The trust gap: why AI in healthcare must feel safe, not just be built safe

AI is rapidly becoming the first layer of engagement in healthcare, powering everything from symptom-checkers to therapy bots. Yet adoption is outpacing trust. In Singapore, a highly digitized and well-regulated environment, AI is ubiquitous (80% of residents use it) – but trust drops sharply once the advice moves into sensitive or emotionally charged domains like mental health. This pattern holds across South-East Asia, where one in four people in Indonesia and Hong Kong has tried an AI mental-health tool, yet concerns about safety and empathy persist. Globally, nearly 60% of Americans feel uneasy about AI-aided diagnosis. The bottleneck is no longer technical accuracy; it is a crisis of emotional assurance.

This trust deficit arises from three systemic gaps, visible even in advanced implementations:

1. Patients and clinicians often cannot see how AI risk scores are generated, creating a foundational trust gap. This structural opacity isn't merely an inconvenience; it can have direct clinical consequences and be a key factor in patient harm. For instance, the OECD AI Incidents Monitor documents cases in healthcare where flawed AI design led to biased outcomes, such as one system that unintentionally prioritized white patients over black patients by using healthcare costs as a proxy for medical needs. This real-world risk is confirmed by regulatory audits, such as a npj Digital Medicine 2025 study that found that over 90% of FDA-approved AI devices fail to report basic information about their training data or architecture. When the "why" behind a decision is invisible, safety becomes a matter of guesswork.

2. Accountability. The ITU’s AI Governance Report 2025 documents how responsibility fragments across developers, hospitals and ministries when private AI models enter public health workflows. This creates a tangible accountability gap: When an error occurs, there is no clear owner. The result is that patient grievances can enter a bureaucratic void, with no single party obligated to investigate, explain or provide redress, eroding trust following system failures.

3. Human control. When AI suggests a diagnosis first, it can reverse the traditional clinical workflow. Recent research confirms that when clinicians engage with AI-proposed diagnoses, their role is shifting toward verification, with their acceptance hinging on the model's ability to explain its reasoning. But the assumption of human oversight means little without explicit, operationalized checkpoints, such as mandating that clinicians must explicitly read, then accept or override AI diagnoses.

In healthcare, digital trust is a prerequisite for effectiveness. Evidence from digital mental-health deployments shows that unease with AI leads to lower engagement and earlier dropout – even when clinical accuracy is high. Users disengage not because the model is wrong, but because the experience feels unsafe.

Since AI cannot genuinely experience empathy, trust cannot be built on its capacity for human-like connection. Instead, as research into human-AI interaction clarifies, trust is established through respect for the patient's vulnerability; a dynamic defined by predictability, clarity and user control. This requires making three elements clear:

Health systems are moving from designing trust on paper to demonstrating it in practice. These reforms make safety observable – shifting trust from a promise to proof:

Building trust at scale requires minimum viable assurance – turning governance principles into visible, user-facing signals of safety. Three metrics, already within reach, do this:

As nations deploy AI across health systems, a new priority must guide governance: patient trust as a core performance indicator. This requires a fundamental shift from measuring only technical efficacy to evaluating human confidence.

Current assessments still prioritize model accuracy and efficiency. Yet real-world adoption hinges on whether tools feel safe and fair to use. To close this gap, policy must mandate continuous trust assurance alongside technical validation, moving beyond one-time audits to ongoing monitoring of real-world impact.

Critical actions include:

The next phase of digital health will be defined not by algorithmic performance in labs, but by confidence in clinics. Governance must be experienced, not just documented. For AI to fulfill its promise, trust must be engineered into every interaction.