Clinician Experiences with Ambient AI Scribe Technology in Singapore: A Qualitative Study

This qualitative study of 28 clinicians at Singapore's Alexandra Hospital reveals that while ambient AI scribe technology offers significant potential to reduce administrative burden and enhance patient engagement, its successful implementation in Singapore's multilingual healthcare system requires addressing critical challenges related to documentation accuracy, workflow adaptation, and compliance with local privacy regulations.

The Gist

Imagine a doctor's office as a busy kitchen. For years, the chef (the doctor) has been trying to cook a delicious meal for the patient while simultaneously writing down the entire recipe, the exact temperature of the oven, and a list of every ingredient used—all while the customer is watching. This is what clinical documentation feels like: the doctor is so busy typing notes on a computer that they can't look the patient in the eye. This leads to a burnt-out chef who is tired, stressed, and less able to connect with the person they are trying to help.

This paper is a story about a new "kitchen assistant" called Ambient AI Scribe.

The New Assistant

Imagine a super-smart, invisible robot that sits in the corner of the kitchen. It listens to the conversation between the chef and the customer. Instead of the chef stopping to write, the robot listens, understands the story, and automatically writes down the recipe on a notepad for the chef to check later.

The researchers in Singapore wanted to see how this robot actually worked in a real hospital (Alexandra Hospital) and what the doctors and nurses thought about it. They talked to 28 different healthcare workers to get their honest opinions.

The Good News: Putting the Chef Back in the Kitchen

The biggest surprise was how much the doctors loved being able to look up.

• The "Eye Contact" Effect: Before, doctors were staring at a screen, typing furiously. With the robot, they could finally look their patients in the eye, listen to them, and build a real connection. One doctor said, "For the first time in years, I'm actually listening, not typing."
• The "Super-Listener": Patients felt heard. Especially for elderly patients, having a doctor who isn't distracted by a keyboard made them feel much more cared for.

The Bumpy Road: When the Robot Gets Confused

However, the robot isn't perfect yet. The study found several "glitches" that made the doctors nervous:

1. The "Hallucinating" Robot: Sometimes, the AI makes things up. One doctor found the robot wrote down that a patient was taking a specific heart medicine, but the patient had never taken it. It's like a robot chef writing "I added salt" when you never asked for salt. Doctors have to double-check everything, which sometimes takes just as much time as writing it themselves.
2. The "Language Barrier": Singapore is a melting pot of languages (English, Mandarin, Malay, Tamil, and many dialects). The robot is great at English, but when a patient switches to Mandarin or a local dialect, the robot gets garbled. It's like a translator who only speaks perfect French but gets confused when you switch to Italian mid-sentence.
3. The "Lost Voice": Doctors felt the notes sounded a bit robotic. They lacked the "clinical intuition" or the special way an experienced doctor would summarize a case. It felt like the notes were written by a trainee rather than a master chef.
4. The "Privacy Paranoia": Some patients were worried that the robot was recording everything, even the private family secrets they didn't want in their official medical file. Doctors had to be very careful to tell patients, "Hey, this robot is listening," which sometimes felt awkward.

The Verdict: A Promising Tool, But Not Ready for Prime Time

The study concludes that this AI assistant is a game-changer for reducing stress and helping doctors connect with patients. It's like giving the chef a helper who can handle the paperwork.

But to make it work perfectly, we need to:

• Teach the robot better languages so it understands Singapore's mix of dialects.
• Train the chefs on how to work with the robot without losing their own style.
• Make sure the robot doesn't lie (hallucinate) about medical facts.
• Be honest with the patients about what the robot is doing.

In short, the technology is a powerful tool that can save the day, but it needs a little more tuning and training before it can fully replace the human touch in the doctor's office. It's a step in the right direction, but we can't just flip a switch and expect perfection.

Technical Summary

1. Problem Statement

Clinical documentation is a primary driver of clinician burnout, with physicians spending approximately two hours on Electronic Health Record (EHR) tasks for every hour of direct patient care. While Ambient Artificial Intelligence (AI) scribe technology—using Large Language Models (LLMs) to passively transcribe and summarize clinical encounters—has shown promise in reducing documentation time in Western contexts (e.g., US, Australia), there is a critical gap in understanding its efficacy and implementation in the Asia-Pacific region.

Specifically, no studies have examined this technology within Singapore's unique healthcare ecosystem, characterized by:

• A highly multilingual patient population (English, Mandarin, Malay, Tamil, and dialects).
• Distinct regulatory frameworks, specifically the Personal Data Protection Act (PDPA).
• An integrated public healthcare delivery model.

