Human Oversight in Clinical AI in Viet Nam: What Meaningful Oversight Looks Like in Practice and Why Responsibility Remains Human
IMPLEMENTATION INSIGHTS
Human Oversight in Clinical AI in Viet Nam: What Meaningful Oversight Looks Like in Practice and Why Responsibility Remains Human
GOVERNANCE & RESPONSIBLE ADOPTION
10/1/2026 · 19 min read
Last reviewed: 1 October 2026
When healthcare organizations discuss artificial intelligence, one reassurance appears again and again: “A human is still in the loop.” That sounds sensible, but it does not tell us very much.
A clinician may technically remain somewhere in the workflow while having little practical ability to challenge what the AI recommends. The output may arrive at a point when the decision has effectively already been made. The user may not understand the system's limitations. Workload may make meaningful review unrealistic. An interface may encourage rapid acceptance. Or the clinician may technically have an override button but face a workflow in which disagreeing with the AI is difficult, poorly documented or discouraged. In all of these situations, a person is present, but meaningful human oversight may still be weak.
This distinction has become particularly important in Viet Nam. The Law on Artificial Intelligence No. 134/2025/QH15, effective from 1 March 2026, establishes a human-centered principle for AI activities. It states that AI serves people and does not replace human authority and responsibility, while requiring human control and the ability to intervene in AI decisions and behavior. For healthcare specifically, the Law requires particular attention to patient safety, reliability under real-world conditions of use and protection of health data.[1]
The practical question for healthcare organizations is therefore not simply whether a person remains somewhere in the process. It is whether the right person has the information, competence, time and authority needed to exercise independent judgment and intervene when intervention matters.
Human oversight is not the same as human presence
A person clicking “accept” does not necessarily constitute meaningful oversight. Imagine an imaging system that prioritizes examinations it considers urgent. A radiologist remains responsible for reviewing the images, so technically a human remains involved. But if clinicians gradually assume that examinations not flagged by the AI are unlikely to contain urgent findings, the system may influence attention before anyone consciously considers its recommendation.
An AI-supported endoscopy system provides another example. The endoscopist remains at the center of the procedure, but real-time highlights may influence what receives attention, when a lesion is reconsidered and how confidence develops during the examination. Generative AI creates a different version of the same problem. A clinician may review an AI-generated clinical summary before signing it, yet if the summary is long, plausible and usually correct, review can gradually become superficial. A human signature at the end of the process does not demonstrate that every clinically important statement was independently verified.
Meaningful oversight is therefore a functional property of the clinical system, not simply evidence that a human name appears somewhere in the workflow.
Viet Nam's AI framework makes human control a foundational principle
Article 4 of Viet Nam's AI Law does not reserve human control only for high-risk systems. Human-centeredness is one of the fundamental principles governing AI activities generally: AI serves people and does not replace human authority and responsibility, while human control and the ability to intervene in decisions and behavior of AI systems must be maintained. For high-risk AI, the requirements become more specific. Providers must design systems to support human oversight and intervention, while deployers must maintain the ability for human intervention during use.[1]
Decree No. 142/2026/NĐ-CP gives an unusually practical indication of what substantive human oversight can mean. In its framework for determining which systems should be proposed for the high-risk list, the Decree refers to “giám sát thực chất của con người”—substantive human oversight—as one relevant circumstance. It describes technical mechanisms and operational procedures under which an authorized person can independently review a system decision and intervene, reject or change it before that decision takes effect.[2]
In this article, meaningful human oversight is used as a practical governance concept closely reflecting that idea. It should not be treated as a separate statutory classification or as an official English defined term in Vietnamese law.
The Decree's provision also does not mean that simply adding a human reviewer automatically takes a system outside the high-risk regime. Risk classification depends on the full statutory framework and the characteristics of the particular system. What the provision demonstrates is that human involvement is expected to have substance: the person must have real capacity to influence what happens rather than merely confirm an automated result.
