Artificial intelligence is reshaping nearly every industry, but few domains carry the stakes of healthcare. When a model suggests a cancer diagnosis that turns out to be wrong, the consequences are not just inconvenient—they are life-altering. That reality sits at the heart of a compelling conversation between Don Finley and Dr. Jade Zhou, an academic with over a decade of experience in electrical engineering, computer science, and the intersection of technology and healthcare.
From Engineering to Healthcare: A Natural Pivot
Dr. Zhou’s journey into healthcare AI was not a straight line. After earning her bachelor’s, master’s, and PhD in electrical engineering and computer science, she faced a familiar crossroads: follow her classmates into roles at major tech companies, or find a path where technology could serve a deeper human purpose. She chose the latter, drawn to healthcare by a simple but powerful observation: the United States spends an enormous portion of its GDP on healthcare, yet the system remains burdened by inefficiency, expensive labor, and administrative overhead. Technology, she believed, could serve as a critical assistant—not a replacement—for the humans operating within that system.
Two Pillars of Healthcare AI
When asked where technology could have the greatest impact in healthcare, Dr. Zhou pointed to two key applications. The first is AI-assisted clinical decision-making: using models to help physicians collect evidence, evaluate diagnostic options, and arrive at treatment plans more quickly. The second is using technology to reduce the labor burden on healthcare professionals, gradually lowering costs by automating portions of work that do not require human expertise.
The emphasis, however, is always on the word assist. AI can gather and synthesize evidence in a day or two what might previously have taken weeks. But the final decision—the judgment—must remain with the physician. This is what Dr. Zhou calls “trustworthy AI.”
What Makes AI Trustworthy?
Trustworthiness in AI is not a single metric. It is a combination of confidence, evidence, and human oversight. In a healthcare setting, an AI model might produce a high-confidence recommendation, but without transparent, traceable evidence backing that recommendation, it remains unreliable. The physician needs to see not just the conclusion but the reasoning path that led there.
Dr. Zhou draws a sharp distinction between casual AI use—asking a chatbot for writing suggestions—and clinical AI use. In healthcare, the margin for error is vanishingly small. A false positive cancer diagnosis is not a typo; it is a trauma. That is why trustworthy AI in medicine requires systems that can present their evidence, allow human verification at each step, and defer to the clinician for the final call.
The Human as the Magic Ingredient
Don Finley shared an observation from his own company that reinforced this idea: two people using the same AI platform can produce dramatically different results. The difference is not the model—it is the human. The prompts, the parameters, the depth of questioning, and the domain expertise the user brings all shape what the AI returns. In other words, the human is not just a reviewer of AI output; the human is an active participant whose expertise unlocks the model’s potential.
This has profound implications for how we think about AI in professional settings. The model is not the product. The interaction between the model and a skilled human is the product.
Vibe Coding vs. Agentic Engineering
The conversation turned to a growing phenomenon in software development: “vibe coding”—the practice of letting AI generate code with minimal human oversight. Dr. Zhou and Finley both observed that this approach creates a dangerous divide. On one side are engineers who use AI as a tool within a structured, step-by-step process: they plan, they break problems into manageable pieces, they validate each step before moving forward, and they maintain control over the direction of the work. On the other side are engineers who accept AI output at face value, sometimes with disastrous results.
Finley shared a striking example: an engineer allowed an AI to rewrite an entire application’s codebase rather than making a targeted fix, resulting in two parallel code paths that became a maintenance nightmare. The AI took the path of least resistance; the engineer failed to ask the right questions or spot-check the output. The lesson is clear: AI can generate enormous volumes of work in seconds, but that speed amplifies the need for human judgment, not diminishes it.
The Barrier Is Not Technology—It Is People
When asked about the biggest barriers to AI adoption in healthcare, Dr. Zhou identified two cultural gaps. First, many clinicians with deep domain expertise completely reject AI. They may use it casually—to draft emails or revise essays—but they do not embrace it as a clinical tool. They have not reached the point of seeing AI as an external neural system, a collaborator that can deepen their understanding and accelerate their work.
Second, many technical professionals lack the domain knowledge to build effective healthcare software. They can assemble systems and connect data sources, but when something goes wrong—when the quality of a clinical trial protocol design is off, or a drug dynamic measurement is misinterpreted—they have no judgment to fall back on. The software becomes technically functional but practically useless.
This dual gap—domain experts who resist AI and technologists who lack domain understanding—is the real bottleneck. The technology is ready. The people are not.
Rethinking Education and Hiring
The conversation closed with a provocative idea: AI may force us to rethink how we educate and hire. Traditional universities still categorize students into rigid majors, but in a world where a clinician can learn to build useful software with AI tools, and a computer scientist can use LLMs to design drug compounds, the boundaries between disciplines are blurring.
Dr. Zhou shared an extraordinary story: an IT professional in Australia whose dog was diagnosed with cancer and told there was no available treatment. Using LLMs and other AI tools to cross-validate evidence and design a drug compound, this person developed a treatment—and the dog survived. It is a single anecdote, but it signals something larger: domain expertise is no longer locked behind years of specialized training. AI can democratize access to knowledge, but only if the person wielding it has the judgment to use it responsibly.
Finley noted that traditional hiring paradigms—designed around domain-specific knowledge—are already breaking down. If an AI can pass most technical interviews, the question shifts from “What do you know?” to “How do you think?” How do you decompose a problem? How do you identify your own biases? How do you evaluate information from a place of neutrality? These are the skills that will define the next generation of professionals, regardless of their field.
The Takeaway
The message from this conversation is neither optimistic nor pessimistic—it is pragmatic. AI has enormous potential to improve healthcare, reduce costs, and accelerate discovery. But that potential is only realized when skilled humans remain in the loop, providing direction, validating evidence, and making the final call. The future of AI in healthcare is not about replacing human judgment. It is about building systems that make human judgment faster, better, and more informed than ever before.
The technology is ready. The question is whether we are ready to use it wisely.
