AI and Human Collaboration Explored at Whiting School Event

On October 23, 2025, the Whiting School of Engineering’s Department of Computer Science at Johns Hopkins University hosted a pivotal talk by Aaron Roth, a professor of computer and cognitive science at the University of Pennsylvania. His presentation, titled “Agreement and Alignment for Human-AI Collaboration,” addressed the findings presented in three significant research papers: “Tractable Agreement Protocols,” “Collaborative Prediction: Tractable Information Aggregation via Agreement,” and “Emergent Alignment from Competition.”

As artificial intelligence increasingly permeates various sectors, researchers are keen to understand how AI can enhance human decision-making. Roth illustrated this through the example of AI assisting doctors in diagnosing patients. In this scenario, the AI processes data such as previous diagnoses, blood types, and symptoms to make predictions. The physician then reviews these predictions and can either agree or disagree based on their clinical expertise and knowledge of the patient.

Roth emphasized that in cases of disagreement, the AI and the physician can iterate their opinions in a finite number of rounds. Each round integrates their unique viewpoints until a consensus is reached. This concept is rooted in what Roth describes as a “common prior,” which assumes both parties start with the same foundational understanding of the world, even if they possess different evidence. This mutual knowledge leads to what Roth terms “Perfect Bayesian Rationality,” where both parties leverage their respective knowledge to facilitate agreement.

Despite its theoretical appeal, Roth acknowledged several challenges with this approach. Establishing a common prior is inherently complex, and navigating the intricacies of real-world scenarios can hinder reaching consensus, especially in multi-dimensional topics like hospital diagnostic codes.

In addressing the issue of agreement, Roth introduced the concept of calibration, which he likened to a test for accuracy. For instance, a weather forecaster can be evaluated based on their ability to predict true probabilities. “You can sort of design tests such that they would pass those tests if they were forecasting true probabilities,” he explained.

Roth then introduced the idea of conversation calibration between the doctor and the AI. This involves one party’s claims being influenced by the previous round’s statements from the other party. If the AI assesses a 40% risk in treatment and the doctor estimates it at 35%, the AI’s next assessment will fall between these two figures. This iterative process continues until both parties reach an agreement, thereby facilitating quicker resolutions.

While this discussion assumed a shared goal between the AI and the physician, Roth pointed out that this alignment is not always guaranteed. For example, if an AI model is developed by a pharmaceutical company, it may prioritize recommending its own products over others, potentially misaligning with the doctor’s objective for patient care. To mitigate this risk, Roth suggested that doctors consult multiple AI models. This would create competition among AI providers, compelling them to align their recommendations with patient welfare, as each provider strives to be the most trusted source.

Roth concluded by discussing the significance of real probabilities, which represent the true underlying factors that influence the world. While achieving perfect precision is rare, it is often adequate for probabilities to be unbiased under specific conditions. In these instances, reliable data can yield accurate estimations without necessitating the complexities of ideal reasoning.

Through these frameworks, AI and medical professionals can collaborate effectively, enhancing the accuracy of treatments, diagnoses, and ultimately improving patient outcomes.