TLDR: A study compared dermatology treatment plans generated by human experts and two AI models (GPT-4o, o3). Human dermatologists scored human-generated plans higher, while an AI judge (Gemini 2.5 Pro) scored AI-generated plans, especially o3, significantly higher. This “evaluator effect” highlights a fundamental difference in how humans (experience-based) and AI (data-driven) assess clinical quality, suggesting a critical need for explainable AI and human-AI collaboration for effective integration into clinical practice.
The integration of artificial intelligence (AI) into healthcare is rapidly advancing, moving beyond simple diagnostics to more complex areas like treatment planning. However, evaluating the quality of these AI-generated clinical plans presents a unique challenge, especially when AI models offer solutions that might differ from traditional human practices.
Understanding the Study
A recent study titled “Divergent Realities: A Comparative Analysis of Human Expert vs. Artificial Intelligence Based Generation and Evaluation of Treatment Plans in Dermatology” aimed to compare treatment plans created by human experts and two different AI models: GPT-4o, a state-of-the-art generalist AI, and o3, an advanced reasoning AI. The research involved ten board-certified dermatologists and these two AI models, who independently generated treatment plans for five complex dermatological cases. These cases were designed to be open-ended, requiring nuanced clinical judgment rather than a single correct answer.
To ensure fairness, all 60 generated plans (5 cases x 12 participants) were anonymized and standardized in style and length by another AI model (GPT-4) before evaluation. The evaluation process was conducted in two distinct phases. In Phase 1, the ten human dermatologists scored the plans, blinded to their source. In Phase 2, a superior AI model, Gemini 2.5 Pro, evaluated the same plans using an identical scoring rubric, acting as an AI judge.
The Striking Results: A Tale of Two Evaluations
The study uncovered a profound and statistically significant phenomenon dubbed the “evaluator effect.” This means that the perceived quality of a treatment plan was fundamentally dependent on whether the evaluator was human or AI. The human experts and the AI judge held almost opposite views on the quality of plans generated by humans versus AI.
In the human-led evaluation, plans created by human experts were consistently scored higher than those from AI models. Human-generated plans received an average score of 7.62, while AI-generated plans averaged 7.16. Among all participants, human experts secured the top five positions. GPT-4o ranked 6th, and surprisingly, the advanced reasoning model, o3, ranked 11th out of 12 participants.
However, the AI judge’s evaluation completely reversed these findings. The Gemini 2.5 Pro model scored AI-generated plans significantly higher than human-generated ones. AI-generated plans had a mean score of 7.75, compared to 6.79 for human-generated plans. In this phase, the o3 model, which was ranked near last by humans, was elevated to 1st place by the AI judge, and GPT-4o ranked 2nd. Consequently, all ten human experts were ranked below the two AI models by the AI judge.
Why the Discrepancy?
The study suggests that this divergence likely stems from the inherent nature of human clinical judgment and the influence of cognitive biases. Human expertise is shaped by real-world experience, including resource limitations and local practice norms. Humans might unconsciously favor plans that align with their familiar approaches, a phenomenon known as confirmation bias or status quo bias. An AI’s knowledge, on the other hand, is derived from a vast amount of literature and datasets, which can lead to recommendations that are technically optimal but might seem unconventional or impractical to human experts.
The “o3 paradox” highlights this point: o3, by leveraging the latest global literature, might have proposed highly advanced or novel strategies not yet widely practiced by the human experts in India. While a human expert might perceive such an unfamiliar plan as incorrect, the AI judge, operating on similar data-driven logic, would recognize its evidence-based optimality.
Also Read:
- The Unseen Cost of AI Transparency: How Disclosure Shapes Perceptions of Writing and Authors
- Beyond Sentiment: Unpacking AI’s Journey Towards Empathetic Conversation
Looking Ahead: The Path to Synergy
The most crucial takeaway from these findings is that neither human experts nor AI systems are independently sufficient for optimal clinical planning in their current states. Each brings unique strengths. The ideal path forward appears to be a “human-in-the-loop” system, combining AI’s consistency and data-processing power with the human clinician’s holistic understanding and contextual judgment.
For such collaboration to be effective, the study emphasizes the critical importance of Explainable AI (XAI). If an AI proposes a correct but unconventional plan without a clear explanation, human experts, influenced by their biases, are likely to reject it. AI systems must not only provide recommendations but also articulate the reasoning behind them to bridge the gap in understanding and foster trust.
It’s important to note that this study had limitations, including a small sample size of experts and cases, and its focus solely on dermatology. The findings might differ in other medical fields.
In conclusion, this research highlights a fundamental difference in how humans and AI evaluate clinical treatment plans. Humans assess through the lens of experience and practical heuristics, while AI evaluates through data-driven logic. The “better” plan is relative to the evaluator. The future of clinical decision support lies not in determining which is superior, but in creating synergistic systems where AI enhances human expertise with its vast knowledge, and humans guide AI with their irreplaceable understanding of nuance and patient context.