Decision No. 33/2026/QĐ-TTg reinforces the same principle. It states that use of AI does not change, transfer or exclude the legally assigned authority and responsibility of competent organizations or individuals and requires human supervision, control and the ability to intervene during system operation.[3]
Meaningful oversight begins with knowing what decision the AI is influencing
Healthcare organizations cannot design effective oversight unless they first identify what decision or action the AI actually influences. This is not always obvious.
Some systems make explicit recommendations such as “high risk,” “possible lesion,” “refer,” or “consider diagnosis X.” Others influence care indirectly. A triage algorithm may affect which patient is reviewed first. An imaging worklist system may influence which study receives earlier attention. A deterioration model may generate an alert that changes escalation behavior. A documentation system may determine which information becomes most prominent in a clinical summary.
Even when AI does not make the final clinical decision, it can influence attention, timing, framing and confidence. Oversight should therefore be designed around the real point of influence: what the system does, what happens immediately afterward, and what could happen if the output is wrong, late, missing or followed too readily.
Oversight should be proportionate to clinical consequence
Not every AI output needs the same level of human review. An administrative tool that helps categorize appointments presents a very different safety profile from an algorithm influencing emergency triage, treatment selection or a physical intervention. Applying the same oversight process to all of them would be inefficient and could itself create unnecessary workload.
A useful principle is that the greater the potential clinical consequence of an AI error, the stronger and more timely the human oversight should be. For a relatively low-consequence system, periodic review and clear escalation procedures may be sufficient. For AI influencing an important diagnostic or treatment decision, independent clinical review before action may be appropriate. For technology capable of controlling or directly influencing a physical intervention, intervention mechanisms and fail-safe controls may need to be considerably stronger.
This risk-proportionate approach is consistent with Viet Nam's broader AI framework, which differentiates regulatory obligations according to risk. It is also consistent with international healthcare AI governance. WHO places protection of human autonomy among its core principles for AI in health and states that humans should remain in control of healthcare systems and medical decisions. IMDRF's final Good Machine Learning Practice principles similarly emphasize evaluation of the human–AI team, rather than treating an AI-enabled medical device in isolation.[6,9]
Timing determines whether intervention is real
One of the most practical questions in oversight is when human review occurs. If the reviewer sees an AI recommendation only after an irreversible clinical action has already taken place, the review is retrospective rather than preventive. Retrospective review can still support audit and learning, but it cannot protect the patient from that particular decision.
This is why the wording of Decree No. 142 matters. Its substantive-human-oversight criterion describes an authorized person being able to independently review and intervene, reject or change a system decision before the decision takes effect.[2]
Healthcare organizations should apply the same practical reasoning even where that particular classification criterion does not directly determine the legal status of the system. If AI can influence an important or potentially irreversible action, human review needs to occur early enough to change what happens.
If AI supports patient prioritization, can a clinician change the priority before another patient's care is materially delayed? If AI recommends a treatment, does the treating clinician have an opportunity to review the relevant clinical evidence before the action is carried out? If AI is integrated into a robotic intervention, at which stages can an authorized professional stop, modify or confirm what the system is doing?
A theoretical right to intervene after the important event has already occurred provides limited protection.
The reviewer must be able to exercise independent judgment
Oversight also becomes weak when the human is expected simply to confirm what AI has already concluded. Meaningful oversight requires a realistic possibility that the reviewer can reach a different judgment.
The clinician therefore needs access to the clinical information required to assess the case rather than being shown only the AI output. In some use cases, the organization may also need to consider the order in which information is presented. Seeing an AI conclusion before forming an independent assessment can contribute to anchoring or automation bias.
There is no universal rule that clinicians must always remain blinded to AI until they have completed their own assessment. In many applications that would undermine the purpose of decision support. The point is that healthcare organizations should deliberately consider how the presentation and timing of AI information affect human reasoning, rather than assuming that simply displaying an AI output beside the medical record is neutral.
WHO has specifically highlighted automation bias and the risk of inappropriate delegation of difficult health decisions in relation to large multimodal and generative AI systems.[7]
The objective is not to make clinicians distrust AI. It is to build appropriately calibrated trust—neither accepting an output merely because it was produced by AI nor rejecting useful assistance simply because it came from a machine.