2. Methodology

• Study Design: A qualitative descriptive study guided by the RE-AIM/PRISM implementation science framework and adhering to the COREQ (Consolidated Criteria for Reporting Qualitative Research) checklist.
• Setting: Alexandra Hospital (AH), a tertiary hospital within the National University Health System (NUHS) in Singapore. The hospital had implemented an enterprise-wide ambient AI scribe platform across outpatient clinics, selected inpatient wards, and the emergency department.
• Participants: 28 clinicians purposively sampled for diversity in role, specialty, and usage intensity:
  • 18 Physicians (64.3%)
  • 6 Advanced Practice Providers (21.4%)
  • 4 Allied Health Professionals (14.3%)
  • Usage levels ranged from high (>70% activation) to discontinued use.
• Data Collection: Semi-structured interviews conducted between January and April 2026 (30–55 minutes each) via secure video conferencing.
• Data Analysis: Reflexive thematic analysis following Braun and Clarke's six-phase approach. Three researchers independently coded transcripts using Dedoose software, achieving consensus through iterative discussion.

3. Key Contributions

This study provides the first qualitative exploration of ambient AI scribe technology in the Asia-Pacific context. Its primary technical contributions include:

• Multilingual Validation Gap: It identifies specific technical failures in code-switching and dialect handling (e.g., Mandarin, Hokkien, Malay) that are unique to Singapore and not fully captured in Western literature.
• Regulatory Context: It analyzes the intersection of AI documentation with Singapore's PDPA, highlighting specific tensions regarding consent capacity (especially in dementia patients) and data residency.
• Nuance in Accuracy: It moves beyond simple error rates to explore "clinical voice" loss, where AI-generated notes lack the prioritization and nuance of experienced clinicians, particularly in mental health and complex cases.
• Implementation Framework: It applies the RE-AIM/PRISM framework to a non-Western setting, offering a model for evaluating AI adoption in diverse healthcare systems.

4. Key Results

Five major themes emerged from the analysis:

A. Reclaiming Presence in the Clinical Encounter

• Benefit: Clinicians reported a significant restoration of eye contact and active listening, shifting focus from the screen to the patient.
• Impact: Patients, particularly the elderly, reported feeling "heard."
• Limitation: Benefits were less pronounced in procedure-heavy specialties (e.g., surgery, rehabilitation) where structured templates and exact measurements are required.

B. Navigating Accuracy and Trust

• Hallucinations: Clinicians reported AI fabricating medication histories or misattributing symptoms to the wrong person.
• Loss of Nuance: In psychiatry, the AI failed to distinguish between passive thoughts of death and active self-harm planning.
• The "Editing Paradox": While writing time decreased, time spent reading and correcting notes increased. Some clinicians felt the net time savings were negligible due to the cognitive load of verifying AI output.

C. Workflow Disruption and Adaptation

• Multilingual Failure: Accuracy dropped significantly when patients switched languages or used dialects. Clinicians often had to rewrite notes from scratch for non-English speakers.
• Behavioral Modification: Some clinicians altered their natural speech patterns (e.g., narrating clinical reasoning aloud) to "train" the AI, which some felt disrupted the therapeutic alliance.
• Setting Mismatch: The technology fit outpatient conversations well but created friction in inpatient wards where pre-templated workflows were already streamlined.

D. Privacy, Consent, and Ethical Tensions

• Consent Capacity: Concerns were raised about obtaining valid consent from patients with cognitive impairments (e.g., dementia).
• PDPA Compliance: Uncertainty existed regarding whether patients understood the implications of AI recording and where data was processed (cloud vs. local).
• Sensitive Data: Clinicians noted the difficulty of capturing sensitive family or personal disclosures that patients might not want in the permanent medical record, requiring active manual redaction.

E. Envisioning Sustainable Integration

• Training: Success was highly correlated with robust onboarding, peer support, and "clinical champions."
• Customization: Clinicians demanded the ability to customize templates (e.g., brief notes for follow-ups vs. detailed histories for new patients).
• Governance: Participants called for structured error reporting mechanisms and regular audits of note accuracy.

5. Significance and Implications

• For Practice: Deployment of ambient AI scribes in Singapore must be staged and specialty-sensitive. It is most viable for conversation-driven outpatient settings. Investment in multilingual AI capabilities is urgent to ensure equity for non-English speaking patients.
• For Policy: Consent processes must be transparent and aligned with the PDPA, with special protocols for vulnerable populations.
• For Research: Future work must focus on longitudinal studies of sustained benefits, patient perspectives on AI documentation, and the development of AI models specifically trained on Asian languages and dialects.

Conclusion: While ambient AI scribes hold substantial promise for alleviating documentation burdens in Singapore, realizing this potential requires moving beyond efficiency metrics to address safety validation, linguistic inclusivity, and ethical compliance. The technology's success depends on a human-centric implementation strategy that respects the complex cultural and regulatory landscape of Singapore's healthcare system.