Singapore launches 'nutrition label' guidelines for GenAI chatbots

Singapore launches 'nutrition label' guidelines for GenAI chatbots Companies are encouraged to give users a single place to check what their chatbot can do, how safe it is and how their data is used, IMDA says. SINGAPORE: "Nutrition labels" for generative artificial intelligence (GenAI) chatbots could soon become more common, with Singapore introducing voluntary guidelines for such platforms on Monday (Jul 20). The new guidelines encourage companies to include infocards for their GenAI chatbots, which could take different forms, such as a dedicated information page or a disclosure document. “Like a nutrition label or medicine label, it gives users a single, consolidated place to find essential information about the chatbot they are using,” said the Infocomm Media Development Authority (IMDA) in a press release. Among GenAI applications, people encounter chatbots the most, said Minister for Digital Development and Information Josephine Teo, announcing the guidelines at the Singapore Data Festival. Individuals use the application but may not know its limitations or what happens to their data, she said, noting that this information usually already exists but is scattered across terms of service and privacy notices. "The label does not tell us every scientific detail. Instead, it tells us the essentials. What the 'medicine' is for, how to take it, how much is recommended, what side effects to watch out for, when not to use it," said Mrs Teo. “We are starting with a voluntary framework and will refine it with industry inputs as practices mature.” These infocards should be relevant, accessible and timely, IMDA said. They should answer questions users actually care about – what the chatbot can and cannot do, how reliable and safe it is, how their data is handled and how to raise concerns or report issues. The information should be in plain language, easy to find and navigate, and kept up to date, it added. “As GenAI chatbots become embedded in everyday life, transparency is crucial as it allows users to understand what these applications can and cannot do, and how key risks are managed,” IMDA said. While the guidelines are voluntary, several organisations have indicated they intend to refer to them to strengthen transparency practices for their public-facing chatbots over the next six to 12 months, the authority said. The public sector will lead by example, with agencies including the National Library Board, Health Promotion Board and the Land Transport Authority planning to do the same. Companies including DBS, Google, Meta, OCBC and Singapore Airlines will also refer to the guidelines as they continue improving their chatbots' transparency, Mrs Teo said. USING DATA TO DEVELOP GENAI Mrs Teo also announced new advisory guidelines on how organisations should responsibly collect and use personal data to develop GenAI. “We take for example, a customer service team that wants to improve a GenAI model using call recordings, so they can respond quickly and accurately to customer queries. We've all been on the receiving end of these calls and being told that the call may be recorded for quality checks and improvement. "But we also know the recordings may contain personal data, such as our names, our addresses, billing details,” she said. “What are the obligations that these customer service teams will have to the customers before using their data for model training?” With more organisations incorporating GenAI into their work, IMDA said it was important to clarify how the Personal Data Protection Act applies – particularly around web scraping and the reuse of data originally collected for non-GenAI purposes. Organisations developing GenAI models can scrape publicly accessible personal data without consent, IMDA said. But if this data sits behind digital barriers such as paywalls or registration requirements, organisations need to assess whether it still qualifies as "publicly available". If it does not, and no other exception applies, consent must be obtained. Organisations have always needed consent to process personal data, including notifying individuals why their data is being collected. The new guidelines clarify that these notifications must now specifically flag when personal data will be used for GenAI development – a broad statement about "personalisation of services" or "product development" is no longer sufficient, IMDA said. Organisations should also spell out what data will be used, what it will be used for and how individuals can decline or withdraw consent. Using the customer service example, Mrs Teo said the company could update its privacy policy to state that recorded calls with consenting customers would be used to train and improve AI models. Staff scripts for seeking consent could also be updated, she added. Many organisations already provide similar AI-specific notices today, which is why the guidelines go further, Mrs Teo said, covering the roles and responsibilities of parties across the AI value chain and due diligence when relying on publicly available data. "With greater clarity on how existing requirements apply to GenAI, companies can design better processes with the right safeguards." Individuals can also request to access or correct their personal data even after it has been used to develop a GenAI model, according to the guidelines. IMDA acknowledged this may be difficult given the scale of data involved, and recommended organisations adopt practices such as upstream data handling, case-by-case review of requests and appropriate tracking and technical measures.