Humans need enough information to supervise the system
A clinician cannot meaningfully oversee a system that they fundamentally misunderstand. Before clinical use, users need to know what the AI is intended to do, which population and setting it was designed for, what important limitations are known, which information it uses, when its output may be unreliable and what they should do when AI and clinical judgment conflict.
That does not mean every user needs to understand model architecture, source code or advanced machine-learning mathematics. The information required for oversight should be relevant to the clinical decision.
For a diagnostic-support tool, users may need to understand what the output represents and what it does not represent. For a risk score, they should understand its target outcome and time horizon. For imaging AI, they need to know whether the product is designed to identify specified abnormalities or support broader interpretation. For generative AI, clinicians need to understand that fluent language does not establish factual accuracy.
Decree No. 142 requires providers of high-risk AI to give deployers information necessary for safe operation, including the intended purpose, safe operating conditions, identified risks and corresponding risk-control measures.[2] The National AI Ethics Framework likewise emphasizes transparency concerning purpose, scope and limitations, and calls for safe and responsible use of AI.[4]
Training should therefore focus not only on how to operate the software, but on how to recognize when its output deserves caution, verification or escalation.
Authority matters as much as knowledge
A clinician may recognize that an AI output is wrong and still lack practical authority to do anything about it. Meaningful oversight therefore requires real ability to intervene.
Depending on the use case, intervention may include rejecting a recommendation, changing a proposed action, requesting additional review, escalating to a senior clinician, returning temporarily to a manual process or stopping use of the system when a credible safety concern emerges.
Healthcare organizations should be cautious about workflows in which disagreeing with AI requires several steps of documentation or managerial approval while accepting the AI recommendation takes one click. Interface design and organizational culture can quietly turn a formally optional recommendation into a de facto default.
Authority must also exist at the organizational level. If several clinicians report unexpected system behavior, someone must have the authority to investigate, increase human review, restrict the use case, contact the provider or temporarily suspend the system.
Human oversight fails when everyone can see a problem but nobody has authority to act.
Oversight needs enough time to happen
Good oversight procedures on paper can become impossible in real clinical work. If a clinician is expected to review dozens of AI alerts during an already overloaded shift, the likely result may be rapid acceptance, alert fatigue or complete disregard. If every AI-generated note requires detailed line-by-line verification but the organization simultaneously expects documentation time to fall dramatically, those expectations may conflict.
Human oversight therefore has a workload dimension. Healthcare organizations need to estimate how much review the system creates, when that work occurs and who is expected to perform it. Oversight should be realistic not only during a controlled pilot but also during nights, weekends, high-volume periods and staff shortages.
Where the oversight requirement becomes excessively burdensome, the problem may lie in the AI threshold, interface, workflow or choice of use case rather than in clinician behavior. A safety model that depends indefinitely on extraordinary vigilance from exhausted staff is unlikely to be reliable.
Competence is part of oversight
The person supervising AI also needs the right competence for the decision involved. A technically skilled IT professional may understand system availability and integration but may not be qualified to determine whether a clinical recommendation is appropriate. A physician may understand the clinical context but require support from data or informatics specialists when investigating systematic performance changes.
Meaningful oversight is therefore often multidisciplinary. At the point of care, the relevant clinician needs sufficient AI literacy to understand the system's role and limitations. At the organizational level, clinical leaders, quality and patient-safety professionals, information technology, data or AI specialists, biomedical engineering, legal or compliance teams and vendors may each have different responsibilities.
Viet Nam's National AI Ethics Framework explicitly includes human control and also encourages capability building, AI-risk awareness and safe use among staff.[4]
A hospital does not need every clinician to become an AI engineer. It does need clinicians to recognize when an AI output deserves skepticism and know what to do when something does not look right.
Automation bias is only one side of the problem
Much discussion about human oversight focuses on automation bias—people accepting machine recommendations too readily. That risk is important, but the opposite can also occur. Clinicians may distrust a new system, ignore valid alerts or develop workarounds that prevent useful AI from contributing to care.
Poorly calibrated trust can therefore produce both over-reliance and underuse. Oversight should not aim for the highest possible override rate, nor should a healthcare organization assume that a very high AI acceptance rate proves successful implementation.
