SINGAPORE: Proposed “nutrition labels” for artificial intelligence services could help users better understand how such systems work – but only if they are clearly designed, regularly updated and backed by the industry, experts told CNA.
The idea, which Digital Development and Information Minister Josephine Teo raised last month, is being “actively explored” by the government as part of broader efforts to strengthen trust and safety in the digital space.
While such labels have potential to improve transparency, Professor Simon Chesterman, who leads AI governance and policy at the National University of Singapore’s AI Institute, said poorly designed ones risk becoming another “box-ticking exercise” that burdens responsible developers while being ignored by everyone else.
To be effective, Prof Chesterman said, the labels should explain what a system is designed to do, what data it relies on, what its main limitations are, how often it is updated and who is accountable when something goes wrong.
CNA Games
Show More Show Less
“For consumer-facing tools, the most important point is not to create the illusion of precision, but to give people a better sense of when they should trust the output and when they should be cautious,” he added.
Echoing this, Professor Bo An, who heads the AI division at Nanyang Technological University’s (NTU) College of Computing and Data Science, cautioned that labels must strike a balance between clarity and detail.
“If they are too vague, too long or treated as a box-ticking exercise, most users will ignore them,” said Prof An.
Motorola Solutions, which introduced “AI nutrition labels” across its safety and security technologies in July last year, said it addressed the potential issue of information overload by using a layered approach.
The labels have a “glanceable” summary for general understanding but provide more detailed information through links, said the technology company’s senior vice-president Jehan Wickramasuriya, who leads Motorola Solutions’ AI research and development teams globally.
An example of an AI "nutrition label" launched by Motorola Solutions. (Image: Motorola Solutions website)
Another key issue is keeping labels up to date as AI systems evolve.
Experts highlighted that static labels could quickly become outdated, especially as models are frequently updated.
To address this, Prof An and Prof Chesterman suggested linking labels to regularly updated pages that show changes and revision dates.
This is similar to Motorola Solutions’ approach, where label updates are embedded into its development process so that they evolve alongside the technology.
Standardising AI labels across different systems will also be challenging, given the wide range of applications and risk levels.
"Explainability means something different to a data scientist than it does to a police chief or a concerned citizen,” said Mr Wickramasuriya, sharing that the company’s experience in standardising the language across diverse stakeholders, while ensuring that it remains understandable to all, has been difficult yet rewarding.
Despite this, there is broad agreement that some level of standardisation is needed.
Prof An said regulation should set minimum disclosure requirements, while the industry develops formats that are practical and easy for users to compare.
Prof Chesterman shared a similar view, saying baseline standards are necessary even if full uniformity is difficult.
Ultimately, experts said the goal of AI nutrition labels is to help users make better-informed decisions, rather than to achieve perfect transparency.
“In the misinformation context especially, the real goal is not perfect transparency, but better calibrated trust,” said Prof Chesterman.
Continue reading...
The idea, which Digital Development and Information Minister Josephine Teo raised last month, is being “actively explored” by the government as part of broader efforts to strengthen trust and safety in the digital space.
While such labels have potential to improve transparency, Professor Simon Chesterman, who leads AI governance and policy at the National University of Singapore’s AI Institute, said poorly designed ones risk becoming another “box-ticking exercise” that burdens responsible developers while being ignored by everyone else.
To be effective, Prof Chesterman said, the labels should explain what a system is designed to do, what data it relies on, what its main limitations are, how often it is updated and who is accountable when something goes wrong.
CNA Games
Show More Show Less
“For consumer-facing tools, the most important point is not to create the illusion of precision, but to give people a better sense of when they should trust the output and when they should be cautious,” he added.
Echoing this, Professor Bo An, who heads the AI division at Nanyang Technological University’s (NTU) College of Computing and Data Science, cautioned that labels must strike a balance between clarity and detail.
“If they are too vague, too long or treated as a box-ticking exercise, most users will ignore them,” said Prof An.
Motorola Solutions, which introduced “AI nutrition labels” across its safety and security technologies in July last year, said it addressed the potential issue of information overload by using a layered approach.
The labels have a “glanceable” summary for general understanding but provide more detailed information through links, said the technology company’s senior vice-president Jehan Wickramasuriya, who leads Motorola Solutions’ AI research and development teams globally.
An example of an AI "nutrition label" launched by Motorola Solutions. (Image: Motorola Solutions website)
KEEPING LABELS UPDATED AND UNDERSTANDABLE
Another key issue is keeping labels up to date as AI systems evolve.
Experts highlighted that static labels could quickly become outdated, especially as models are frequently updated.
To address this, Prof An and Prof Chesterman suggested linking labels to regularly updated pages that show changes and revision dates.
This is similar to Motorola Solutions’ approach, where label updates are embedded into its development process so that they evolve alongside the technology.
Standardising AI labels across different systems will also be challenging, given the wide range of applications and risk levels.
"Explainability means something different to a data scientist than it does to a police chief or a concerned citizen,” said Mr Wickramasuriya, sharing that the company’s experience in standardising the language across diverse stakeholders, while ensuring that it remains understandable to all, has been difficult yet rewarding.
Despite this, there is broad agreement that some level of standardisation is needed.
Prof An said regulation should set minimum disclosure requirements, while the industry develops formats that are practical and easy for users to compare.
Prof Chesterman shared a similar view, saying baseline standards are necessary even if full uniformity is difficult.
Ultimately, experts said the goal of AI nutrition labels is to help users make better-informed decisions, rather than to achieve perfect transparency.
“In the misinformation context especially, the real goal is not perfect transparency, but better calibrated trust,” said Prof Chesterman.
Related:
Continue reading...