The Role of Artificial Intelligence in Medical Diagnosis and Its Implications for Medical Negligence Liability

In February 2024, the Hospital Authority (HA) confirmed that it had deployed artificial intelligence (AI) software for the triage of CT brain scans across all public accident and emergency departments, processing over 130,000 scans in its first year of operation. This marks a concrete shift from pilot testing to routine clinical workflow in Hong Kong’s public healthcare system. The HA’s own 2023-2024 annual report states that the AI triage tool reduced the median time to radiology report for abnormal scans by 17 minutes. As AI systems move from assisting radiologists to making autonomous or semi-autonomous diagnostic decisions, a critical question arises for Hong Kong’s legal framework: who bears liability when the algorithm is wrong? The current law, grounded in the Medical Registration Ordinance (Cap. 161) and common law principles of medical negligence, was not designed for a clinical environment where a machine, not a human, interprets the primary data. This article examines the legal implications of AI-assisted diagnosis for medical negligence claims in Hong Kong, focusing on the standard of care, institutional liability, and the evidentiary challenges that litigants and their solicitors will face.

The Current Legal Framework for Medical Negligence in Hong Kong

The Bolam Test and Its Application to AI-Assisted Decisions

Hong Kong courts apply the Bolam test (derived from Bolam v Friern Hospital Management Committee [1957] 1 WLR 582) to determine the standard of care in medical negligence claims. The test asks whether the doctor acted in accordance with a practice accepted as proper by a responsible body of medical professionals skilled in that particular field. The Court of Final Appeal in Rogers v Whitaker (1992) 175 CLR 479, as adopted in Hong Kong, added a further layer: the court must be satisfied that the professional opinion is logical and reasonable.

The legislation provides no express provision for AI-assisted diagnosis. The standard of care under Cap. 161 therefore remains the conduct of a reasonable doctor, not a reasonable algorithm. This creates a tension. If a radiologist relies on an AI triage tool that flags a scan as normal, and the patient suffers harm from a missed intracranial haemorrhage, the court procedure is to assess whether the radiologist’s reliance on the AI output was itself a practice accepted as proper by a responsible body of peers. The HA’s own clinical guidelines, which mandate that AI outputs must be reviewed by a human radiologist before a final report is issued, become central to this inquiry.

The Distinction Between Clinical Judgment and Administrative Decision

The court procedure distinguishes between errors in clinical judgment and errors in administrative or systemic decisions. In Lai Kam Yin v The Incorporated Management Committee of the Kwong Wah Hospital (2011) 14 HKCFAR 795, the Court of Final Appeal held that a hospital’s systemic failure to implement adequate protocols could sound in direct institutional liability, separate from the individual doctor’s negligence.

This distinction is critical for AI-related claims. An error in the AI algorithm’s output—such as a false-negative classification of a pulmonary nodule—is not an error of clinical judgment by any human doctor. The doctor who overrides the AI output may still be negligent if the override was unreasonable. But the hospital or health authority that procured, deployed, and maintained the AI system may face direct liability for a systemic failure if the algorithm was not properly validated for the local patient population. The HA’s 2023-2024 annual report notes that the AI triage tool was validated on a dataset of 15,000 Hong Kong CT scans before deployment—a fact that a plaintiff’s solicitor would scrutinise in discovery.

Standard of Care When AI Is Involved

The Human-in-the-Loop Requirement

The HA’s current deployment model for the CT brain triage tool is a “human-in-the-loop” system. The AI generates a priority score and a preliminary classification (normal, abnormal, or urgent abnormal), but a radiologist must review every scan and issue the final report. The court procedure is that the standard of care remains that of the reviewing radiologist. The AI tool is merely an aid, analogous to a decision-support tool like a clinical decision rule or a laboratory reference range.

The legislation provides that a doctor cannot delegate the duty of care to a machine. In Cassidy v Ministry of Health [1951] 2 KB 343, the English Court of Appeal established that a hospital owes a non-delegable duty of care to its patients. This principle was affirmed by the Hong Kong Court of Appeal in Yuen Kwok Fung v Hospital Authority (1998) 1 HKLRD 142. Where the hospital delegates a diagnostic function to an AI system, the hospital remains liable for any failure in that function. The doctor who reviews the AI output must exercise independent clinical judgment. A doctor who blindly accepts an AI recommendation without applying reasonable scrutiny may be found negligent, even if the AI output was technically correct.