A clinician who disagrees with AI is not automatically correct, and an AI system that disagrees with a clinician is not automatically unsafe. The useful information lies in the pattern of disagreement. If clinicians repeatedly override the system in a particular patient group, department or clinical circumstance, that pattern may identify a performance problem, an inappropriate threshold, a workflow issue or a training need. If clinicians almost never disagree, the organization may need to determine whether the technology is genuinely highly reliable or whether staff have stopped exercising independent judgment.
Oversight itself therefore needs monitoring.
Generative AI makes meaningful review especially important
Generative AI deserves particular attention because its outputs can appear unusually convincing. Traditional predictive systems often produce a score, classification or highlighted image region. A large language or multimodal model can produce a fluent narrative, differential diagnosis, patient explanation, discharge summary or clinical note.
That fluency changes the human factors. An incorrect medication dose, omitted allergy or misleading clinical statement embedded within an otherwise professional-looking summary may be more difficult to notice than an obvious software failure. Users may also infer confidence from writing style even when the model has no reliable basis for the statement.
WHO's guidance on large multimodal models highlights risks associated with their use in healthcare and specifically addresses the governance challenges created by increasingly capable generative systems.[7] The current WHO publication page dates this guidance 25 March 2025.
For generative clinical tools, meaningful oversight may therefore require verification of clinically important facts against authoritative patient information, appropriate identification of AI-generated content, restrictions on unsupported autonomous actions and workflows that do not treat generated documentation as correct merely because it reads well.
At the same time, verification needs to remain operationally realistic. If clinicians must reconstruct the entire medical record every time AI generates a summary, the organization should question whether the system is actually delivering the value that justified its adoption.
Human oversight should not become a way to transfer vendor risk to clinicians
A healthcare technology company may describe its product as “decision support only” and state that the clinician remains responsible for clinical decisions. That may appropriately describe the intended role of the technology, but it does not automatically transfer every system-related responsibility to the individual clinician.
Providers, deployers and healthcare organizations have their own obligations. Under Viet Nam's high-risk AI framework, providers must establish and maintain risk-management systems, provide necessary information and maintain mechanisms for human oversight. Deployers must operate and monitor systems within their intended purpose and risk classification, maintain the ability for human intervention, respond to risks and cooperate with providers when significant changes or incidents occur.[1,2]
A clinician cannot reasonably compensate for inadequate model design, missing performance information, unsafe integration, deficient cybersecurity, poor procurement decisions or an organizational decision to use AI outside its intended purpose.
“Human oversight” should therefore never become shorthand for “the doctor assumes all the risk.”
Responsibility remains human, but responsibility is distributed
The title of this article requires an important qualification. Viet Nam's AI Law states that AI does not replace human authority and responsibility, while Decision No. 33 states that use of AI does not change, transfer or exclude legally assigned authority and responsibility.[1,3] That does not mean responsibility rests exclusively with whichever clinician happens to be the last person in the workflow.
The Law on Medical Examination and Treatment No. 15/2023/QH15 remains relevant. Article 45 requires practitioners to comply with professional and technical requirements and makes them responsible for the medical examination and treatment they provide. Article 62 further requires practitioners assigned to examination, treatment selection and prescribing to make timely and accurate decisions and take responsibility for those decisions. At the same time, Article 60 places responsibilities on healthcare establishments, including responsibility for the results of medical examination and treatment provided by practitioners under their management and for ensuring the conditions necessary for care.[5]
The AI Law creates another layer of responsibility. Article 29 establishes a specific rule for a high-risk AI system that has been managed, operated and used in accordance with the law but nevertheless causes damage: the deployer must compensate the affected person. After compensation, the deployer may seek reimbursement from the provider, developer or other relevant parties if there is an agreement between the parties. The Article also contains specified exemptions and separately addresses unlawful third-party interference.[1]
This provision should not be read as resolving every possible liability scenario involving clinical AI. Different facts may engage civil law, professional obligations, healthcare law, contractual arrangements and other legislation. Its relevance here is that Vietnamese law does not support the simplistic idea that placing a doctor “in the loop” automatically makes every system-related responsibility the doctor's alone.
