Transparency Guidelines for Generative AI Chatbots

Key Findings

• — Deployers are 'strongly encouraged' to meet a minimum of at least one substantive disclosure in each of the four areas (capabilities and limits, reliability and safety, data handling, issue reporting), plus a high-level safety statement; the guidelines create no legal obligations and do not displace sector rules such as medical-device regulation
• — The safety area asks deployers to name the common risks their chatbot type poses, the safeguards in place, how they know the safeguards work, residual limitations, and user precautions; the document says emotional-safety risks such as addiction or self-harm are more salient for companion chatbots while content safety dominates for general-purpose chatbots, and that risks to younger users cut across types
• — An 'Emotional Safety and Healthy Usage' pointer, marked as particularly relevant to companion chatbots, covers measures to prevent the chatbot from encouraging excessive emotional reliance, how it responds to users expressing distress or self-harm ideation, and whether it offers session limits or usage nudges
• — Deployers are asked to consider whether the chatbot is accessible to minors when calibrating disclosures; the document cites parents' concern about inappropriate content and 'potential self-harm' and names children among vulnerable users at risk of detrimental effects
• — Timeliness provisions: the info card should be published by the time a chatbot is made available, kept current, and versioned; material changes that warrant an update include widespread adoption for high-risk uses the chatbot was not designed for, while minor model upgrades and interface changes may not
• — The guidelines classify chatbots into general-purpose (naming ChatGPT, Claude and the Gemini app), social-interaction and emotional-engagement (naming Character.AI and Replika) and specialised (customer-service) types, and state that future work may extend the approach to other generative-AI applications with sector regulators

Methodology Notes

Policy guidance document, 27 pages, with an executive summary, three principle chapters, a future-work section, Annex A (suggested disclosure pointers by risk category) and Annex B (a sample chatbot info card). The PDF carries no publication date; the date 2026-07-20 is that of the IMDA press release 'New Transparency Guidelines to help consumers use Generative AI Chatbots safely and responsibly', which is blocked to automated access on imda.gov.sg and was confirmed through three independent republications and law-firm notes (TNGlobal 2026-07-20, Baker McKenzie July 2026, Rajah & Tann). The imda.gov.sg host answers HTTP 202 with an AWS WAF challenge to direct fetches of both the HTML and the PDF; the PDF text was read in full through a reader proxy on 2026-09-21 and the title, page count and structure verified.

Sources