The Emerging “Reasonable Algorithm” Standard

Some academic commentators have proposed a separate standard of care for AI systems: the “reasonable algorithm” standard. This would ask whether the AI’s performance was within the range expected of a competent human practitioner in the same specialty. The Hong Kong courts have not yet adopted this standard. The court procedure is to apply the Bolam test to the human doctor, not to the machine.

However, the legislation provides a potential pathway through the Control of Exemption Clauses Ordinance (Cap. 71) and the Supply of Services (Implied Terms) Ordinance (Cap. 457) . If the AI system is procured as a service from a third-party vendor, the hospital may have a contractual claim against the vendor for breach of implied terms of reasonable care and skill. This does not assist the injured patient directly, but it creates an incentive for hospitals to pursue recovery from AI vendors, which in turn affects the allocation of liability in settlement negotiations.

Institutional Liability and the Role of the Hospital Authority

Direct Liability for System Design and Procurement

The Court of Final Appeal in Lai Kam Yin (2011) established that a hospital can be directly liable for a systemic failure, independent of any individual doctor’s negligence. The court procedure for such claims requires the plaintiff to prove that the hospital’s system of care fell below a reasonable standard. For AI-assisted diagnosis, this opens three specific avenues of inquiry:

1. Validation and calibration: Was the AI algorithm validated on a dataset representative of the Hong Kong population? The HA’s 2023-2024 annual report states its CT brain tool was validated on local scans, but a plaintiff’s solicitor would request the full validation protocol and the demographic breakdown of the training data.

2. Monitoring and updating: Did the hospital have a system in place to monitor the AI’s real-world performance and to update the algorithm when errors were detected? The HA’s own internal audit reports, if disclosed, would be relevant.

3. Human oversight protocols: Were the human-in-the-loop protocols adequate? For example, if the AI triage tool classifies a scan as “normal” and the radiologist, relying on that classification, does not scrutinise the scan as carefully as they would without the AI, the hospital’s protocol may itself be negligent.

Vicarious Liability for AI Errors

The legislation provides that an employer is vicariously liable for the negligence of its employees acting in the course of employment. For AI errors, the question is whether the AI system is an “employee” or an “instrument”. The Hong Kong courts have not addressed this directly. In the English case of The Ocean Victory [2017] UKSC 35, the Supreme Court held that a ship’s navigation system was an instrument, not an employee. By analogy, an AI diagnostic tool is an instrument of the hospital, not a person for whom the hospital is vicariously liable.

The practical consequence is that the hospital’s liability for an AI error will almost certainly be direct, not vicarious. The plaintiff must prove that the hospital itself—through its procurement, maintenance, or oversight of the AI system—failed to meet the standard of care. This places a heavier evidential burden on the plaintiff than a standard vicarious liability claim, where the employee’s negligence is often easier to prove.

Evidential Challenges in AI-Related Medical Negligence Claims

The Black Box Problem and Discovery

The court procedure for discovery in personal injury actions is governed by Order 24 of the Rules of the High Court (Cap. 4A). The plaintiff is entitled to inspect documents that are relevant to the issues in the action. For AI-related claims, the key documents include:

• The AI algorithm’s source code (or a functionally equivalent description)
• The training dataset and its demographic composition
• The validation protocols and performance metrics
• The system logs showing the specific AI output for the plaintiff’s scan
• The hospital’s internal protocols for AI-assisted diagnosis

The “black box” problem refers to the difficulty of explaining why a particular AI system reached a particular output. Some deep learning models are inherently opaque. The Hong Kong courts have not yet ruled on whether a hospital must disclose the source code of a proprietary AI system. The legislation provides that the court may order disclosure of any document that is relevant and necessary for the fair disposal of the matter. In Lau Tak Wo v HKSAR (2005) 8 HKCFAR 304, the Court of Final Appeal held that the court has inherent jurisdiction to regulate its own procedure. A plaintiff’s solicitor should apply for specific discovery of the AI system’s technical documentation at an early stage.

Causation and the Counterfactual

The court procedure for causation in medical negligence requires the plaintiff to prove, on the balance of probabilities, that the defendant’s breach of duty caused the plaintiff’s injury. For AI-related claims, the counterfactual inquiry is: what would have happened if the AI system had not been used?