Responsibility in clinical AI can therefore involve the practitioner, the healthcare organization as deployer, the provider, the developer and other relevant parties depending on what happened and the role each played.
Meaningful human oversight preserves human judgment. It does not erase organizational accountability.
Human oversight should be designed as a patient-safety control
Healthcare organizations may find it more useful to think about human oversight in the same way they think about other patient-safety controls. Safe medication processes do not rely only on telling clinicians to “be careful.” They may combine prescribing rules, pharmacist review, patient-identification checks, dose limits, escalation procedures, monitoring and incident review.
Clinical AI requires the same systems thinking. Human review is one safety control among several. A responsible implementation may combine appropriate validation, technical safeguards, human review, clinical protocols, logging, post-deployment monitoring, escalation arrangements, vendor support and a fallback pathway. If one control fails, another should still reduce the likelihood that an erroneous AI output contributes to patient harm.
This matters because humans themselves are imperfect safety barriers. People become tired, hurried and distracted. They develop habits. If an AI system produces a highly reliable recommendation thousands of times, clinicians can gradually become less likely to challenge it on the rare occasion when it is wrong.
The safer model is therefore not AI versus humans. It is a well-designed human–AI system in which technology and people compensate for each other's limitations.
Different forms of oversight may be appropriate for different systems
Terms such as human-in-the-loop, human-on-the-loop and human-in-command are widely used in AI governance discussions, but they are not separate statutory risk classifications under Viet Nam's AI Law. They can nevertheless be useful design concepts.
In some clinical workflows, a person should review and approve an AI recommendation before it becomes effective. In others, AI may operate continuously while a qualified person monitors system behavior and retains authority to intervene. Lower-consequence applications may rely more heavily on periodic audit, monitoring and escalation.
The appropriate model depends on what the AI does and what can happen when it is wrong. A hospital should therefore avoid treating “human-in-the-loop” as a universal slogan. The more useful question is what level and timing of human involvement are necessary to keep the actual clinical risk appropriately controlled.
Oversight needs a fallback when AI is unavailable or unreliable
Human control also means that care should not become unsafe simply because AI disappears. If diagnostic AI becomes unavailable, can clinicians continue interpreting cases appropriately? If an AI triage system goes offline, is there a manual prioritization pathway? If monitoring indicates that performance has deteriorated, can the organization restrict or suspend the system without losing its ability to provide essential care?
These questions are part of oversight because intervention is of limited value without an alternative. An organization that can technically shut down an unsafe AI system but cannot operationally deliver care without it may have much less real control than it assumes.
Fallback pathways should therefore be designed during implementation, not invented during a safety incident.
Meaningful oversight can be measured
Human oversight should not remain an abstract governance principle. Healthcare organizations can evaluate whether it actually works.
Depending on the use case, useful measures might include whether required reviews occur within the expected time, patterns of disagreement or override, escalation rates, use outside intended scope, AI-related near misses, user-reported concerns, time required for review and whether intervention mechanisms function during exercises or real incidents.
These numbers need interpretation. A high override rate is not inherently good or bad. A low escalation rate may mean the system is safe, or it may indicate that staff do not recognize or report problems. Quantitative measures should therefore be combined with structured feedback from clinicians and other users, who may detect workflow pressure, unexpected behavior or subtle changes before these become visible in formal performance statistics.
In other words, the organization should monitor whether human oversight remains meaningful after go-live, not simply document that an oversight process was designed before launch.
A practical test for healthcare organizations
Before deploying clinical AI, a healthcare organization should be able to answer several questions in plain language: Who is expected to supervise this system? What exactly are they supervising? At what point can they intervene? What information do they receive? What authority do they have if they disagree? How much time does the task realistically require? What happens when the AI repeatedly behaves unexpectedly?
If clinicians can override the AI, the organization should know whether those disagreements are captured and reviewed. If the system can be restricted or suspended, somebody should have defined authority to make that decision. The organization should also test whether the oversight process remains realistic during a busy clinical shift, not merely during a controlled demonstration.
