A new wave of research at Dartmouth College is shedding light on the complexities of agentic artificial intelligence, which is evolving beyond traditional reactive models to systems capable of autonomous decision-making. This shift raises important questions about the reliability and bias inherent in AI agents, particularly as they are deployed in critical areas such as healthcare, energy management, and quantum physics.
The term 'AI agent' originates from John McCarthy, a pioneer in the field who organized the landmark 1956 Dartmouth Summer Research Project on Artificial Intelligence. Today, researchers like Nikhil Singh, an assistant professor of computer science, are building on McCarthy's foundation. Singh's lab, the Science and Art of Human-AI Systems, focuses on understanding how these agents make decisions and the implications of those decisions in real-world scenarios.
The Challenge of Autonomous Decision-Making
Singh points out that an effective AI agent must make autonomous decisions under uncertainty. For example, if tasked with finding an ideal backpack, the agent must navigate the web, compare products, and shortlist the best options. However, research shows that AI agents display heightened biases, particularly when faced with default choices, often selecting these defaults more readily than humans. This raises concerns about the reliability of AI agents in decision-making roles.
In a recent study presented at the 2026 International Conference on Learning Representations, Singh and colleagues discovered that AI agents are significantly influenced by factors such as product pricing and user ratings. Their experiments in a simulated shopping environment revealed that marketing nudges, designed to attract attention to specific items, can systematically bias agents' choices. This highlights the need for a deeper understanding of how these agents process information and make decisions, especially before they are entrusted with significant responsibilities.
Evaluating Visual Decision-Making
The Dartmouth team has broadened its research to include visual agents that use computer vision instead of traditional text-based processing. This approach aims to understand how visual stimuli impact decision-making. By utilizing image generation models, the researchers found that even minor visual adjustments—such as enhanced lighting or contextual cues—could sway the choices made by AI agents. Remarkably, similar visual adjustments were also found to affect human decision-making, suggesting that the biases present in agentic AI could reflect those of their human counterparts.
Addressing Risks and Building Trust
Despite advancements, concerns about the potential misuse of AI agents remain. Eugene Santos Jr., a professor of engineering at Dartmouth, underscores the importance of transparency and reliability in AI systems, drawing parallels to engineering disciplines where reliability is essential. He cautions that AI technology, particularly in conversational agents, may provide confident responses even when lacking accurate knowledge, leading to harmful outcomes.
Santos advocates for a rigorous engineering approach to AI, emphasizing the need to understand the underlying decision-making processes to mitigate risks. By unpacking the incentives that drive AI systems, researchers can begin to tackle biases and enhance trust in human-AI collaborations.
Diverse Applications and Future Directions
Dartmouth’s research spans various fields, showcasing the versatility of AI agents. For instance, undergraduate student Catherine Chu is investigating how these agents can optimize experimental controls in quantum physics, aiming to streamline complex setups with minimal human input. Meanwhile, assistant professor Cong Chen is using AI agents as proxies for different types of energy consumers to provide insights that could reshape energy pricing and policy.
Healthcare is another critical area where agentic AI could significantly impact. Singh’s recent work on AI agents for diabetes management aims to transform how patients engage with their health data, offering personalized insights that traditional applications cannot provide. This innovative approach seeks to empower patients while ensuring their data privacy is protected.
As Dartmouth researchers continue to expand the capabilities of AI agents, their work reveals a dual narrative: the potential for AI to assist in complex problem-solving while highlighting the urgent need for ongoing scrutiny of these systems. Balancing technological advances with a critical assessment of their implications will be essential as society navigates the future of AI.
While agentic AI holds promise for driving innovation across various sectors, it also presents challenges that require careful consideration and proactive solutions to ensure safe and effective implementation.
The stories that move AI & crypto markets — before the market reacts.
Free. 7am ET. Five stories. 62,400 readers.