If the AI system generated a false-negative result and the human radiologist also missed the abnormality, the hospital may argue that the AI did not cause the harm—the human error did. The plaintiff must then prove that the human radiologist would have detected the abnormality if they had not been influenced by the AI output. This is a difficult evidential burden. The court procedure allows expert evidence on this point. A radiologist with experience in AI-assisted workflow can opine on whether the AI output likely altered the human reviewer’s level of scrutiny.

Expert Evidence and the Standard of Proof

The legislation provides that expert evidence in civil proceedings is governed by Order 38 of the Rules of the High Court (Cap. 4A). The expert’s duty is to the court, not to the party instructing them. For AI-related claims, the expert must address two distinct questions:

1. The standard of care: Did the AI system meet the standard expected of a competent human practitioner? This requires the expert to compare the AI’s performance metrics (sensitivity, specificity, positive predictive value) against published benchmarks for human radiologists in Hong Kong.

2. Causation: Did the AI output cause the human error? The expert must consider the cognitive psychology of human-AI interaction, including automation bias (the tendency to over-rely on automated outputs).

The court procedure is that the expert must disclose the factual basis for their opinion. An expert who cannot explain why the AI reached a particular output may be of limited assistance to the court.

Practical Implications for Claimants and Their Solicitors

The Importance of Early Case Assessment

A solicitor handling a medical negligence claim involving AI must assess the case at the earliest possible stage. The key questions are:

• Was the AI system used in the diagnostic pathway?
• What was the specific AI output for the plaintiff’s case?
• Did a human clinician review the AI output, and if so, what was their independent assessment?
• Are there system logs or audit trails that document the AI’s output and the human’s response?

The court procedure for pre-action discovery under Order 24, rule 7A of the Rules of the High Court (Cap. 4A) allows a potential plaintiff to obtain documents before issuing a writ. This is particularly important for AI-related claims, where the relevant technical documents may be held by the hospital or the AI vendor.

The Role of the AI Vendor as a Potential Defendant

The legislation provides that a plaintiff may join multiple defendants in a single action where there is a common question of fact or law. In AI-related claims, the AI vendor may be a potential defendant, either in contract (if the hospital has a direct contractual relationship with the vendor) or in tort (if the vendor owed a duty of care to the patient).

The Court of Final Appeal in White v Jones [1995] 2 AC 207, as applied in Hong Kong, established that a professional can owe a duty of care to a third party who is not their client, provided the three-part test in Caparo Industries plc v Dickman [1990] 2 AC 605 is satisfied: foreseeability of harm, proximity of relationship, and whether it is fair, just, and reasonable to impose a duty. An AI vendor that designs a diagnostic algorithm for use in clinical settings may owe a duty of care to the patients whose scans are processed by that algorithm. This is a developing area of law, and no Hong Kong court has yet ruled on the point.

The Limitation Period

The legislation provides that the limitation period for personal injury claims is three years from the date of the injury or from the date of knowledge (section 27 of the Limitation Ordinance, Cap. 347). For AI-related claims, the date of knowledge may be difficult to determine if the patient does not know that an AI system was involved in their diagnosis. The court procedure is that the limitation period runs from the date the plaintiff knew (or ought reasonably to have known) that the injury was attributable to the defendant’s act or omission. A solicitor should investigate the use of AI in the diagnostic pathway as soon as the claim is intimated, to avoid any limitation issues.

Conclusion and Actionable Takeaways

The integration of AI into clinical diagnosis in Hong Kong’s public hospitals is a concrete reality, not a future hypothetical. The HA’s deployment of AI triage tools, the validation of these tools on local datasets, and the routine use of AI in radiology and pathology all point to a healthcare system that is increasingly reliant on algorithmic decision-making. The legal framework, grounded in the Bolam test and the principles of institutional liability, is not yet adapted to this new reality. Claimants and their solicitors must approach AI-related medical negligence claims with a clear understanding of the evidential challenges, the potential for direct institutional liability, and the importance of early discovery.

Three actionable takeaways for claimants and their solicitors:

1. Request the AI system logs and validation protocols at the pre-action discovery stage — these documents are essential to establish whether the AI system met the standard of care and whether the hospital’s oversight protocols were adequate.

2. Instruct an expert with specific experience in AI-assisted clinical workflows — the expert must address both the algorithmic performance and the cognitive impact of AI on human decision-making, including automation bias.

3. Consider joining the AI vendor as a potential defendant — the vendor may owe a direct duty of care to the patient under the Caparo test, particularly if the algorithm was not properly validated for the local population.

---

Disclaimer: This does not constitute legal advice. Consult a solicitor for your specific case.