If these questions cannot be answered, the organization may have a human somewhere in the workflow without yet having a meaningful human-oversight system.
The current high-risk list does not define the boundary of clinical oversight
As of 1 October 2026, Decision No. 33/2026/QĐ-TTg contains two healthcare entries in Viet Nam's current high-risk AI list, both relating to AI-assisted surgery or robotic surgical systems under specified conditions. The list does not currently classify every diagnostic, imaging, triage or clinical decision-support AI system as high-risk.[3]
That limited list should not be interpreted as meaning that meaningful oversight matters only for those systems. The AI Law's human-centered principle applies more broadly, while clinical governance should respond to the actual consequences of a technology even when it is not legally classified as high-risk.
A diagnostic algorithm, AI endoscopy system, imaging prioritization tool, deterioration model or generative clinical assistant may significantly influence patient care without appearing on the current high-risk list. The intensity of oversight should therefore reflect clinical consequence as well as regulatory classification.
There is also a transitional context. Under Article 35 of the AI Law, AI systems in healthcare that were already operating before 1 March 2026 have 18 months from the Law's effective date to complete applicable compliance obligations, resulting in a deadline of 1 September 2027. They may continue operating during that period unless the competent AI regulatory authority identifies a risk of serious harm and requires suspension or termination.[1]
Decision No. 33 provides additional transition rules for systems appearing on its high-risk list. Listed healthcare systems already operating before the Decision took effect on 15 August 2026 must complete applicable compliance obligations before 1 September 2027, and those pre-existing systems may continue operating during the transition unless the competent authority determines that they present a risk of serious harm. A separate rule applies to listed systems put into operation during the six months following the Decision's effective date: their providers and deployers must complete the applicable compliance obligations before 1 March 2027.[3]
Healthcare organizations should use these transition periods to build genuine governance and oversight capability rather than treating them only as documentation deadlines.
Viet Nam's National AI Ethics Framework supports the same direction
The National Artificial Intelligence Ethics Framework, issued under Circular No. 05/2026/TT-BKHCN and effective from 10 March 2026, provides another useful national reference. It emphasizes safety and reliability, appropriate testing, human control, accountability and responsible AI practices. In particular, it requires within its direct scope that organizations establish mechanisms for human supervision and intervention appropriate to system impact and maintain human control over AI decisions and behavior.[4]
Its legal scope must be stated carefully. The Circular applies directly to state-management agencies and to organizations or individuals acting as developers, providers, deployers or users of AI systems serving state management or the provision of public services. Vietnamese and foreign organizations and individuals undertaking AI research, development, provision, deployment or use in Viet Nam outside that direct scope are encouraged to apply the Framework.[4]
Healthcare organizations should therefore assess whether their particular activity falls within the Circular's direct scope rather than assuming that the Framework is universally mandatory. Even when applied on an encouraged basis, however, it provides a useful Vietnamese reference for designing human oversight and organizational accountability.
International guidance reaches a similar conclusion
WHO's 2021 guidance on AI for health places protection of human autonomy among its six core principles and states that humans should remain in control of healthcare systems and medical decisions. It also emphasizes responsibility and accountability across stakeholders rather than assuming responsibility belongs to one person at the end of the workflow.[6]
WHO's guidance on large multimodal models addresses the additional governance challenges created by increasingly capable generative systems in healthcare.[7] In the broader context of evidence-informed health policy, WHO's April 2026 discussion paper similarly states that AI should augment rather than replace human judgment and identifies human oversight and multidisciplinary collaboration among the practical considerations for responsible implementation. That document is directed primarily to policy-making rather than clinical AI, but the underlying governance principle is consistent.[8]
The IMDRF's final 2025 Good Machine Learning Practice principles likewise emphasize assessment of human–AI interactions and the performance of the human–AI team, rather than evaluating the device solely in isolation.[9]
These international materials do not replace Vietnamese law. They do, however, point in the same direction: human oversight works only when it is built into the design, workflow and governance of the system rather than added as a disclaimer after deployment.
