthiness against key indicators such as hate, toxic and violent content, regardless of the context or use cases,” she said.
To facilitate its AI governance ambitions, Singapore has redesignated its Digital Trust Centre as the country’s AI Safety Institute, which is part of an international network of similar institutes.
Digital Trust Centre was established in 2022 to lead research and development in technology that protects and secures digital information and systems.
While some experts have argued for an overarching AI law to protect societies, Mrs Teo said Singapore has no immediate plans to introduce one.
This is because existing laws and regulations already address some of the harms associated with AI, such as AI-generated fake news spread online, she added. These include laws for personal data protection and against online misinformation and disinformation.
“Regardless of how the fake news is produced, as long as there is public interest to debunk it, our laws already allow us to issue correction notices to alert people,” said Mrs Teo, who is also Second Minister for Home Affairs.
She added that in other cases, refreshing existing laws has allowed the authorities to address AI harms, such as when the Penal Code was updated to introduce the offence of sextortion – a form of blackmail where someone threatens to share intimate videos or images of another person.
This is illegal whether the images are real or AI-generated deepfakes, she noted. While the Republic is not defenceless against such AI-enabled harms, proper design and upstream measures are needed to prevent them from happening in the first place, Mrs Teo said.