Singapore put AI ethics inside religious education. Turn values into usable choices

THE SIGNAL IN ONE SENTENCE

A value becomes useful around technology when it changes what someone does next. That sounds obvious. It is also where many AI ethics conversations quietly wander off into the decorative shrubbery. On October 25, the Islamic Religious Council of Singapore, known as MUIS, published a report on the conclusion of QuranSG 2026. The five-month season ran from June through October and recorded an estimated 15,000 attendances across competitions, talks, roadshows and a festival. Its closing seminar brought roughly 400 participants, including religious teachers and members of the public, to Suntec Convention Centre on October 24. For the first time, the seminar included breakout sessions on three contemporary subjects: artificial intelligence, environmental sustainability, and peace and conflict. The AI room paired Ustaz Ridhwan Mohd Basor, founder of DigitalWell.ai and a Cambridge researcher in responsible AI and governance, with Ustaz Khairul Zaman of Darul Tafsir Singapore. According to MUIS's official annex, they examined the ethical implications of AI, the importance of preserving human judgment and independent thought, and the role Quranic values can play in responsible decisions about technology. That is a thoughtful place to begin. It is not yet a public curriculum, an evaluated behavior-change program or a complete Muslim position on AI. MUIS describes a seminar discussion. It does not publish the room's full questions, cases, conclusions or participant feedback. The plain signal is this: turn the values into usable choices. A person deciding whether to use AI rarely encounters a giant button labeled ethics. The choice arrives disguised as convenience. Should I paste this private family message into a chatbot? Should I forward this polished answer without opening its sources? Should I let an image generator depict something a community treats with care? Should I rely on a generated religious explanation when the question deserves a qualified teacher? Should I send this automated reply to someone who is grieving? Should I let the machine act, or should it merely help me think? Those are not abstract governance puzzles. They are ordinary moments with consequences. The useful community tool would be a short decision card built around reflection before action. Not a terms-of-service novel. Not a list of forbidden product names. Not twelve cheerful icons explaining that AI is both exciting and risky. Six questions will do more work. What am I asking the system to do? What information am I giving it? What evidence supports the answer? Who could be affected if it is wrong or shared? Can the action be undone? Which person remains responsible? That card would not claim to translate a religious tradition into six boxes. It would turn the seminar's stated concern for reflection, human judgment and responsible decision-making into a repeatable pause. The first question separates assistance from authority. Using a tool to reorganize notes is different from asking it to decide what is true. Drafting a checklist is different from issuing religious guidance. Translating a public notice is different from interpreting a disputed passage. Generating options is different from choosing on behalf of another person. The interface may look identical. The responsibility is not. Ask participants to sort familiar uses into three buckets: help me prepare, help me compare, and decide or act for me. Then discuss where each use needs evidence, consent or a qualified human. The point is not to make everyone agree on every boundary. The point is to notice that a boundary exists before the output arrives wearing a very confident blazer. The second question protects the person whose information makes the task possible. People paste more than prompts. They paste conversations, schoolwork, workplace files, health concerns, photographs, account details and the private stories of people who never agreed to join the experiment. A practical exercise should use a fictional prompt and ask participants to circle every piece of information that should be removed, generalized or kept out. Then run the task without the sensitive material. If the tool cannot help without receiving information it should not have, the correct action may be to close the tab. That is AI literacy too. The third question makes verification physical. Give the group a generated answer containing a mixture of checkable claims, vague advice and one invented citation. Do not ask whether it sounds wise. Ask which sentence can be traced to a real source, which source actually supports it and which part remains uncertain. Then make the participant open the source. Links are not evidence merely because they are blue. A source can exist and still fail to support the sentence attached to it. A generated summary can flatten context, merge viewpoints or quietly promote a minority interpretation into a universal rule. For religious material, the handoff matters. A tool can help locate, translate or organize material. A qualified teacher or trusted institution should remain visible when a question turns on interpretation, pastoral care or community practice. The system should not impersonate authority it does not possess. The fourth question widens the frame beyond the person holding the phone. An AI-made image might be funny to one person and degrading to another. An automated reply might save the sender time while making a vulnerable recipient feel brushed aside. A summary