What meaningful human oversight should look like in practice
A healthcare organization with mature AI governance would not simply write “the final decision remains with the physician” into a policy and consider the matter closed. It would identify which decisions or actions the AI influences and who is responsible for supervising them. Reviewers would have access to relevant clinical information and understand the system's intended use and limitations. The workflow would give authorized people a genuine ability to disagree, intervene, escalate and stop use when necessary, and staff would receive practical training appropriate to their role.
The organization would also recognize that oversight cannot be carried entirely by frontline clinicians. Providers need to supply usable information about system purpose, limitations, risks and changes. Healthcare organizations need appropriate procurement, evaluation, integration, monitoring and incident-management processes. Leadership needs to allocate authority for escalation and suspension. Quality and patient-safety teams need to consider AI when it contributes to an adverse event or near miss.
The oversight process itself should then be monitored after go-live to determine whether it continues to work as intended.
Meaningful human oversight is not a person standing behind the AI. It is an organizational capability.
The main message
AI can support clinicians by identifying patterns, prioritizing information, detecting possible abnormalities and processing volumes of information that would be difficult for people to review unaided. The purpose of human oversight is not to eliminate those benefits by requiring clinicians to independently redo everything the system has already done.
The purpose is to preserve something different: human clinical judgment, context, authority and accountability at the points where those qualities matter for patient care.
Viet Nam's 2026 AI framework makes this principle particularly clear. AI serves people and does not replace legally assigned human authority and responsibility. High-risk AI is subject to specific requirements for human supervision and intervention, while Decree No. 142's concept of substantive human oversight provides a useful practical benchmark: where such oversight is relevant, an authorized person should have genuine capacity to independently review and intervene in a system decision rather than merely confirm it.[1–3]
For healthcare organizations, the most useful question is therefore not simply “Is there still a human in the loop?” It is “Can the right human understand the role and limitations of the AI, exercise independent judgment, intervene at the right time and act when the system may be wrong?”
If the answer is no, human involvement may exist on paper, but meaningful oversight has not yet been achieved.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical or medical-device advice; determine the regulatory classification of a particular AI system; establish a Vietnamese clinical guideline; or determine civil, professional or organizational liability in a particular case.
The term clinical AI is used as a practical description and is not a formal risk category under Viet Nam's Law on Artificial Intelligence. The expression meaningful human oversight is used here as a governance concept, including as a practical English description of substantive human oversight where relevant; it should not be treated as a separate statutory classification. Likewise, terms such as human-in-the-loop, human-on-the-loop and human-in-command are implementation concepts rather than separate risk categories under Vietnamese law.
The principle that responsibility remains human should not be interpreted to mean that responsibility rests exclusively with an individual clinician. Depending on the circumstances, legal and professional responsibilities may involve practitioners, healthcare establishments, AI providers, deployers, developers and other relevant parties. Allocation of responsibility in a specific case requires consideration of the actual system, clinical decision, deployment arrangement, conduct of the parties, contractual relationships and applicable law.
The National Artificial Intelligence Ethics Framework has a defined scope under Circular No. 05/2026/TT-BKHCN. Organizations and activities outside its direct scope are encouraged, rather than universally required by that Circular, to apply the Framework. International guidance referenced in this article provides useful governance benchmarks but does not replace Vietnamese legal requirements.
References
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Artificial Intelligence. 30 April 2026; effective 1 May 2026.
Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. 30 June 2026; effective 15 August 2026.
Ministry of Science and Technology of Viet Nam. Circular No. 05/2026/TT-BKHCN Promulgating the National Artificial Intelligence Ethics Framework. 10 March 2026; effective 10 March 2026.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment. Adopted 9 January 2023; effective 1 January 2024.
World Health Organization. Ethics and Governance of Artificial Intelligence for Health: WHO Guidance. Geneva: WHO; 2021.
World Health Organization. Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models. Geneva: WHO; publication page dated 25 March 2025.
World Health Organization. Artificial Intelligence and Evidence-Informed Policy: Emerging Challenges and Opportunities. Discussion paper. 25 April 2026.
International Medical Device Regulators Forum. Good Machine Learning Practice for Medical Device Development: Guiding Principles. IMDRF/AIML WG/N88 FINAL:2025. 27 January 2025.