The recent STAT+ article 'When the Path to Good AI is Littered with Bad Data' by Brittany Trang, Ph.D., highlights a critical yet under-discussed issue in healthcare AI: the reliance on flawed or biased datasets that can undermine diagnostic and predictive tools. While Trang’s piece identifies real-world examples—such as AI models trained on incomplete electronic health records (EHRs) leading to misdiagnoses—it stops short of dissecting the systemic roots of poor data quality, the ethical implications for patient trust, and the broader societal impact of deploying unvetted AI systems in high-stakes environments.

First, let’s contextualize the problem. Healthcare AI often pulls from datasets riddled with inconsistencies—think missing patient demographics, coding errors in EHRs, or underrepresentation of minority groups. A 2021 study in The Lancet Digital Health (Vol. 3, Issue 10) analyzed 81 AI tools for clinical use and found that 92% were trained on datasets with significant bias, often skewed toward majority populations in high-income countries (observational study, n=81 tools, no conflicts of interest disclosed). This isn’t just a technical glitch; it’s a pipeline issue stemming from decades of fragmented healthcare systems and underfunded data infrastructure. Trang’s article misses this historical angle, framing bad data as a contemporary tech problem rather than a symptom of long-standing inequities.

Second, the ethical stakes are higher than mainstream coverage suggests. When AI misdiagnoses due to bad data, it doesn’t just err—it erodes trust in medical systems already strained by disparities. Consider the 2019 case of Optum’s algorithm, which underestimated the needs of Black patients for care management due to biased historical spending data (as reported in Science, DOI: 10.1126/science.aax2342, observational analysis, n=unspecified, no conflicts noted). Trang’s piece glosses over how such failures disproportionately harm vulnerable groups, missing the chance to connect AI’s data problem to broader calls for health equity.

Third, there’s a pattern of tech-driven optimism blinding stakeholders to these risks. The rush to integrate AI into healthcare—spurred by venture capital and regulatory leniency—often prioritizes speed over scrutiny. The FDA’s 2023 guidance on AI/ML devices, while promising, still lacks robust mandates for data transparency, leaving gaps that bad data can slip through. Combining this with the Lancet findings, it’s clear that without systemic reform (e.g., mandatory bias audits or open-access datasets), AI’s promise in healthcare remains a mirage for many patients.

What’s missing from Trang’s coverage is a call to action. Beyond identifying bad data as a problem, we need to ask: Who is accountable when AI fails due to poor inputs? Is it the developer, the hospital, or the regulator? And how do we balance innovation with patient safety when the financial incentives lean so heavily toward deployment over diligence? These questions tie into a larger trend of tech exceptionalism in medicine, where tools are hailed as revolutionary before their flaws are fully understood—a pattern seen in everything from early telemedicine platforms to wearable health trackers.

Ultimately, the path to 'good AI' isn’t just littered with bad data; it’s paved with ethical blind spots and systemic neglect. If we don’t address these root causes—through policy, transparency, and equity-focused data collection—healthcare AI risks becoming another tool that widens disparities rather than closing them.