This post takes that idea further by grounding it in the work of philosopher Miranda Fricker. She distinguishes two key forms of wrong: testimonial injustice and hermeneutical injustice. When we look at clinical AI through this lens, especially in global health contexts, a troubling pattern emerges. Our models are not just biased. They are systematically silencing certain ways of knowing.
And I have seen this tension firsthand in my work on dengue triage systems in Bangladesh and fairness aware ECG models.
When Data Speaks, Who Is Allowed to Be Heard?
Clinical AI systems are often trained on datasets from high income countries. These datasets are treated as neutral, objective, and generalizable. But they are anything but.
They encode specific healthcare infrastructures, culturally shaped symptom reporting, and demographic distributions that rarely reflect the Global South. When such data becomes the foundation of intelligent systems deployed globally, we are not just exporting models. We are exporting epistemologies.
Testimonial Injustice in Clinical Data
Fricker describes testimonial injustice as a credibility deficit assigned to someone's word due to prejudice. In clinical AI, this happens in a subtle but powerful way.
In my dengue triage chatbot work, patients in Bangladesh often describe symptoms differently than what is captured in structured datasets. Expressions of fatigue, pain, or warning signs of dengue are shaped by language, access to care, and lived experience. But training data does not treat all expressions equally. Symptoms common in Western datasets are weighted as reliable signals. Locally expressed or less formally documented symptoms are treated as noise. Informal care seeking patterns are often excluded entirely.
The result is a system that implicitly says some patients are more credible than others. Not because of malicious intent, but because their experiences were never meaningfully encoded. This is testimonial injustice at scale.
Hermeneutical Injustice: When Entire Realities Are Missing
If testimonial injustice is about not being believed, hermeneutical injustice is about not being understood at all. This is even more dangerous.
In fairness aware ECG modeling, we often focus on demographic parity or representation learning. But what if the problem is deeper? What if certain cardiac patterns common in underrepresented populations are not well studied? What if wearable device data reflects usage patterns tied to socioeconomic status? What if clinical labels themselves are biased or incomplete?
In such cases, the issue is not just imbalance. It is absence. The system lacks the conceptual resources to interpret certain signals correctly. Entire physiological or experiential patterns remain invisible.
This aligns with recent work in medical AI ethics that argues epistemic harm is not only about biased outputs but about structural gaps in meaning making. When AI systems learn from incomplete clinical records or historically skewed datasets, they inherit not just bias but ignorance. And unlike human clinicians, they cannot question what they do not know.
Scaling Injustice Through Automation
One of the most unsettling insights from recent literature is that AI does not just replicate epistemic injustice. It scales it.
A misinterpretation in a single clinical encounter is harmful. The same misinterpretation embedded in an AI system becomes infrastructure. It affects thousands of patients. It standardizes exclusion. It becomes harder to detect because it appears consistent. In this sense, clinical AI transforms localized epistemic failures into global ones.
Why Fairness Metrics Are Not Enough
In my ECG work, fairness aware representation learning improved performance across demographic groups. But even then, something felt incomplete.
Fairness metrics answer questions like: Are predictions equally accurate across groups? Is error distributed evenly? They do not answer: Whose knowledge shaped the model? Which experiences are missing? What kinds of uncertainty remain invisible?
This is where epistemic injustice frameworks push us beyond technical fixes. They force us to ask whether the system is capable of understanding all the populations it serves.
Toward Epistemically Just Clinical AI
If epistemic injustice is built into data and design, then addressing it requires more than post hoc correction. It requires rethinking how we build systems from the ground up. Here are some directions that emerge from both research and practice.
Five Directions for Epistemically Just AI
- Participatory Data Collection in Low Resource Settings: Data should not be extracted. It should be co created. Engage local clinicians, patients, and health workers. Capture culturally specific symptom expressions. Include informal and community based care knowledge. This is not just about representation. It is about restoring epistemic agency.
- Context Aware Model Design: Models should not assume universality. Incorporate region specific features and priors. Allow adaptive decision pathways, especially in triage systems. Design hybrid systems where local rules complement learned patterns. This aligns with my earlier argument for arguable, hybrid clinical AI.
- Epistemic Auditing, Not Just Fairness Auditing: We need to audit not only outputs but knowledge structures. What data sources dominate the model? Which populations are under described or missing? Where does the model exhibit high uncertainty, and why? This extends the idea of epistemic responsibility I discussed in earlier posts.
- Designing for Uncertainty and Humility: A system that does not know should be able to say so. Explicit uncertainty reporting, referral mechanisms instead of forced predictions, and interfaces that invite clinician judgment rather than replace it. Epistemic justice requires humility, not just accuracy.
- Building Hermeneutical Bridges: When gaps in understanding exist, systems should help surface them. Flag atypical patterns rather than suppress them. Enable feedback loops from clinicians in diverse settings. Continuously update models with locally grounded knowledge. This is how systems evolve from static predictors to learning participants in global health ecosystems.
Closing Thought
It showed me that every dataset carries voices. And every model decides which of those voices matter.
Epistemic injustice reminds us that silence in data is not absence. It is often exclusion. If we want clinical AI to be truly global, it cannot just scale models. It must learn to listen.