might remove the disagreement that gives a difficult issue its meaning. A convincing false claim can travel through a family or community group before anyone asks where it came from. The user should name the affected people before pressing send. Who is represented? Who is missing? Who bears the embarrassment, exclusion or repair work if the output fails? Who deserves to know that a machine helped produce it? The fifth question sorts low-stakes experiments from consequential actions. If a generated dinner plan is terrible, the family can order noodles. If a generated message damages a relationship, a deletion does not rewind the conversation. If a system submits a form, transfers money, publishes a claim or speaks for an institution, the cost of a mistake rises sharply. Reversibility is a useful community-level test because it does not require technical expertise. Before a system acts, ask whether the result can be inspected, stopped, corrected and explained. If the answer is no, move the decision back to a person. This aligns with a broader strand of Singapore's own AI governance work. The Infocomm Media Development Authority's 2026 framework for agentic AI emphasizes bounded use cases, meaningful human accountability, approval checkpoints, technical controls and end-user transparency and training. That framework is written for organizations, not religious education. The connection is an editorial inference, but a useful one: responsibility needs an owner and a stopping point in both settings. The sixth question prevents the familiar magic trick where everyone supervises the system and nobody owns the result. Name the responsible person before use. That person should have enough information, time and authority to reject the output. A human signature applied after an automated decision is not meaningful oversight if the reviewer cannot see the source, change the result or stop the action. Community education should test this with role play. One participant uses the tool. One verifies the output. One represents the affected person. One decides whether to publish or act. Give the team a time limit, an ambiguous source and a tempting shortcut. Then change one condition. The source disappears. The answer conflicts with a teacher. The private data belongs to someone else. The action cannot be reversed. Watch whether the decision changes. That is a better measure than asking whether participants enjoyed the seminar. Attendance tells us that people came. It does not tell us whether they can now recognize a risky prompt, preserve confidential information, trace a claim, disclose AI assistance or stop a system at the right moment. An evaluation can stay modest. Before the session, give participants three short scenarios and ask what they would do. Repeat the exercise afterward with different examples. Score the steps, not the slogan: identify the purpose, protect the data, inspect the source, name the affected person, check reversibility and keep a human owner. Follow up a month later with an optional anonymous exercise. Which habits survived? Which questions felt impractical? Where did participants still need a teacher, technologist or institution to help? Publish the aggregate findings without turning a religious education program into a surveillance project. The program should also preserve disagreement. Muslim communities are not one audience with one job, level of technical confidence or interpretation of every ethical question. Religious teachers, students, parents, creators, workers, older adults and software builders may see different benefits and risks. A useful curriculum can share a method for reflection without pretending every case has one automatic answer. Invite participants to document where the card is insufficient. Perhaps a teacher needs guidance for generated learning materials. A mosque administrator needs rules for membership data. A family needs help identifying a synthetic recording. A creator wants to know when disclosure is appropriate. A young person needs a private route to ask whether a chatbot interaction has crossed a line. Those gaps should feed the next session. The strongest version of this work would connect religious educators, AI practitioners, youth leaders, privacy experts and community members. The religious educator brings moral vocabulary and pastoral context. The practitioner explains what the system does and cannot promise. The privacy expert shows where information travels. The community member tests whether the advice survives an ordinary Tuesday night on a phone. No single expert covers all four. The output could remain pleasantly small: a one-page decision card, a facilitator guide, six local scenarios, a source-check exercise, a private-data exercise, a referral list and a short evaluation rubric. Publish it in the languages participants actually use. Make it printable, accessible and product-neutral. Product-neutral matters. Today's popular interface will change. The habit should survive it. Pause. Define the task. Protect the person. Trace the claim. Consider the consequence. Keep the human responsible. A laminated values poster is lovely. It is not a brake pedal. QuranSG's AI session matters because it placed emerging technology inside an existing practice of reflection rather than treating ethics as an appendix written after deployment. The next step is to show what that reflection looks like when a real person is one tap away from trusting, forwarding or automating something. Values should not merely decorate the conversation. They should change the next click.