Understanding AI in Medicine: The Crucial Principle of GIGO
By Yoon Hang Kim MD MPH
Over the past year, I have dedicated an hour a day to engaging with various AI models. Throughout this journey, one principle has consistently emerged: GIGO—“garbage in, garbage out.” This simple yet profound concept is particularly relevant in the realm of medical AI, where the quality of input, the structure of workflows, and the verification processes are essential determinants of the reliability of the output.Safeguarding the effective use of AI in medicine requires a careful approach to “stacking” tools and implementing built-in veracity checks. It’s vital to recognize that success does not hinge on a singular model; rather, it lies in comprehensively understanding each system's strengths and limitations.Take OpenEvidence, for instance. Designed specifically for medical applications, this model adheres closely to established medical protocols and provides answers that are well-supported by credible sources. In my observations, I have not encountered hallucinations—misleading or fabricated outputs—within its results.
However, this strength also presents a limitation: the model's heavy reliance on conventional guidelines can hinder its ability to tackle questions that lie outside standardized frameworks or those that require a more exploratory approach.
Other AI models, such as Gemini and Claude, also demonstrate robust performance metrics, marked by relatively low hallucination rates when utilized correctly. One effective strategy is implementing AI stacking—an approach that involves using tools like Perplexity for initial literature searches, followed by cross-validation of findings with models like OpenEvidence, Gemini, and Claude. This multi-layered workflow can significantly mitigate errors and enhance the reliability of the information retrieved.The error rate can be further diminished when users directly upload primary references into systems like NotebookLM (paired with Gemini) or Project Files (associated with Claude). By anchoring responses to specific source documents rather than relying solely on general training data, we can maintain a higher standard of veracity in the outputs.
Ultimately, the safety and efficacy of AI applications in medicine depend significantly on the operator's competency. While effective prompting is important, recognizing the inherent limitations of these systems is even more critical. The real safeguard against the pitfalls of AI lies in knowing when to step outside the AI environment and consult primary sources directly. AI has the potential to augment clinical reasoning, but it should never replace professional judgment or the necessity for independent source validation.As we navigate the ever-evolving landscape of medical AI, let us embrace the principle of GIGO and remain vigilant in our approach, ensuring we harness these powerful tools to their fullest potential while upholding the sanctity of medical practice and patient care.
