The short answer
Picture an AI tool that suggests discharges, and a ward that starts accepting them without a second look. Singapore’s updated guidelines say plainly that AI should not make clinical decisions on its own without human oversight. They also ask hospitals to list their AI tools, watch them, and plan how to switch one off.
AIHGle 2.0 is the updated Artificial Intelligence in Healthcare Guidelines from Singapore’s Ministry of Health and Health Sciences Authority, launched on 10 March 2026. It builds on the 2021 framework, adds guidance on generative AI, and spells out separate responsibilities for developers, deployers and users.2 For organisations deploying AI, the sharpest line is in its risk framework: deployers should not deploy AI solutions that make clinical decisions independently without human oversight.1 It also asks deployers to keep a registry of AI solutions, monitor them, and have a plan that includes shutting an AI solution down when something goes wrong.1
Deedglass
How Deedglass keeps an AI agent in check
-
1Before it runs
Your rules decide
- Read a recordAllowed
- Delete a tableNever
- Change a medicationNeeds a person
-
2If it’s risky
A named person approves
NSDr. N. ShahApproved in SlackDeclineApprove -
3After it runs
Checks what really changed
MatchesRecord sealedCheck it offline
What changed in 2.0
MOH’s circular to licensees lists three key updates:2
- Clearer accountability, through defined responsibilities for developers, deployers and users.
- Guidance on transparency, to support informed decisions.
- Updated guidance on AI deployment, including how to assess and mitigate risks.
The guidelines are accompanied by a deployers’ toolkit for new adopters and an infographic of roles. MOH calls AIHGle a living document to be updated as technology changes.1 AIHGle is a set of guidelines, and the circular says it complements HSA’s regulatory guidelines for software as a medical device, which cover AI that is a medical device.
Who the guidelines cover
AIHGle applies broadly to AI in healthcare but is aimed at the more complex subset that uses machine learning or deep learning, including generative AI, because of risks from complexity, opacity and scale.1 It groups use cases into three kinds: clinical, clinical-ops and ops.1
The three roles matter because each gets its own section:
| Role | Who it means in practice |
|---|---|
| Developer | The company or team that builds the AI solution |
| Deployer | The healthcare organisation that puts it into use |
| User | The healthcare professional working with it |
A hospital that builds its own tool is both developer and deployer, and carries both sets of duties.
Seven ethical principles
The guidelines adopt seven principles: safety, fairness, transparency, explainability, robustness, security and data protection, and alignment of AI to human values or goals.1 Each is mapped to examples of what developers, deployers and users should do. The principles are general; the deployment sections make them concrete.
Human oversight: the three levels
The risk assessment framework uses three broad categories of human oversight.1
| Level | What it means | Guidance in the framework |
|---|---|---|
| Human-in-the-loop | The human keeps full control; the AI only recommends | Example risk levels range from negligible to moderate depending on the use |
| Human-over-the-loop | The human monitors and can take over when the AI fails | For clinical decisions, deployers should not deploy AI that decides without a human making the final decision |
| Human-out-of-the-loop | No human oversight; the AI acts on its own output | Deployers should not deploy AI that makes clinical decisions independently without human oversight |
The examples in the table help. An imaging tool that highlights possible abnormalities while doctors make the final call is human-in-the-loop. A speech-to-text tool that drafts case notes for a doctor’s review is also human-in-the-loop, with the doctor in full control of what gets saved.1
The framework treats clinical decisions without human oversight as something deployers should not do at all. For clinical-ops uses, its examples include human-over-the-loop supervision, such as a pharmacist monitoring a medication-education tool, provided deployers assess risks, train users and put governance in place.
What deployers are asked to do
The deployer section runs from governance to end of life. The points most relevant to IT and operations teams:1
- A governance platform. Set up internal governance that assesses and approves each AI solution before deployment and documents the decision.
- A registry. Keep a registry of all AI solutions deployed in the organisation.
- Monitoring. Monitor performance on both an ad hoc and a regular basis, so that it at least meets the clinical practice baseline, and set up user feedback channels.
- Adverse events. Set up a process to detect, respond to, investigate and report adverse events. Contingency plans should include shutting down the AI solution and switching to protocols that don’t involve it.
- External reporting. Follow existing reporting duties: data incidents to the Personal Data Protection Commission, medical device adverse events to HSA, and critical cybersecurity incidents to the Cyber Security Agency.
The shutdown expectation is the one many deployments miss. If an AI tool is wired into clinical workflows, someone needs to be able to turn it off quickly, and staff need to know the fallback process. Our posts on AI agent governance in hospitals and incident response for AI errors go further on both.
Generative AI
AIHGle 2.0 has a section on generative AI. It lists four risks that GenAI amplifies:1
| Risk | Example from the guidelines’ framing |
|---|---|
| Hallucination | Incomplete or made-up content presented as fact, which users may not notice |
| Undesirable content | Harmful outputs that affect clinical outcomes or patient safety |
| Data disclosure | Sensitive data leaked through prompts or model behaviour |
| Adversarial prompts | Deliberate attempts to manipulate the model’s output or behaviour |
The guidelines say automated text metrics have significant limits for healthcare content and that developers should use evaluation methods specific to the use case, combining objective metrics with human judgment. The adversarial prompt item is worth underlining: it is prompt injection by another name, and it applies to any GenAI tool that reads untrusted text such as patient messages, referral letters or web pages.
What this means for AI agents
AIHGle 2.0 doesn’t use the word “agent,” but agents that take actions in clinical or operational systems fall under its definitions. Three practical consequences:
- Write access needs a human gate for clinical actions. If an agent can place an order, change a medication list or update a record, the human oversight framework says a clinician should make the final decision.1 That means a real approval step, not a log reviewed later.
- The registry should list agents and what they can do. A registry entry that says “documentation assistant” without listing its write permissions doesn’t support the risk assessment.1
- Shutdown must be fast and tested. An agent wired into the EHR needs a switch that stops its actions, and a fallback process staff can follow.1
Approval fatigue is a real risk when every action needs sign-off; our post on automation bias in health IT approvals covers how to keep reviews meaningful. For Singapore’s data rules that sit alongside AIHGle, see our Health Information Act post, and our healthcare page shows how agent actions and approvals get recorded.
A checklist for deployers
- Decide whether each AI tool is a medical device; if so, HSA’s device rules apply too.2
- Classify each use case as clinical, clinical-ops or ops, and set its oversight level.1
- Don’t deploy AI that makes clinical decisions without human oversight.1
- Approve each AI solution through an internal governance process and record the decision.1
- Keep a registry of AI solutions, including agents and their permissions.1
- Monitor performance against the current clinical baseline and collect user feedback.1
- Write and test a shutdown and fallback plan.1
- Test GenAI tools against adversarial prompts and hallucination.1
This post summarises the published guidelines and is not legal or clinical advice. To see agent actions held for approval in your browser, open the playground; to grade what an agent can reach on your own machine, run npx deedglass scan (the grade, explained).
Written with AI assistance and checked against our claims register.