AI safety takes centre stage

THE whipsawing artificial intelligence (AI) trade saw another knee-jerk selloff this week – this time over AI safety fears.

Pioneers of the burgeoning technology have called for a slowdown in the pace at which AI companies build and improve systems and for stronger safety controls in the interim.

While some say the warnings should be heeded and that it is “high time” for companies to beef up the safety parameters of their AI models, others have dismissed them as “hype and puffery” and a self-serving attempt to mask slowing growth, as the initial public offerings of some AI giants look to be pushed out.

Whether it is fearmongering or prescient warnings remains to be seen.

Either way, Malaysia is moving ahead with its own AI governance framework, with the AI Governance Bill expected to be tabled by the first quarter of next year.

Data privacy, cybersecurity, misinformation and deepfakes, bias and discrimination, human oversight and accountability are among the key areas the government will need to address.

The level of autonomy granted to AI systems will also be an important consideration as AI agents become more capable.

Singapore, for instance, has developed AI Verify, a testing framework that allows companies to assess AI systems for safety, security, robustness, fairness, accountability and human agency.

The city-state has also established an AI Safety Institute focused on testing and evaluation, safe model development and governance.

The European Union has gone further with a risk-based regulatory regime, requiring higher-risk AI systems to undergo risk assessments, maintain human oversight, and meet requirements on transparency, cybersecurity, robustness and traceability.

Its rules for general-purpose AI models with systemic risk also require adversarial testing and mitigation of systemic risks.